Tag Archives: Kinesis Data Streams

Optimize write throughput for Amazon Kinesis Data Streams

Post Syndicated from Buddhike de Silva original https://aws.amazon.com/blogs/big-data/optimize-write-throughput-for-amazon-kinesis-data-streams/

Amazon Kinesis Data Streams is used by many customers to capture, process, and store data streams at any scale. This level of unparalleled scale is enabled by dividing each data stream into multiple shards. Each shard in a stream has a 1 Mbps or 1,000 records per second write throughput limit. Whether your data streaming application is collecting clickstream data from a web application or recording telemetry data from billions of Internet of Things (IoT) devices, streaming applications are highly susceptible to a varying amount of data ingestion. Sometimes such a large and unexpected volume of data could be the thing we least expect. For instance, consider application logic with a retry mechanism when writing records to a Kinesis data stream. In case of a network failure, it’s common to buffer data locally and write them when connectivity is restored. Depending on the rate that data is buffered and the duration of connectivity issue, the local buffer can accumulate enough data that could saturate the available write throughput quota of a Kinesis data stream.

When an application attempts to write more data than what is allowed, it will receive write throughput exceeded errors. In some instances, not being able to address these errors in a timely manner can result in data loss, unhappy customers, and other undesirable outcomes. In this post, we explore the typical reasons behind write throughput exceeded errors, along with methods to identify them. We then guide you on swift responses to these events and provide several solutions for mitigation. Lastly, we delve into how on-demand capacity mode can be valuable in addressing these errors.

Why do we get write throughput exceeded errors?

Write throughput exceeded errors are generally caused by three different scenarios:

  • The simplest is the case where the producer application is generating more data than the throughput available in the Kinesis data stream (the sum of all shards).
  • Next, we have the case where data distribution is not even across all shards, known as hot shard issue.
  • Write throughout errors can also be caused by an application choosing a partition key to write records at a rate exceeding the throughput offered by a single shard. This situation is somewhat similar to hot shard issue, but as we see later in this post, unlike a hot shard issue, you can’t solve this problem by adding more shards to the data stream. This behavior is commonly known as a hot key issue.

Before we discuss how to diagnose these issues, let’s look at how Kinesis data streams organize data and its relationship to write throughput exceeded errors.

A Kinesis data stream has one or more shards to store data. Each shard is assigned a key range in 128-bit integer space. If you view the details of a data stream using the describe-stream operation in the AWS Command Line Interface (AWS CLI), you can actually see this key range assignment:

$ aws kinesis describe-stream --stream-name my-data-stream
"StreamDescription": {
  "Shards": [
    {
      "ShardId": "shardId-000000000000",
      "HashKeyRange": {
        "StartingHashKey": "0",
        "EndingHashKey": 
        "85070591730234615865843651857942052863"
       }
    },
    {
       "ShardId": "shardId-000000000001",
       "HashKeyRange": {
       "StartingHashKey": 
          "85070591730234615865843651857942052864",
       "EndingHashKey": 
         "170141183460469231731687303715884105727"
       }
    }
  ]
}

When a producer application invokes the PutRecord or PutRecords API, the service calculates a MD5 hash for the PartitionKey specified in the record. The resulting hash is used to determine which shard to store that record. You can take more control over this process by setting the ExplicitHashKey property in the PutRecord request to a hash key that falls within a specific shard’s key range. For instance, setting ExplicitHashKey to 0 will guarantee that record is written to shard ID shardId-0 in the stream described in the preceding code snippet.

How partition keys are distributed across available shards plays a vital role in maximizing the available throughput in a Kinesis data stream. When the partition key being used is repeated frequently in a way that some keys are more frequent than the others, shards storing those records will be utilized more. We also get the same net effect if we use ExplicitHashKey and our logic for choosing the hash key is biased towards a subset of shards.

Imagine you have a fleet of web servers logging performance metrics for each web request served into a Kinesis data stream with two shards and you used a request URL as the partition key. Each time a request is served, the application makes a call to the PutRecord API carrying a 10-bytes record. Let’s say that you have a total of 10 URLs and each receives 10 requests per second. Under these circumstances, total throughput required for the workload is 1,000 bytes per second and 100 requests per second. If we assume perfect distribution of 10 URLs across the two shards, each shard will receive 500 bytes per second and 50 requests per second.

Now imagine one of these URLs went viral and it started receiving 1,000 requests per second. Although the situation is positive from a business point of view, you’re now on the brink of making users unhappy. After the page gained popularity, you’re now counting 1,040 requests per second for the shard storing the popular URL (1000 + 10 * 4). At this point, you’ll receive write throughput exceeded errors from that shard. You’re throttled based on the requests per second quota because even with increased requests, you’re still generating approximately 11 KB of data.

You can solve this problem either by using a UUID for each request as the partition key so that you share the total load across both shards, or by adding more shards to the Kinesis data stream. The method you choose depends on how you want to consume data. Changing the partition key to a UUID would be problematic if you want performance metrics from a given URL to be always processed by the same consumer instance or if you want to maintain the order of records on a per-URL basis.

Knowing the exact cause of write throughout exceeded errors is an important step in remediating them. In the next sections, we discuss how to identify the root cause and remediate this problem.

Identifying the cause of write throughput exceeded errors

The first step in solving a problem is that knowing that it exists. You can use the WriteProvisionedThrougputExceeded metric in Amazon CloudWatch in this case. You can correlate the spikes in the WriteProvisionedThrougputExceeded metric to the IncomingBytes and IncomingRecords metrics to identify whether an application is getting throttled due to the size of data or the number of records written.

Let’s look at a few tests we performed in a stream with two shards to illustrate various scenarios. In this instance, with two shards in our stream, total throughput available to our producer application is either 2 Mbps or 2,000 records per second.

In the first test, we ran a producer to write batches of 30 records, each being 100 KB, using the PutRecords API. As you can see in the graph on the left of the following figure, our WriteProvisionedThroughputExceedded errors count went up. The graph on the right shows that we are reaching the 2 Mbps limit, but our incoming records rate is much lower than the 2,000 records per second limit (Kinesis metrics are published at 1-minute intervals, hence 125.8 and 120,000 as upper limits).Record size based throttling example

The following figures show how the same three metrics changed when we changed the producer to write batches of 500 records, each being 50 bytes, in the second test. This time, we exceeded the 2,000 records per second throughput limit, but our incoming bytes rate is well under the limit.

Record count based throttling

Now that we know that problem exists, we should look for clues to see if we’re exceeding the overall throughput available in the stream or if we’re having a hot shard issue due to an imbalanced partition key distribution as discussed earlier. One approach to this is to use enhanced shard-level metrics. Prior to our tests, we enabled enhanced shard-level metrics, and we can see in the following figure that both shards equally reached their quota in our first test.

Enhanced shard level metrics

We have seen Kinesis data streams containing thousands of shards harnessing the power of infinite scale in Kinesis data streams. However, plotting enhanced shard-level metrics on a such large stream may not provide an easy to way to find out which shards are over-utilized. In that instance, it’s better to use CloudWatch Metrics Insights to run queries to view top-n items, as shown in the following code (adjust the LIMIT 5 clause accordingly):

-- Show top 5 shards with highest incoming bytes
SELECT
SUM(IncomingBytes)
FROM "AWS/Kinesis"
GROUP BY ShardId, StreamName
ORDER BY MAX() DESC
LIMIT 5

-- Show top 5 shards with highest incoming records
SELECT
SUM(IncomingRecords)
FROM "AWS/Kinesis"
GROUP BY ShardId, StreamName
ORDER BY MAX() DESC
LIMIT 5

Enhanced shard-level metrics are not enabled by default. If you didn’t enable them and you want to perform root cause analysis after an incident, this option isn’t very helpful. In addition, you can only query the most recent 3 hours of data. Enhanced shard-level metrics also incur additional costs for CloudWatch metrics and it may be cost prohibitive to have it always on in data streams with a lot of shards.

One interesting scenario is when the workload is bursty, which can make the resulting CloudWatch metrics graphs rather baffling. This is because Kinesis publishes CloudWatch metric data aggregated at 1-minute intervals. Consequently, although you can see write throughput exceeded errors, your incoming bytes/records graphs may be still within the limits. To illustrate this scenario, we changed our test to create a burst of writes exceeding the limits and then sleep for a few seconds. Then we repeated this cycle for several minutes to yield the graphs in the following figure, which show write throughput exceeded errors on the left, but the IncomingBytes and IncomingRecords graphs on the right seem fine.

Effect of one data aggregated at 1-minute intervals

To enhance the process of identifying write throughput exceeded errors, we developed a CLI tool called Kinesis Hot Shard Advisor (KHS). With KHS, you can view shard utilization when shard-level metrics are not enabled. This is particularly useful for investigating an issue retrospectively. It can also show most frequently written keys to a particular shard. KHS reports shard utilization by reading records and aggregating them per second intervals based on the ApproximateArrivalTimestamp in the record. Because of this, you can also understand shard utilization drivers during bursty write workloads.

By running the following command, we can get KHS to inspect the data that arrived in 1 minute during our first test and generate a report:

khs -stream my-data-stream -from "2023-06-22 17:35:00" -to "2023-06-22 17:36:00"

For the given time window, the summary section in the generated report shows the maximum bytes per second rate observed, total bytes ingested, maximum records per second observed, and the total number of records ingested for each shard.

KHS report summary

Choosing a shard ID in the first column will display a graph of incoming bytes and records for that shard. This is similar to the graph you get in CloudWatch metrics, except the KHS graph reports on a per-second basis. For instance, in the following figure, we can see how the producer was going through a series of bursty writes followed by a throttling event during our test case.

KHS shard level metrics display

Running the same command with the -aggregate-key option enables partition key distribution analysis. It generates an additional graph for each shard showing the key distribution, as shown in the following figure. For our test scenario, we can only see each key being used one time because we used a new UUID for each record.

KHS key distribution graph

Because KHS reports based on data stored in streams, it creates an enhanced fan-out consumer at startup to prevent using the read throughput quota available for other consumers. When the analysis is complete, it deletes that enhanced fan-out consumer.

Due its nature of reading data streams, KHS can transfer a lot of data during analysis. For instance, assume you have a stream with 100 shards. If all of them are fully utilized during a minute window specified using -from and -to arguments, the host running KHS will receive at least 1 MB * 100 * 60 = 6000 MB = approximately 6 GB data. To avoid this kind of excessive data transfer and speed up the analysis process, we recommend first using the WriteProvisionedThroughoutExceeded CloudWatch metric to identify a time period when you experienced throttling and use a small window (such as 10 seconds) with KHS. You can also run KHS in an Amazon Elastic Compute Cloud (Amazon EC2) instance in the same AWS Region as your Kinesis data stream to minimize network latency during reads.

KHS is designed to run in a single machine to diagnose large-scale workloads. Using a naive in-memory-based counting algorithm (such as a hash map storing the partition key and count) for partition key distribution analysis could easily exhaust the available memory in the host system. Therefore, we use a probabilistic data structure called count-min-sketch to estimate the number of times a key has been used. As a result, the number you see in the report should be taken as an approximate value rather than an absolute value. After all, with this report, we just want to find out if there’s an imbalance in the keys written to a shard.

Now that we understand what causes hot shards and how to identify them, let’s look at how to deal with this in producer applications and remediation steps.

Remediation steps

Having producers retry writes is a step towards making our producers resilient to write throughput exceeded errors. Consider our earlier sample application logging performance metrics data for each web request served by a fleet of web servers. When implementing this retry mechanism, you should remember that records that are not written to the Kinesis stream are going to be in host system’s memory. The first issue with this is, if the host crashes before the records could be written, you’ll experience data loss. Scenarios such as tracking web request performance data might be more forgiving for this type of data loss than scenarios like financial transactions. You should evaluate durability guarantees required for your application and employ techniques to achieve them.

The second issue is that records waiting to be written to the Kinesis data stream are going to consume the host system’s memory. When you start getting throttled and have some retry logic in place, you should notice that your memory utilization is going up. A retry mechanism should have a way to avoid exhausting the host system’s memory.

With the appropriate retry logic in place, if you receive write throughput exceeded errors, you can use the methods we discussed earlier to identify the cause. After you identify the root cause, you can choose the appropriate remediation step:

  • If the producer application is exceeding the overall stream’s throughput, you can add more shards to the stream to increase its write throughput capacity. When adding shards, the Kinesis data stream makes the new shards available incrementally, minimizing the time that producers experience write throughput exceeded errors. To add shards to a stream, you can use the Kinesis console, the update-shard-count operation in the AWS CLI, the UpdateShardCount API through the AWS SDK, or the ShardCount property in the AWS CloudFormation template used to create the stream.
  • If the producer application is exceeding the throughput limit of some shards (hot shard issue), pick one of the following options based on consumer requirements:
    • If locality of data is required (records with the same partition key are always processed by the same consumer) or an order based on partition key is required, use the split-shard operation in the AWS CLI or the SplitShard API in the AWS SDK to split those shards.
    • If locality or order based on the current partition key is not required, change the partition key scheme to increase its distribution.
  • If the producer application is exceeding the throughput limit of a shard due to a single partition key (hot key issue), change the partition key scheme to increase its distribution.

Kinesis Data Streams also has an on-demand capacity mode. In on-demand capacity mode, Kinesis Data Streams automatically scales streams when needed. Additionally, you can switch between on-demand and provisioned capacity modes without causing an outage. This could be particularly useful when you’re experiencing write throughput exceeded errors but require immediate reaction to keep your application available to your users. In such instances, you can switch a provisioned capacity mode data stream to an on-demand data stream and let Kinesis Data Streams handle the required scale appropriately. You can then perform root cause analysis in the background and take corrective actions. Finally, if necessary, you can change the capacity mode back to provisioned.

Conclusion

You should now have a solid understanding of the common causes of write throughput exceeded errors in Kinesis data streams, how to diagnose them, and what actions to take to appropriately deal with them. We hope that this post will help you make your Kinesis Data Streams applications more robust. If you are just starting with Kinesis Data Streams, we recommend referring to the Developer Guide.

If you have any questions or feedback, please leave them in the comments section.

About the Authors

Buddhike de Silva is a Senior Specialist Solutions Architect at Amazon Web Services. Buddhike helps customers run large scale streaming analytics workloads on AWS and make the best out of their cloud journey.

Nihar Sheth is a Senior Product Manager at Amazon Web Services. He is passionate about developing intuitive product experiences that solve complex customer problems and enable customers to achieve their business goals.

Architectural Patterns for real-time analytics using Amazon Kinesis Data Streams, Part 2: AI Applications

Post Syndicated from Raghavarao Sodabathina original https://aws.amazon.com/blogs/big-data/architectural-patterns-for-real-time-analytics-using-amazon-kinesis-data-streams-part-2-ai-applications/

Welcome back to our exciting exploration of architectural patterns for real-time analytics with Amazon Kinesis Data Streams! In this fast-paced world, Kinesis Data Streams stands out as a versatile and robust solution to tackle a wide range of use cases with real-time data, from dashboarding to powering artificial intelligence (AI) applications. In this series, we streamline the process of identifying and applying the most suitable architecture for your business requirements, and help kickstart your system development efficiently with examples.

Before we dive in, we recommend reviewing Architectural patterns for real-time analytics using Amazon Kinesis Data Streams, part 1 for the basic functionalities of Kinesis Data Streams. Part 1 also contains architectural examples for building real-time applications for time series data and event-sourcing microservices.

Now get ready as we embark on the second part of this series, where we focus on the AI applications with Kinesis Data Streams in three scenarios: real-time generative business intelligence (BI), real-time recommendation systems, and Internet of Things (IoT) data streaming and inferencing.

Real-time generative BI dashboards with Kinesis Data Streams, Amazon QuickSight, and Amazon Q

In today’s data-driven landscape, your organization likely possesses a vast amount of time-sensitive information that can be used to gain a competitive edge. The key to unlock the full potential of this real-time data lies in your ability to effectively make sense of it and transform it into actionable insights in real time. This is where real-time BI tools such as live dashboards come into play, assisting you with data aggregation, analysis, and visualization, therefore accelerating your decision-making process.

To help streamline this process and empower your team with real-time insights, Amazon has introduced Amazon Q in QuickSight. Amazon Q is a generative AI-powered assistant that you can configure to answer questions, provide summaries, generate content, and complete tasks based on your data. Amazon QuickSight is a fast, cloud-powered BI service that delivers insights.

With Amazon Q in QuickSight, you can use natural language prompts to build, discover, and share meaningful insights in seconds, creating context-aware data Q&A experiences and interactive data stories from the real-time data. For example, you can ask “Which products grew the most year-over-year?” and Amazon Q will automatically parse the questions to understand the intent, retrieve the corresponding data, and return the answer in the form of a number, chart, or table in QuickSight.

By using the architecture illustrated in the following figure, your organization can harness the power of streaming data and transform it into visually compelling and informative dashboards that provide real-time insights. With the power of natural language querying and automated insights at your fingertips, you’ll be well-equipped to make informed decisions and stay ahead in today’s competitive business landscape.

Build real-time generative business intelligence dashboards with Amazon Kinesis Data Streams, Amazon QuickSight, and Amazon Qtreaming & inferencing pipeline with AWS IoT & Amazon SageMaker

The steps in the workflow are as follows:

  1. We use Amazon DynamoDB here as an example for the primary data store. Kinesis Data Streams can ingest data in real time from data stores such as DynamoDB to capture item-level changes in your table.
  2. After capturing data to Kinesis Data Streams, you can ingest the data into analytic databases such as Amazon Redshift in near-real time. Amazon Redshift Streaming Ingestion simplifies data pipelines by letting you create materialized views directly on top of data streams. With this capability, you can use SQL (Structured Query Language) to connect to and directly ingest the data stream from Kinesis Data Streams to analyze and run complex analytical queries.
  3. After the data is in Amazon Redshift, you can create a business report using QuickSight. Connectivity between a QuickSight dashboard and Amazon Redshift enables you to deliver visualization and insights. With the power of Amazon Q in QuickSight, you can quickly build and refine the analytics and visuals with natural language inputs.

For more details on how customers have built near real-time BI dashboards using Kinesis Data Streams, refer to the following:

Real-time recommendation systems with Kinesis Data Streams and Amazon Personalize

Imagine creating a user experience so personalized and engaging that your customers feel truly valued and appreciated. By using real-time data about user behavior, you can tailor each user’s experience to their unique preferences and needs, fostering a deep connection between your brand and your audience. You can achieve this by using Kinesis Data Streams and Amazon Personalize, a fully managed machine learning (ML) service that generates product and content recommendations for your users, instead of building your own recommendation engine from scratch.

With Kinesis Data Streams, your organization can effortlessly ingest user behavior data from millions of endpoints into a centralized data stream in real time. This allows recommendation engines such as Amazon Personalize to read from the centralized data stream and generate personalized recommendations for each user on the fly. Additionally, you could use enhanced fan-out to deliver dedicated throughput to your mission-critical consumers at even lower latency, further enhancing the responsiveness of your real-time recommendation system. The following figure illustrates a typical architecture for building real-time recommendations with Amazon Personalize.

Build real-time recommendation systems with Kinesis Data Streams and Amazon Personalize

The steps are as follows:

  1. Create a dataset group, schemas, and datasets that represent your items, interactions, and user data.
  2. Select the best recipe matching your use case after importing your datasets into a dataset group using Amazon Simple Storage Service(Amazon S3), and then create a solution to train a model by creating a solution version. When your solution version is complete, you can create a campaign for your solution version.
  3. After a campaign has been created, you can integrate calls to the campaign in your application. This is where calls to the GetRecommendations or GetPersonalizedRanking APIs are made to request near-real-time recommendations from Amazon Personalize. Your website or mobile application calls a AWS Lambda function over Amazon API Gateway to receive recommendations for your business apps.
  4. An event tracker provides an endpoint that allows you to stream interactions that occur in your application back to Amazon Personalize in near-real time. You do this by using the PutEvents API. You can build an event collection pipeline using API Gateway, Kinesis Data Streams, and Lambda to receive and forward interactions to Amazon Personalize. The event tracker performs two primary functions. First, it persists all streamed interactions so they will be incorporated into future retrainings of your model. This is also how Amazon Personalize cold starts new users. When a new user visits your site, Amazon Personalize will recommend popular items. After you stream in an event or two, Amazon Personalize immediately starts adjusting recommendations.

To learn how other customers have built personalized recommendations using Kinesis Data Streams, refer to the following:

Real-time IoT data streaming and inferencing with AWS IoT Core and Amazon SageMaker

From office lights that automatically turn on as you enter the room to medical devices that monitors a patient’s health in real time, a proliferation of smart devices is making the world more automated and connected. In technical terms, IoT is the network of devices that connect with the internet and can exchange data with other devices and software systems. Many organizations increasingly rely on the real-time data from IoT devices, such as temperature sensors and medical equipment, to drive automation, analytics, and AI systems. It’s important to choose a robust streaming solution that can achieve very low latency and handle high volumes of data throughputs to power the real-time AI inferencing.

With Kinesis Data Streams, IoT data across millions of devices can simultaneously write to a centralized data stream. Alternatively, you can use AWS IoT Core to securely connect and easily manage the fleet of IoT devices, collect the IoT data, and then ingest to Kinesis Data Streams for real-time transformation, analytics, and event-driven microservices. Then, you can use integrated services such as Amazon SageMaker for real-time inference. The following diagram depicts the high-level streaming architecture with IoT sensor data.

Build real-time IoT data streaming & inferencing pipeline with AWS IoT & Amazon SageMaker

The steps are as follows:

  1. Data originates in IoT devices such as medical devices, car sensors, and industrial IoT sensors. This telemetry data is collected using AWS IoT Greengrass, an open source IoT edge runtime and cloud service that helps your devices collect and analyze data closer to where the data is generated.
  2. Event data is ingested into the cloud using edge-to-cloud interface services such as AWS IoT Core, a managed cloud platform that connects, manages, and scales devices effortlessly and securely. You can also use AWS IoT SiteWise, a managed service that helps you collect, model, analyze, and visualize data from industrial equipment at scale. Alternatively, IoT devices could send data directly to Kinesis Data Streams.
  3. AWS IoT Core can stream ingested data into Kinesis Data Streams.
  4. The ingested data gets transformed and analyzed in near real time using Amazon Managed Service for Apache Flink. Stream data can further be enriched using lookup data hosted in a data warehouse such as Amazon Redshift. Managed Service for Apache Flink can persist streamed data into Amazon Redshift after the customer’s integration and stream aggregation (for example, 1 minute or 5 minutes). The results in Amazon Redshift can be used for further downstream BI reporting services, such as QuickSight. Managed Service for Apache Flink can also write to a Lambda function, which can invoke SageMaker models. After the ML model is trained and deployed in SageMaker, inferences are invoked in a microbatch using Lambda. Inferenced data is sent to Amazon OpenSearch Service to create personalized monitoring dashboards using OpenSearch Dashboards. The transformed IoT sensor data can be stored in DynamoDB. You can use AWS AppSync to provide near real-time data queries to API services for downstream applications. These enterprise applications can be mobile apps or business applications to track and monitor the IoT sensor data in near real time.
  5. The streamed IoT data can be written to an Amazon Data Firehose delivery stream, which microbatches data into Amazon S3 for future analytics.

To learn how other customers have built IoT device monitoring solutions using Kinesis Data Streams, refer to:

Conclusion

This post demonstrated additional architectural patterns for building low-latency AI applications with Kinesis Data Streams and its integrations with other AWS services. Customers looking to build generative BI, recommendation systems, and IoT data streaming and inferencing can refer to these patterns as the starting point of designing your cloud architecture. We will continue to add new architectural patterns in the future posts of this series.

For detailed architectural patterns, refer to the following resources:

If you want to build a data vision and strategy, check out the AWS Data-Driven Everything (D2E) program.

About the Authors

Raghavarao Sodabathina is a Principal Solutions Architect at AWS, focusing on Data Analytics, AI/ML, and cloud security. He engages with customers to create innovative solutions that address customer business problems and to accelerate the adoption of AWS services. In his spare time, Raghavarao enjoys spending time with his family, reading books, and watching movies.

Hang Zuo is a Senior Product Manager on the Amazon Kinesis Data Streams team at Amazon Web Services. He is passionate about developing intuitive product experiences that solve complex customer problems and enable customers to achieve their business goals.

Shwetha Radhakrishnan is a Solutions Architect for AWS with a focus in Data Analytics. She has been building solutions that drive cloud adoption and help organizations make data-driven decisions within the public sector. Outside of work, she loves dancing, spending time with friends and family, and traveling.

Brittany Ly is a Solutions Architect at AWS. She is focused on helping enterprise customers with their cloud adoption and modernization journey and has an interest in the security and analytics field. Outside of work, she loves to spend time with her dog and play pickleball.

Build Spark Structured Streaming applications with the open source connector for Amazon Kinesis Data Streams

Post Syndicated from Idan Maizlits original https://aws.amazon.com/blogs/big-data/build-spark-structured-streaming-applications-with-the-open-source-connector-for-amazon-kinesis-data-streams/

Apache Spark is a powerful big data engine used for large-scale data analytics. Its in-memory computing makes it great for iterative algorithms and interactive queries. You can use Apache Spark to process streaming data from a variety of streaming sources, including Amazon Kinesis Data Streams for use cases like clickstream analysis, fraud detection, and more. Kinesis Data Streams is a serverless streaming data service that makes it straightforward to capture, process, and store data streams at any scale.

With the new open source Amazon Kinesis Data Streams Connector for Spark Structured Streaming, you can use the newer Spark Data Sources API. It also supports enhanced fan-out for dedicated read throughput and faster stream processing. In this post, we deep dive into the internal details of the connector and show you how to use it to consume and produce records from and to Kinesis Data Streams using Amazon EMR.

Introducing the Kinesis Data Streams connector for Spark Structured Streaming

The Kinesis Data Streams connector for Spark Structured Streaming is an open source connector that supports both provisioned and On-Demand capacity modes offered by Kinesis Data Streams. The connector is built using the latest Spark Data Sources API V2, which uses Spark optimizations. Starting with Amazon EMR 7.1, the connector comes pre-packaged on Amazon EMR on Amazon EKS, Amazon EMR on Amazon EC2, and Amazon EMR Serverless, so you don’t need to build or download any packages. For using it with other Apache Spark platforms, the connector is available as a public JAR file that can be directly referred to while submitting a Spark Structured Streaming job. Additionally, you can download and build the connector from the GitHub repo.

Kinesis Data Streams supports two types of consumers: shared throughput and dedicated throughput. With shared throughput, 2 Mbps of read throughput per shard is shared across consumers. With dedicated throughput, also known as enhanced fan-out, 2 Mbps of read throughput per shard is dedicated to each consumer. This new connector supports both consumer types out of the box without any additional coding, providing you the flexibility to consume records from your streams based on your requirements. By default, this connector uses a shared throughput consumer, but you can configure it to use enhanced fan-out in the configuration properties.

You can also use the connector as a sink connector to produce records to a Kinesis data stream. The configuration parameters for using the connector as a source and sink differ—for more information, see Kinesis Source Configuration. The connector also supports multiple storage options, including Amazon DynamoDB, Amazon Simple Service for Storage (Amazon S3), and HDFS, to store checkpoints and provide continuity.

For scenarios where a Kinesis data stream is deployed in an AWS producer account and the Spark Structured Streaming application is in a different AWS consumer account, you can use the connector to do cross-account processing. This requires additional Identity and Access Management (IAM) trust policies to allow the Spark Structured Streaming application in the consumer account to assume the role in the producer account.

You should also consider reviewing the security configuration with your security teams based on your data security requirements.

How the connector works

Consuming records from Kinesis Data Streams using the connector involves multiple steps. The following architecture diagram shows the internal details of how the connector works. A Spark Structured Streaming application consumes records from a Kinesis data stream source and produces records to another Kinesis data stream.

A Kinesis data stream is composed of set of shards. A shard is a uniquely identified sequence of data records in a stream and provides a fixed unit of capacity. The total capacity of the stream is the sum of the capacity of all of its shards.

A Spark application consists of a driver and a set of executor processes. The Spark driver acts as a coordinator, and the tasks running in executors are responsible for producing and consuming records to and from shards.

The solution workflow includes the following steps:

  1. Internally, by default, Structured Streaming queries are processed using a micro-batch processing engine, which processes data streams as a series of small batch jobs. At the beginning of a micro-batch run, the driver uses the Kinesis Data Streams ListShard API to determine the latest description of all available shards. The connector exposes a parameter (kinesis.describeShardInterval) to configure the interval between two successive ListShard API calls.
  2. The driver then determines the starting position in each shard. If the application is a new job, the starting position of each shard is determined by kinesis.startingPosition. If it’s a restart of an existing job, it’s read from last record metadata checkpoint from storage (for this post, DynamoDB) and ignores kinesis.startingPosition.
  3. Each shard is mapped to one task in an executor, which is responsible for reading data. The Spark application automatically creates an equal number of tasks based on the number of shards and distributes it across the executors.
  4. The tasks in an executor use either polling mode (shared) or push mode (enhanced fan-out) to get data records from the starting position for a shard.
  5. Spark tasks running in the executors write the processed data to the data sink. In this architecture, we use the Kinesis Data Streams sink to illustrate how the connector writes back to the stream. Executors can write to more than one Kinesis Data Streams output shard.
  6. At the end of each task, the corresponding executor process saves the metadata (checkpoint) about the last record read for each shard in the offset storage (for this post, DynamoDB). This information is used by the driver in the construction of the next micro-batch.

Solution overview

The following diagram shows an example architecture of how to use the connector to read from one Kinesis data stream and write to another.

In this architecture, we use the Amazon Kinesis Data Generator (KDG) to generate sample streaming data (random events per country) to a Kinesis Data Streams source. We start an interactive Spark Structured Streaming session and consume data from the Kinesis data stream, and then write to another Kinesis data stream.

We use Spark Structured Streaming to count events per micro-batch window. These events for each country are being consumed from Kinesis Data Streams. After the count, we can see the results.

Prerequisites

To get started, follow the instructions in the GitHub repo. You need the following prerequisites:

After you deploy the solution using the AWS CDK, you will have the following resources:

  • An EMR cluster with the Kinesis Spark connector installed
  • A Kinesis Data Streams source
  • A Kinesis Data Streams sink

Create your Spark Structured Streaming application

After the deployment is complete, you can access the EMR primary node to start a Spark application and write your Spark Structured Streaming logic.

As we mentioned earlier, you use the new open source Kinesis Spark connector to consume data from Amazon EMR. You can find the connector code on the GitHub repo along with examples on how to build and set up the connector in Spark.

In this post, we use Amazon EMR 7.1, where the connector is natively available. If you’re not using Amazon EMR 7.1 and above, you can use the connector by running the following code:

cd /usr/lib/spark/jars 
sudo wget https://awslabs-code-us-east-1.s3.amazonaws.com/spark-sql-kinesis-connector/spark-streaming-sql-kinesis-connector_2.12-1.2.1.jar
sudo chmod 755 spark-streaming-sql-kinesis-connector_2.12-1.2.1.jar

Complete the following steps:

  1. On the Amazon EMR console, navigate to the emr-spark-kinesis cluster.
  2. On the Instances tab, select the primary instance and choose the Amazon Elastic Compute Cloud (Amazon EC2) instance ID.

You’re redirected to the Amazon EC2 console.

  1. On the Amazon EC2 console, select the primary instance and choose Connect.
  2. Use Session Manager, a capability of AWS Systems Manager, to connect to the instance.
  3. Because the user that is used to connect is the ssm-user, we need to switch to the Hadoop user: 
    sudo su hadoop

  4. Start a Spark shell either using Scala or Python to interactively build a Spark Structured Streaming application to consume data from a Kinesis data stream.

For this post, we use Python for writing to a stream using a PySpark shell in Amazon EMR.

  1. Start the PySpark shell by entering the command pyspark.

Because you already have the connector installed in the EMR cluster, you can now create the Kinesis source.

  1. Create the Kinesis source with the following code: 
    kinesis = spark.readStream.format("aws-kinesis") \
    .option("kinesis.region", "<aws-region>") \
    .option("kinesis.streamName", "kinesis-source") \
    .option("kinesis.consumerType", "GetRecords") \
    .option("kinesis.endpointUrl", "https://kinesis.<aws-region>.amazonaws.com") \
    .option("kinesis.startingposition", "LATEST") \
    .load()

For creating the Kinesis source, the following parameters are required:

  • Name of the connector – We use the connector name aws-kinesis
  • kinesis.region – The AWS Region of the Kinesis data stream you are consuming
  • kinesis.consumerType – Use GetRecords (standard consumer) or SubscribeToShard (enhanced fan-out consumer)
  • kinesis.endpointURL – The Regional Kinesis endpoint (for more details, see Service endpoints)
  • kinesis.startingposition – Choose LATEST, TRIM_HORIZON, or AT_TIMESTAMP (refer to ShardIteratorType)

For using an enhanced fan-out consumer, additional parameters are needed, such as the consumer name. The additional configuration can be found in the connector’s GitHub repo.

kinesis_efo = spark \
.readStream \
.format("aws-kinesis") \
.option("kinesis.region", "<aws-region>") \
.option("kinesis.streamName", "kinesis-source") \
.option("kinesis.consumerType", "SubscribeToShard") \
.option("kinesis.consumerName", "efo-consumer") \
.option("kinesis.endpointUrl", "https://kinesis.<aws-region>.amazonaws.com") \
.option("kinesis.startingposition", "LATEST") \
.load()

Deploy the Kinesis Data Generator

Complete the following steps to deploy the KDG and start generating data:

  1. Choose Launch Stack:
    launch stack 1

You might need to change your Region when deploying. Make sure that the KDG is launched in the same Region as where you deployed the solution.

  1. For the parameters Username and Password, enter the values of your choice. Note these values to use later when you log in to the KDG.
  2. When the template has finished deploying, go to the Outputs tab of the stack and locate the KDG URL.
  3. Log in to the KDG, using the credentials you set when launching the CloudFormation template.
  4. Specify your Region and data stream name, and use the following template to generate test data: 
    {
    "id": {{random.number(100)}},
    "data": "{{random.arrayElement(
        ["Spain","Portugal","Finland","France"]
    )}}",
    "date": "{{date.now("YYYY-MM-DD hh:mm:ss")}}"
}

  5. Return to Systems Manager to continue working with the Spark application.
  6. To be able to apply transformations based on the fields of the events, you first need to define the schema for the events: 
    from pyspark.sql.types import *

pythonSchema = StructType() \
 .add("id", LongType()) \
 .add("data", StringType()) \
 .add("date", TimestampType())

  7. Run the following the command to consume data from Kinesis Data Streams: 
    from pyspark.sql.functions import *

events= kinesis \
  .selectExpr("cast (data as STRING) jsonData") \
  .select(from_json("jsonData", pythonSchema).alias("events")) \
  .select("events.*")

  8. Use the following code for the Kinesis Spark connector sink: 
    events \
    .selectExpr("CAST(id AS STRING) as partitionKey","data","date") \
    .writeStream \
    .format("aws-kinesis") \
    .option("kinesis.region", "<aws-region>") \
    .outputMode("append") \
    .option("kinesis.streamName", "kinesis-sink") \
    .option("kinesis.endpointUrl", "https://kinesis.<aws-region>.amazonaws.com") \
    .option("checkpointLocation", "/kinesisCheckpoint") \
    .start() \
    .awaitTermination()

You can view the data in the Kinesis Data Streams console.

  1. On the Kinesis Data Streams console, navigate to kinesis-sink.
  2. On the Data viewer tab, choose a shard and a starting position (for this post, we use Latest) and choose Get records.

You can see the data sent, as shown in the following screenshot. Kinesis Data Streams uses base64 encoding by default, so you might see text with unreadable characters.

Clean up

Delete the following CloudFormation stacks created during this deployment to delete all the provisioned resources:

  • EmrSparkKinesisStack
  • Kinesis-Data-Generator-Cognito-User-SparkEFO-Blog

If you created any additional resources during this deployment, delete them manually.

Conclusion

In this post, we discussed the open source Kinesis Data Streams connector for Spark Structured Streaming. It supports the newer Data Sources API V2 and Spark Structured Streaming for building streaming applications. The connector also enables high-throughput consumption from Kinesis Data Streams with enhanced fan-out by providing dedicated throughput up to 2 Mbps per shard per consumer. With this connector, you can now effortlessly build high-throughput streaming applications with Spark Structured Streaming.

The Kinesis Spark connector is open source under the Apache 2.0 license on GitHub. To get started, visit the GitHub repo.

About the Authors


Idan Maizlits is a Senior Product Manager on the Amazon Kinesis Data Streams team at Amazon Web Services. Idan loves engaging with customers to learn about their challenges with real-time data and to help them achieve their business goals. Outside of work, he enjoys spending time with his family exploring the outdoors and cooking.


Subham Rakshit is a Streaming Specialist Solutions Architect for Analytics at AWS based in the UK. He works with customers to design and build search and streaming data platforms that help them achieve their business objective. Outside of work, he enjoys spending time solving jigsaw puzzles with his daughter.

Francisco Morillo is a Streaming Solutions Architect at AWS. Francisco works with AWS customers helping them design real-time analytics architectures using AWS services, supporting Amazon MSK and AWS’s managed offering for Apache Flink.

Umesh Chaudhari is a Streaming Solutions Architect at AWS. He works with customers to design and build real-time data processing systems. He has extensive working experience in software engineering, including architecting, designing, and developing data analytics systems. Outside of work, he enjoys traveling, reading, and watching movies.

Krones real-time production line monitoring with Amazon Managed Service for Apache Flink

Post Syndicated from Florian Mair original https://aws.amazon.com/blogs/big-data/krones-real-time-production-line-monitoring-with-amazon-managed-service-for-apache-flink/

Krones provides breweries, beverage bottlers, and food producers all over the world with individual machines and complete production lines. Every day, millions of glass bottles, cans, and PET containers run through a Krones line. Production lines are complex systems with lots of possible errors that could stall the line and decrease the production yield. Krones wants to detect the failure as early as possible (sometimes even before it happens) and notify production line operators to increase reliability and output. So how to detect a failure? Krones equips their lines with sensors for data collection, which can then be evaluated against rules. Krones, as the line manufacturer, as well as the line operator have the possibility to create monitoring rules for machines. Therefore, beverage bottlers and other operators can define their own margin of error for the line. In the past, Krones used a system based on a time series database. The main challenges were that this system was hard to debug and also queries represented the current state of machines but not the state transitions.

This post shows how Krones built a streaming solution to monitor their lines, based on Amazon Kinesis and Amazon Managed Service for Apache Flink. These fully managed services reduce the complexity of building streaming applications with Apache Flink. Managed Service for Apache Flink manages the underlying Apache Flink components that provide durable application state, metrics, logs, and more, and Kinesis enables you to cost-effectively process streaming data at any scale. If you want to get started with your own Apache Flink application, check out the GitHub repository for samples using the Java, Python, or SQL APIs of Flink.

Overview of solution

Krones’s line monitoring is part of the Krones Shopfloor Guidance system. It provides support in the organization, prioritization, management, and documentation of all activities in the company. It allows them to notify an operator if the machine is stopped or materials are required, regardless where the operator is in the line. Proven condition monitoring rules are already built-in but can also be user defined via the user interface. For example, if a certain data point that is monitored violates a threshold, there can be a text message or trigger for a maintenance order on the line.

The condition monitoring and rule evaluation system is built on AWS, using AWS analytics services. The following diagram illustrates the architecture.

Architecture Diagram for Krones Production Line Monitoring

Almost every data streaming application consists of five layers: data source, stream ingestion, stream storage, stream processing, and one or more destinations. In the following sections, we dive deeper into each layer and how the line monitoring solution, built by Krones, works in detail.

Data source

The data is gathered by a service running on an edge device reading several protocols like Siemens S7 or OPC/UA. Raw data is preprocessed to create a unified JSON structure, which makes it easier to process later on in the rule engine. A sample payload converted to JSON might look like the following:

{
  "version": 1,
  "timestamp": 1234,
  "equipmentId": "84068f2f-3f39-4b9c-a995-d2a84d878689",
  "tag": "water_temperature",
  "value": 13.45,
  "quality": "Ok",
  "meta": {      
    "sequenceNumber": 123,
    "flags": ["Fst", "Lst", "Wmk", "Syn", "Ats"],
    "createdAt": 12345690,
    "sourceId": "filling_machine"
  }
}

Stream ingestion

AWS IoT Greengrass is an open source Internet of Things (IoT) edge runtime and cloud service. This allows you to act on data locally and aggregate and filter device data. AWS IoT Greengrass provides prebuilt components that can be deployed to the edge. The production line solution uses the stream manager component, which can process data and transfer it to AWS destinations such as AWS IoT Analytics, Amazon Simple Storage Service (Amazon S3), and Kinesis. The stream manager buffers and aggregates records, then sends it to a Kinesis data stream.

Stream storage

The job of the stream storage is to buffer messages in a fault tolerant way and make it available for consumption to one or more consumer applications. To achieve this on AWS, the most common technologies are Kinesis and Amazon Managed Streaming for Apache Kafka (Amazon MSK). For storing our sensor data from production lines, Krones choose Kinesis. Kinesis is a serverless streaming data service that works at any scale with low latency. Shards within a Kinesis data stream are a uniquely identified sequence of data records, where a stream is composed of one or more shards. Each shard has 2 MB/s of read capacity and 1 MB/s write capacity (with max 1,000 records/s). To avoid hitting those limits, data should be distributed among shards as evenly as possible. Every record that is sent to Kinesis has a partition key, which is used to group data into a shard. Therefore, you want to have a large number of partition keys to distribute the load evenly. The stream manager running on AWS IoT Greengrass supports random partition key assignments, which means that all records end up in a random shard and the load is distributed evenly. A disadvantage of random partition key assignments is that records aren’t stored in order in Kinesis. We explain how to solve this in the next section, where we talk about watermarks.

Watermarks

A watermark is a mechanism used to track and measure the progress of event time in a data stream. The event time is the timestamp from when the event was created at the source. The watermark indicates the timely progress of the stream processing application, so all events with an earlier or equal timestamp are considered as processed. This information is essential for Flink to advance event time and trigger relevant computations, such as window evaluations. The allowed lag between event time and watermark can be configured to determine how long to wait for late data before considering a window complete and advancing the watermark.

Krones has systems all around the globe, and needed to handle late arrivals due to connection losses or other network constraints. They started out by monitoring late arrivals and setting the default Flink late handling to the maximum value they saw in this metric. They experienced issues with time synchronization from the edge devices, which lead them to a more sophisticated way of watermarking. They built a global watermark for all the senders and used the lowest value as the watermark. The timestamps are stored in a HashMap for all incoming events. When the watermarks are emitted periodically, the smallest value of this HashMap is used. To avoid stalling of watermarks by missing data, they configured an idleTimeOut parameter, which ignores timestamps that are older than a certain threshold. This increases latency but gives strong data consistency.

public class BucketWatermarkGenerator implements WatermarkGenerator<DataPointEvent> {
private HashMap <String, WatermarkAndTimestamp> lastTimestamps;
private Long idleTimeOut;
private long maxOutOfOrderness;
}

Stream processing

After the data is collected from sensors and ingested into Kinesis, it needs to be evaluated by a rule engine. A rule in this system represents the state of a single metric (such as temperature) or a collection of metrics. To interpret a metric, more than one data point is used, which is a stateful calculation. In this section, we dive deeper into the keyed state and broadcast state in Apache Flink and how they’re used to build the Krones rule engine.

Control stream and broadcast state pattern

In Apache Flink, state refers to the ability of the system to store and manage information persistently across time and operations, enabling the processing of streaming data with support for stateful computations.

The broadcast state pattern allows the distribution of a state to all parallel instances of an operator. Therefore, all operators have the same state and data can be processed using this same state. This read-only data can be ingested by using a control stream. A control stream is a regular data stream, but usually with a much lower data rate. This pattern allows you to dynamically update the state on all operators, enabling the user to change the state and behavior of the application without the need for a redeploy. More precisely, the distribution of the state is done by the use of a control stream. By adding a new record into the control stream, all operators receive this update and are using the new state for the processing of new messages.

This allows users of Krones application to ingest new rules into the Flink application without restarting it. This avoids downtime and gives a great user experience as changes happen in real time. A rule covers a scenario in order to detect a process deviation. Sometimes, the machine data is not as easy to interpret as it might look at first glance. If a temperature sensor is sending high values, this might indicate an error, but also be the effect of an ongoing maintenance procedure. It’s important to put metrics in context and filter some values. This is achieved by a concept called grouping.

Grouping of metrics

The grouping of data and metrics allows you to define the relevance of incoming data and produce accurate results. Let’s walk through the example in the following figure.

Grouping of metrics

In Step 1, we define two condition groups. Group 1 collects the machine state and which product is going through the line. Group 2 uses the value of the temperature and pressure sensors. A condition group can have different states depending on the values it receives. In this example, group 1 receives data that the machine is running, and the one-liter bottle is selected as the product; this gives this group the state ACTIVE. Group 2 has metrics for temperature and pressure; both metrics are above their thresholds for more than 5 minutes. This results in group 2 being in a WARNING state. This means group 1 reports that everything is fine and group 2 does not. In Step 2, weights are added to the groups. This is needed in some situations, because groups might report conflicting information. In this scenario, group 1 reports ACTIVE and group 2 reports WARNING, so it’s not clear to the system what the state of the line is. After adding the weights, the states can be ranked, as shown in step 3. Lastly, the highest ranked state is chosen as the winning one, as shown in Step 4.

After the rules are evaluated and the final machine state is defined, the results will be further processed. The action taken depends on the rule configuration; this can be a notification to the line operator to restock materials, do some maintenance, or just a visual update on the dashboard. This part of the system, which evaluates metrics and rules and takes actions based on the results, is referred to as a rule engine.

Scaling the rule engine

By letting users build their own rules, the rule engine can have a high number of rules that it needs to evaluate, and some rules might use the same sensor data as other rules. Flink is a distributed system that scales very well horizontally. To distribute a data stream to several tasks, you can use the keyBy() method. This allows you to partition a data stream in a logical way and send parts of the data to different task managers. This is often done by choosing an arbitrary key so you get an evenly distributed load. In this case, Krones added a ruleId to the data point and used it as a key. Otherwise, data points that are needed are processed by another task. The keyed data stream can be used across all rules just like a regular variable.

Destinations

When a rule changes its state, the information is sent to a Kinesis stream and then via Amazon EventBridge to consumers. One of the consumers creates a notification from the event that is transmitted to the production line and alerts the personnel to act. To be able to analyze the rule state changes, another service writes the data to an Amazon DynamoDB table for fast access and a TTL is in place to offload long-term history to Amazon S3 for further reporting.

Conclusion

In this post, we showed you how Krones built a real-time production line monitoring system on AWS. Managed Service for Apache Flink allowed the Krones team to get started quickly by focusing on application development rather than infrastructure. The real-time capabilities of Flink enabled Krones to reduce machine downtime by 10% and increase efficiency up to 5%.

If you want to build your own streaming applications, check out the available samples on the GitHub repository. If you want to extend your Flink application with custom connectors, see Making it Easier to Build Connectors with Apache Flink: Introducing the Async Sink. The Async Sink is available in Apache Flink version 1.15.1 and later.

About the Authors

Florian Mair is a Senior Solutions Architect and data streaming expert at AWS. He is a technologist that helps customers in Europe succeed and innovate by solving business challenges using AWS Cloud services. Besides working as a Solutions Architect, Florian is a passionate mountaineer, and has climbed some of the highest mountains across Europe.

Emil Dietl is a Senior Tech Lead at Krones specializing in data engineering, with a key field in Apache Flink and microservices. His work often involves the development and maintenance of mission-critical software. Outside of his professional life, he deeply values spending quality time with his family.

Simon Peyer is a Solutions Architect at AWS based in Switzerland. He is a practical doer and is passionate about connecting technology and people using AWS Cloud services. A special focus for him is data streaming and automations. Besides work, Simon enjoys his family, the outdoors, and hiking in the mountains.

Exploring real-time streaming for generative AI Applications

Post Syndicated from Ali Alemi original https://aws.amazon.com/blogs/big-data/exploring-real-time-streaming-for-generative-ai-applications/

Foundation models (FMs) are large machine learning (ML) models trained on a broad spectrum of unlabeled and generalized datasets. FMs, as the name suggests, provide the foundation to build more specialized downstream applications, and are unique in their adaptability. They can perform a wide range of different tasks, such as natural language processing, classifying images, forecasting trends, analyzing sentiment, and answering questions. This scale and general-purpose adaptability are what makes FMs different from traditional ML models. FMs are multimodal; they work with different data types such as text, video, audio, and images. Large language models (LLMs) are a type of FM and are pre-trained on vast amounts of text data and typically have application uses such as text generation, intelligent chatbots, or summarization.

Streaming data facilitates the constant flow of diverse and up-to-date information, enhancing the models’ ability to adapt and generate more accurate, contextually relevant outputs. This dynamic integration of streaming data enables generative AI applications to respond promptly to changing conditions, improving their adaptability and overall performance in various tasks.

To better understand this, imagine a chatbot that helps travelers book their travel. In this scenario, the chatbot needs real-time access to airline inventory, flight status, hotel inventory, latest price changes, and more. This data usually comes from third parties, and developers need to find a way to ingest this data and process the data changes as they happen.

Batch processing is not the best fit in this scenario. When data changes rapidly, processing it in a batch may result in stale data being used by the chatbot, providing inaccurate information to the customer, which impacts the overall customer experience. Stream processing, however, can enable the chatbot to access real-time data and adapt to changes in availability and price, providing the best guidance to the customer and enhancing the customer experience.

Another example is an AI-driven observability and monitoring solution where FMs monitor real-time internal metrics of a system and produces alerts. When the model finds an anomaly or abnormal metric value, it should immediately produce an alert and notify the operator. However, the value of such important data diminishes significantly over time. These notifications should ideally be received within seconds or even while it’s happening. If operators receive these notifications minutes or hours after they happened, such an insight is not actionable and has potentially lost its value. You can find similar use cases in other industries such as retail, car manufacturing, energy, and the financial industry.

In this post, we discuss why data streaming is a crucial component of generative AI applications due to its real-time nature. We discuss the value of AWS data streaming services such as Amazon Managed Streaming for Apache Kafka (Amazon MSK), Amazon Kinesis Data Streams, Amazon Managed Service for Apache Flink, and Amazon Kinesis Data Firehose in building generative AI applications.

In-context learning

LLMs are trained with point-in-time data and have no inherent ability to access fresh data at inference time. As new data appears, you will have to continuously fine-tune or further train the model. This is not only an expensive operation, but also very limiting in practice because the rate of new data generation far supersedes the speed of fine-tuning. Additionally, LLMs lack contextual understanding and rely solely on their training data, and are therefore prone to hallucinations. This means they can generate a fluent, coherent, and syntactically sound but factually incorrect response. They are also devoid of relevance, personalization, and context.

LLMs, however, have the capacity to learn from the data they receive from the context to more accurately respond without modifying the model weights. This is called in-context learning, and can be used to produce personalized answers or provide an accurate response in the context of organization policies.

For example, in a chatbot, data events could pertain to an inventory of flights and hotels or price changes that are constantly ingested to a streaming storage engine. Furthermore, data events are filtered, enriched, and transformed to a consumable format using a stream processor. The result is made available to the application by querying the latest snapshot. The snapshot constantly updates through stream processing; therefore, the up-to-date data is provided in the context of a user prompt to the model. This allows the model to adapt to the latest changes in price and availability. The following diagram illustrates a basic in-context learning workflow.

A commonly used in-context learning approach is to use a technique called Retrieval Augmented Generation (RAG). In RAG, you provide the relevant information such as most relevant policy and customer records along with the user question to the prompt. This way, the LLM generates an answer to the user question using additional information provided as context. To learn more about RAG, refer to Question answering using Retrieval Augmented Generation with foundation models in Amazon SageMaker JumpStart.

A RAG-based generative AI application can only produce generic responses based on its training data and the relevant documents in the knowledge base. This solution falls short when a near-real-time personalized response is expected from the application. For example, a travel chatbot is expected to consider the user’s current bookings, available hotel and flight inventory, and more. Moreover, the relevant customer personal data (commonly known as the unified customer profile) is usually subject to change. If a batch process is employed to update the generative AI’s user profile database, the customer may receive dissatisfying responses based on old data.

In this post, we discuss the application of stream processing to enhance a RAG solution used for building question answering agents with context from real-time access to unified customer profiles and organizational knowledge base.

Near-real-time customer profile updates

Customer records are typically distributed across data stores within an organization. For your generative AI application to provide a relevant, accurate, and up-to-date customer profile, it is vital to build streaming data pipelines that can perform identity resolution and profile aggregation across the distributed data stores. Streaming jobs constantly ingest new data to synchronize across systems and can perform enrichment, transformations, joins, and aggregations across windows of time more efficiently. Change data capture (CDC) events contain information about the source record, updates, and metadata such as time, source, classification (insert, update, or delete), and the initiator of the change.

The following diagram illustrates an example workflow for CDC streaming ingestion and processing for unified customer profiles.

In this section, we discuss the main components of a CDC streaming pattern required to support RAG-based generative AI applications.

CDC streaming ingestion

A CDC replicator is a process that collects data changes from a source system (usually by reading transaction logs or binlogs) and writes CDC events with the exact same order they occurred in a streaming data stream or topic. This involves a log-based capture with tools such as AWS Database Migration Service (AWS DMS) or open source connectors such as Debezium for Apache Kafka connect. Apache Kafka Connect is part of the Apache Kafka environment, allowing data to be ingested from various sources and delivered to variety of destinations. You can run your Apache Kafka connector on Amazon MSK Connect within minutes without worrying about configuration, setup, and operating an Apache Kafka cluster. You only need to upload your connector’s compiled code to Amazon Simple Storage Service (Amazon S3) and set up your connector with your workload’s specific configuration.

There are also other methods for capturing data changes. For example, Amazon DynamoDB provides a feature for streaming CDC data to Amazon DynamoDB Streams or Kinesis Data Streams. Amazon S3 provides a trigger to invoke an AWS Lambda function when a new document is stored.

Streaming storage

Streaming storage functions as an intermediate buffer to store CDC events before they get processed. Streaming storage provides reliable storage for streaming data. By design, it is highly available and resilient to hardware or node failures and maintains the order of the events as they are written. Streaming storage can store data events either permanently or for a set period of time. This allows stream processors to read from part of the stream if there is a failure or a need for re-processing. Kinesis Data Streams is a serverless streaming data service that makes it straightforward to capture, process, and store data streams at scale. Amazon MSK is a fully managed, highly available, and secure service provided by AWS for running Apache Kafka.

Stream processing

Stream processing systems should be designed for parallelism to handle high data throughput. They should partition the input stream between multiple tasks running on multiple compute nodes. Tasks should be able to send the result of one operation to the next one over the network, making it possible for processing data in parallel while performing operations such as joins, filtering, enrichment, and aggregations. Stream processing applications should be able to process events with regards to the event time for use cases where events could arrive late or correct computation relies on the time events occur rather than the system time. For more information, refer to Notions of Time: Event Time and Processing Time.

Stream processes continuously produce results in the form of data events that need to be output to a target system. A target system could be any system that can integrate directly with the process or via streaming storage as in intermediary. Depending on the framework you choose for stream processing, you will have different options for target systems depending on available sink connectors. If you decide to write the results to an intermediary streaming storage, you can build a separate process that reads events and applies changes to the target system, such as running an Apache Kafka sink connector. Regardless of which option you choose, CDC data needs extra handling due to its nature. Because CDC events carry information about updates or deletes, it’s important that they merge in the target system in the right order. If changes are applied in the wrong order, the target system will be out of sync with its source.

Apache Flink is a powerful stream processing framework known for its low latency and high throughput capabilities. It supports event time processing, exactly-once processing semantics, and high fault tolerance. Additionally, it provides native support for CDC data via a special structure called dynamic tables. Dynamic tables mimic the source database tables and provide a columnar representation of the streaming data. The data in dynamic tables changes with every event that is processed. New records can be appended, updated, or deleted at any time. Dynamic tables abstract away the extra logic you need to implement for each record operation (insert, update, delete) separately. For more information, refer to Dynamic Tables.

With Amazon Managed Service for Apache Flink, you can run Apache Flink jobs and integrate with other AWS services. There are no servers and clusters to manage, and there is no compute and storage infrastructure to set up.

AWS Glue is a fully managed extract, transform, and load (ETL) service, which means AWS handles the infrastructure provisioning, scaling, and maintenance for you. Although it’s primarily known for its ETL capabilities, AWS Glue can also be used for Spark streaming applications. AWS Glue can interact with streaming data services such as Kinesis Data Streams and Amazon MSK for processing and transforming CDC data. AWS Glue can also seamlessly integrate with other AWS services such as Lambda, AWS Step Functions, and DynamoDB, providing you with a comprehensive ecosystem for building and managing data processing pipelines.

Unified customer profile

Overcoming the unification of the customer profile across a variety of source systems requires the development of robust data pipelines. You need data pipelines that can bring and synchronize all records into one data store. This data store provides your organization with the holistic customer records view that is needed for operational efficiency of RAG-based generative AI applications. For building such a data store, an unstructured data store would be best.

An identity graph is a useful structure for creating a unified customer profile because it consolidates and integrates customer data from various sources, ensures data accuracy and deduplication, offers real-time updates, connects cross-systems insights, enables personalization, enhances customer experience, and supports regulatory compliance. This unified customer profile empowers the generative AI application to understand and engage with customers effectively, and adhere to data privacy regulations, ultimately enhancing customer experiences and driving business growth. You can build your identity graph solution using Amazon Neptune, a fast, reliable, fully managed graph database service.

AWS provides a few other managed and serverless NoSQL storage service offerings for unstructured key-value objects. Amazon DocumentDB (with MongoDB compatibility) is a fast, scalable, highly available, and fully managed enterprise document database service that supports native JSON workloads. DynamoDB is a fully managed NoSQL database service that provides fast and predictable performance with seamless scalability.

Near-real-time organizational knowledge base updates

Similar to customer records, internal knowledge repositories such as company policies and organizational documents are siloed across storage systems. This is typically unstructured data and is updated in a non-incremental fashion. The use of unstructured data for AI applications is effective using vector embeddings, which is a technique of representing high dimensional data such as text files, images, and audio files as multi-dimensional numeric.

AWS provides several vector engine services, such as Amazon OpenSearch Serverless, Amazon Kendra, and Amazon Aurora PostgreSQL-Compatible Edition with the pgvector extension for storing vector embeddings. Generative AI applications can enhance the user experience by transforming the user prompt into a vector and use it to query the vector engine to retrieve contextually relevant information. Both the prompt and the vector data retrieved are then passed to the LLM to receive a more precise and personalized response.

The following diagram illustrates an example stream-processing workflow for vector embeddings.

Knowledge base contents need to be converted to vector embeddings before being written to the vector data store. Amazon Bedrock or Amazon SageMaker can help you access the model of your choice and expose a private endpoint for this conversion. Furthermore, you can use libraries such as LangChain to integrate with these endpoints. Building a batch process can help you convert your knowledge base content to vector data and store it in a vector database initially. However, you need to rely on an interval to reprocess the documents to synchronize your vector database with changes in your knowledge base content. With a large number of documents, this process can be inefficient. Between these intervals, your generative AI application users will receive answers according to the old content, or will receive an inaccurate answer because the new content is not vectorized yet.

Stream processing is an ideal solution for these challenges. It produces events as per existing documents initially and further monitors the source system and creates a document change event as soon as they occur. These events can be stored in streaming storage and wait to be processed by a streaming job. A streaming job reads these events, loads the content of the document, and transforms the contents to an array of related tokens of words. Each token further transforms into vector data via an API call to an embedding FM. Results are sent for storage to the vector storage via a sink operator.

If you’re using Amazon S3 for storing your documents, you can build an event-source architecture based on S3 object change triggers for Lambda. A Lambda function can create an event in the desired format and write that to your streaming storage.

You can also use Apache Flink to run as a streaming job. Apache Flink provides the native FileSystem source connector, which can discover existing files and read their contents initially. After that, it can continuously monitor your file system for new files and capture their content. The connector supports reading a set of files from distributed file systems such as Amazon S3 or HDFS with a format of plain text, Avro, CSV, Parquet, and more, and produces a streaming record. As a fully managed service, Managed Service for Apache Flink removes the operational overhead of deploying and maintaining Flink jobs, allowing you to focus on building and scaling your streaming applications. With seamless integration into the AWS streaming services such as Amazon MSK or Kinesis Data Streams, it provides features like automatic scaling, security, and resiliency, providing reliable and efficient Flink applications for handling real-time streaming data.

Based on your DevOps preference, you can choose between Kinesis Data Streams or Amazon MSK for storing the streaming records. Kinesis Data Streams simplifies the complexities of building and managing custom streaming data applications, allowing you to focus on deriving insights from your data rather than infrastructure maintenance. Customers using Apache Kafka often opt for Amazon MSK due to its straightforwardness, scalability, and dependability in overseeing Apache Kafka clusters within the AWS environment. As a fully managed service, Amazon MSK takes on the operational complexities associated with deploying and maintaining Apache Kafka clusters, enabling you to concentrate on constructing and expanding your streaming applications.

Because a RESTful API integration suits the nature of this process, you need a framework that supports a stateful enrichment pattern via RESTful API calls to track for failures and retry for the failed request. Apache Flink again is a framework that can do stateful operations in at-memory speed. To understand the best ways to make API calls via Apache Flink, refer to Common streaming data enrichment patterns in Amazon Kinesis Data Analytics for Apache Flink.

Apache Flink provides native sink connectors for writing data to vector datastores such as Amazon Aurora for PostgreSQL with pgvector or Amazon OpenSearch Service with VectorDB. Alternatively, you can stage the Flink job’s output (vectorized data) in an MSK topic or a Kinesis data stream. OpenSearch Service provides support for native ingestion from Kinesis data streams or MSK topics. For more information, refer to Introducing Amazon MSK as a source for Amazon OpenSearch Ingestion and Loading streaming data from Amazon Kinesis Data Streams.

Feedback analytics and fine-tuning

It’s important for data operation managers and AI/ML developers to get insight about the performance of the generative AI application and the FMs in use. To achieve that, you need to build data pipelines that calculate important key performance indicator (KPI) data based on the user feedback and variety of application logs and metrics. This information is useful for stakeholders to gain real-time insight about the performance of the FM, the application, and overall user satisfaction about the quality of support they receive from your application. You also need to collect and store the conversation history for further fine-tuning your FMs to improve their ability in performing domain-specific tasks.

This use case fits very well in the streaming analytics domain. Your application should store each conversation in streaming storage. Your application can prompt users about their rating of each answer’s accuracy and their overall satisfaction. This data can be in a format of a binary choice or a free form text. This data can be stored in a Kinesis data stream or MSK topic, and get processed to generate KPIs in real time. You can put FMs to work for users’ sentiment analysis. FMs can analyze each answer and assign a category of user satisfaction.

Apache Flink’s architecture allows for complex data aggregation over windows of time. It also provides support for SQL querying over stream of data events. Therefore, by using Apache Flink, you can quickly analyze raw user inputs and generate KPIs in real time by writing familiar SQL queries. For more information, refer to Table API & SQL.

With Amazon Managed Service for Apache Flink Studio, you can build and run Apache Flink stream processing applications using standard SQL, Python, and Scala in an interactive notebook. Studio notebooks are powered by Apache Zeppelin and use Apache Flink as the stream processing engine. Studio notebooks seamlessly combine these technologies to make advanced analytics on data streams accessible to developers of all skill sets. With support for user-defined functions (UDFs), Apache Flink allows for building custom operators to integrate with external resources such as FMs for performing complex tasks such as sentiment analysis. You can use UDFs to compute various metrics or enrich user feedback raw data with additional insights such as user sentiment. To learn more about this pattern, refer to Proactively addressing customer concern in real-time with GenAI, Flink, Apache Kafka, and Kinesis.

With Managed Service for Apache Flink Studio, you can deploy your Studio notebook as a streaming job with one click. You can use native sink connectors provided by Apache Flink to send the output to your storage of choice or stage it in a Kinesis data stream or MSK topic. Amazon Redshift and OpenSearch Service are both ideal for storing analytical data. Both engines provide native ingestion support from Kinesis Data Streams and Amazon MSK via a separate streaming pipeline to a data lake or data warehouse for analysis.

Amazon Redshift uses SQL to analyze structured and semi-structured data across data warehouses and data lakes, using AWS-designed hardware and machine learning to deliver the best price-performance at scale. OpenSearch Service offers visualization capabilities powered by OpenSearch Dashboards and Kibana (1.5 to 7.10 versions).

You can use the outcome of such analysis combined with user prompt data for fine-tuning the FM when is needed. SageMaker is the most straightforward way to fine-tune your FMs. Using Amazon S3 with SageMaker provides a powerful and seamless integration for fine-tuning your models. Amazon S3 serves as a scalable and durable object storage solution, enabling straightforward storage and retrieval of large datasets, training data, and model artifacts. SageMaker is a fully managed ML service that simplifies the entire ML lifecycle. By using Amazon S3 as the storage backend for SageMaker, you can benefit from the scalability, reliability, and cost-effectiveness of Amazon S3, while seamlessly integrating it with SageMaker training and deployment capabilities. This combination enables efficient data management, facilitates collaborative model development, and makes sure that ML workflows are streamlined and scalable, ultimately enhancing the overall agility and performance of the ML process. For more information, refer to Fine-tune Falcon 7B and other LLMs on Amazon SageMaker with @remote decorator.

With a file system sink connector, Apache Flink jobs can deliver data to Amazon S3 in open format (such as JSON, Avro, Parquet, and more) files as data objects. If you prefer to manage your data lake using a transactional data lake framework (such as Apache Hudi, Apache Iceberg, or Delta Lake), all of these frameworks provide a custom connector for Apache Flink. For more details, refer to Create a low-latency source-to-data lake pipeline using Amazon MSK Connect, Apache Flink, and Apache Hudi.

Summary

For a generative AI application based on a RAG model, you need to consider building two data storage systems, and you need to build data operations that keep them up to date with all the source systems. Traditional batch jobs are not sufficient to process the size and diversity of the data you need to integrate with your generative AI application. Delays in processing the changes in source systems result in an inaccurate response and reduce the efficiency of your generative AI application. Data streaming enables you to ingest data from a variety of databases across various systems. It also allows you to transform, enrich, join, and aggregate data across many sources efficiently in near-real time. Data streaming provides a simplified data architecture to collect and transform users’ real-time reactions or comments on the application responses, helping you deliver and store the results in a data lake for model fine-tuning. Data streaming also helps you optimize data pipelines by processing only the change events, allowing you to respond to data changes more quickly and efficiently.

Learn more about AWS data streaming services and get started building your own data streaming solution.

About the Authors

Ali Alemi is a Streaming Specialist Solutions Architect at AWS. Ali advises AWS customers with architectural best practices and helps them design real-time analytics data systems which are reliable, secure, efficient, and cost-effective. He works backward from customer’s use cases and designs data solutions to solve their business problems. Prior to joining AWS, Ali supported several public sector customers and AWS consulting partners in their application modernization journey and migration to the Cloud.

Imtiaz (Taz) Sayed is the World-Wide Tech Leader for Analytics at AWS. He enjoys engaging with the community on all things data and analytics. He can be reached via LinkedIn.

Invoke AWS Lambda functions from cross-account Amazon Kinesis Data Streams

Post Syndicated from Amar Surjit original https://aws.amazon.com/blogs/big-data/invoke-aws-lambda-functions-from-cross-account-amazon-kinesis-data-streams/

A multi-account architecture on AWS is essential for enhancing security, compliance, and resource management by isolating workloads, enabling granular cost allocation, and facilitating collaboration across distinct environments. It also mitigates risks, improves scalability, and allows for advanced networking configurations.

In a streaming architecture, you may have event producers, stream storage, and event consumers in a single account or spread across different accounts depending on your business and IT requirements. For example, your company may want to centralize its clickstream data or log data from multiple different producers across different accounts. Data consumers from marketing, product engineering, or analytics require access to the same streaming data across accounts, which requires the ability to deliver a multi-account streaming architecture.

To build a multi-account streaming architecture, you can use Amazon Kinesis Data Streams as the stream storage and AWS Lambda as the event consumer. Amazon Kinesis Data Streams enables real-time processing of streaming data at scale. When integrated with Lambda, it allows for serverless data processing, enabling you to analyze and react to data streams in real time without managing infrastructure. This integration supports various use cases, including real-time analytics, log processing, Internet of Things (IoT) data ingestion, and more, making it valuable for businesses requiring timely insights from their streaming data. In this post, we demonstrate how you can process data ingested into a stream in one account with a Lambda function in another account.

The recent launch of Kinesis Data Streams support for resource-based policies enables invoking a Lambda from another account. With a resource-based policy, you can specify AWS accounts, AWS Identity and Access Management (IAM) users, or IAM roles and the exact Kinesis Data Streams actions for which you want to grant access. After access is granted, you can configure a Lambda function in another account to start processing the data stream belonging to your account. This reduces cost and simplifies the data processing pipeline, because you no longer have to copy streaming data using Lambda functions in both accounts. Sharing access to your data streams or registered consumers does not incur additional charges to your account. Cross-account usage of Kinesis Data Streams resources will continue to be billed to the resource owners.

In this post, we use Kinesis Data Streams with enhanced fan-out feature, empowering consumers with dedicated read throughput tailored to their applications. By default, Kinesis Data Streams offers shared read throughput of 2 MB/sec per shard across consumers, but with enhanced fan-out, each consumer can enjoy dedicated throughput of 2 MB/sec per shard. This flexibility allows you to seamlessly adapt Kinesis Data Streams to your specific requirements, choosing between enhanced fan-out for dedicated throughput or shared throughput according to your needs.

Solution overview

For our solution, we deploy Kinesis Data Streams in Account 1 and Lambda as the consumer in Account 2 to receive data from the data stream. The following diagram illustrates the high-level architecture.

Amazon KDS-Lambda cross acct solution architecture

The setup requires the following key elements:

  • Kinesis data stream in Account 1 and Lambda function in Account 2
  • Kinesis Data Streams resource policies in Account 1, allowing a cross-account Lambda execution role to perform operations on the Kinesis data stream
  • A Lambda execution role in Account 2 and an enhanced fan-out consumer resource policy in Account 1, allowing the cross-account Lambda execution role to perform operations on the Kinesis data stream

For the setup, you use three AWS CloudFormation templates to create the key resources:

  • CloudFormation template 1 creates the following key resources in Account 1:
    • Kinesis data stream
    • Kinesis data stream enhanced fan-out consumer
  • CloudFormation template 2 creates the following key resources in Account 2:
    • Consumer Lambda function
    • Consumer Lambda function execution role
  • CloudFormation template 3 creates the following resource in Account 2:
    • Consumer Lambda function event source mapping

The solution supports single-Region deployment, and the CloudFormation templates must be deployed in the same Region across different AWS accounts. In this solution, we use Kinesis Data Streams enhanced fan-out, which is a best practice for deploying architectures requiring large throughput across multiple consumers. Complete the steps in the following sections to deploy this solution.

Prerequisites

You should have two AWS accounts and the required permissions to run a CloudFormation template to create the services mentioned in the solution architecture. You also need the AWS Command Line Interface (AWS CLI) installed, version 2.15 and above.

Launch CloudFormation template 1

Complete the following steps to launch the first CloudFormation template:

  1. Sign in to the AWS Management Console as Account 1 and select the appropriate AWS Region.
  2. Download and launch CloudFormation template 1 where you want to deploy your Kinesis data stream.
  3. For LambdaConsumerAccountId, enter your Lambda consumer account ID and click submit. The CloudFormation template deployment will take a few minutes to complete.
  4. When the stack is complete, on the AWS CloudFormation console, navigate to the stack Outputs tab and copy the values of following parameters:
    • KinesisStreamArn
    • KinesisStreamEFOConsumerArn
    • KMSKeyArn

You will need these values in later steps.

Launch CloudFormation template 2

Complete the following steps to launch the second CloudFormation template:

  1. Sign in to the console as Account 2 and select the appropriate Region.
  2. Download and launch CloudFormation template 2 where you want to host the Lambda consumer.
  3. Provide the following input parameters captured from the previous step:
    • KinesisStreamArn
    • KinesisStreamEFOConsumerArn
    • KMSKeyArn

The CloudFormation template creates the following key resources:

  • Lambda consumer
  • Lambda execution role

The Lambda function’s execution role is an IAM role that grants the function permission to access AWS services and resources. Here, you create a Lambda execution role that has the required Kinesis Data Streams and Lambda invocation permissions.

The CloudFormation template deployment will take a few minutes to complete.

  1. When the stack is complete, on the AWS CloudFormation console, navigate to the stack Outputs tab and copy the values of following parameters:
    • KinesisStreamCreateResourcePolicyCommand
    • KinesisStreamEFOConsumerCreateResourcePolicyCommand
  2. Run the following AWS CLI commands in Account 1 using AWS CloudShell. We recommend using CloudShell because it will have the latest version of the AWS CLI and avoid any kind of failures.
    • KinesisStreamCreateResourcePolicyCommand – This creates the resource policy in Account 1 for Kinesis Data Stream. The following is a sample resource policy: 
      {
"Version": "2012-10-17",
"Statement": [
{
"Sid": "StreamEFOReadStatementID",
"Effect": "Allow",
"Principal": {
"AWS": [
"arn:aws:iam::<AWS Lambda - Consumer account id>:role/kds-cross-account-stream-consumer-lambda-execution-role"
]
},
"Action": [
"kinesis:DescribeStreamSummary",
"kinesis:ListShards",
"kinesis:DescribeStream",
"kinesis:GetRecords",
"kinesis:GetShardIterator"
],
"Resource": "arn:aws:kinesis:<region id>:<Account 1 - Amazon KDS account id>:stream/kds-cross-account-stream"
}
]
}

    • KinesisStreamEFOConsumerCreateResourcePolicyCommand – This creates the resource policy for the enhanced fan-out consumer for the Kinesis data stream in Account 1. The following is a sample resource policy: 
      {
"Version": "2012-10-17",
"Statement": [
{
"Sid": "ConsumerEFOReadStatementID",
"Effect": "Allow",
"Principal": {
"AWS": [
" arn:aws:iam::<AWS Lambda - Consumer account id>:role/kds-cross-account-stream-consumer-lambda-execution-role"
]
},
"Action": [
"kinesis:DescribeStreamConsumer",
"kinesis:SubscribeToShard"
],
"Resource": "arn:aws:kinesis:<region id>:<Account 1 - Amazon KDS account id>:stream/kds-cross-account-stream/consumer/kds-cross-account-stream-efo-consumer:1706616477"
}
]
}

You can also access this policy on the Kinesis Data Streams console, under Enhanced fan-out, Consumer name, and Consumer sharing resource-based policy.

Launch CloudFormation template 3

Now that you have created resource policies in Account 1 for the Kinesis data stream and its enhanced fan-out consumer, you can create Lambda event source mapping for the consumer Lambda function in Account 2. Complete the following steps:

  1. Sign in to the console as Account 2 and select the appropriate Region.
  2. Download and launch CloudFormation template 3 to update the stack you created using CloudFormation template 2.

The CloudFormation template creates the Lambda event source mapping.

Validate the solution

At this point, the deployment is complete. A Kinesis data stream is available to consume the messages and a Lambda function receives these messages in the destination account. To send sample messages to the data stream in Account 1, run the following AWS CLI command using CloudShell:

aws kinesis put-record --stream-name kds-cross-account-stream --data sampledatarecord --partition-key samplepartitionkey3 --region <region id>

The Lambda function in Account 2 is able to receive the messages, and you should be able to verify the same using Amazon CloudWatch logs:

  1. On the CloudWatch console, choose Log groups in the navigation pane.
  2. Locate the log group /aws/lambda/kds-cross-account-stream-efo-consumer.
  3. Choose Search log group to view the relevant log messages. The following is an example message: 
    "Records": [
{
"kinesis": {
"kinesisSchemaVersion": "1.0",
"partitionKey": "samplepartitionkey3",
"sequenceNumber": "49648798411111169765201534322676841348246990356337393698",
"data": "sampledatarecord",
"approximateArrivalTimestamp": 1706623274.658
},

Clean up

It’s always a good practice to clean up all the resources you created as part of this post to avoid any additional cost.

To clean up your resources, delete the respective CloudFormation stacks from Accounts 1 and 2, and stop the producer from pushing events to the Kinesis data stream. This makes sure that you are not charged unnecessarily.

Summary

In this post, we demonstrated how to configure a cross-account Lambda integration with Kinesis Data Streams using AWS resource-based policies. This enables processing of data ingested into a stream within one AWS account through a Lambda function located in another account. To support customers who use a Kinesis data stream in their central account and have multiple consumers reading data from it, we have used the Kinesis Data Streams enhanced fan-out feature.

To get started, open the Kinesis Data Streams console or use the new API PutResourcePolicy to attach a resource policy to your data stream or consumer.

About the authors

Pratik Patel is Sr. Technical Account Manager and streaming analytics specialist. He works with AWS customers and provides ongoing support and technical guidance to help plan and build solutions using best practices and proactively keep customers’ AWS environments operationally healthy.

Amar is a Senior Solutions Architect at Amazon AWS in the UK. He works across power, utilities, manufacturing and automotive customers on strategic implementations, specializing in using AWS Streaming and advanced data analytics solutions, to drive optimal business outcomes.

Architectural patterns for real-time analytics using Amazon Kinesis Data Streams, part 1

Post Syndicated from Raghavarao Sodabathina original https://aws.amazon.com/blogs/big-data/architectural-patterns-for-real-time-analytics-using-amazon-kinesis-data-streams-part-1/

We’re living in the age of real-time data and insights, driven by low-latency data streaming applications. Today, everyone expects a personalized experience in any application, and organizations are constantly innovating to increase their speed of business operation and decision making. The volume of time-sensitive data produced is increasing rapidly, with different formats of data being introduced across new businesses and customer use cases. Therefore, it is critical for organizations to embrace a low-latency, scalable, and reliable data streaming infrastructure to deliver real-time business applications and better customer experiences.

This is the first post to a blog series that offers common architectural patterns in building real-time data streaming infrastructures using Kinesis Data Streams for a wide range of use cases. It aims to provide a framework to create low-latency streaming applications on the AWS Cloud using Amazon Kinesis Data Streams and AWS purpose-built data analytics services.

In this post, we will review the common architectural patterns of two use cases: Time Series Data Analysis and Event Driven Microservices. In the subsequent post in our series, we will explore the architectural patterns in building streaming pipelines for real-time BI dashboards, contact center agent, ledger data, personalized real-time recommendation, log analytics, IoT data, Change Data Capture, and real-time marketing data. All these architecture patterns are integrated with Amazon Kinesis Data Streams.

Real-time streaming with Kinesis Data Streams

Amazon Kinesis Data Streams is a cloud-native, serverless streaming data service that makes it easy to capture, process, and store real-time data at any scale. With Kinesis Data Streams, you can collect and process hundreds of gigabytes of data per second from hundreds of thousands of sources, allowing you to easily write applications that process information in real-time. The collected data is available in milliseconds to allow real-time analytics use cases, such as real-time dashboards, real-time anomaly detection, and dynamic pricing. By default, the data within the Kinesis Data Stream is stored for 24 hours with an option to increase the data retention to 365 days. If customers want to process the same data in real-time with multiple applications, then they can use the Enhanced Fan-Out (EFO) feature. Prior to this feature, every application consuming data from the stream shared the 2MB/second/shard output. By configuring stream consumers to use enhanced fan-out, each data consumer receives dedicated 2MB/second pipe of read throughput per shard to further reduce the latency in data retrieval.

For high availability and durability, Kinesis Data Streams achieves high durability by synchronously replicating the streamed data across three Availability Zones in an AWS Region and gives you the option to retain data for up to 365 days. For security, Kinesis Data Streams provide server-side encryption so you can meet strict data management requirements by encrypting your data at rest and Amazon Virtual Private Cloud (VPC) interface endpoints to keep traffic between your Amazon VPC and Kinesis Data Streams private.

Kinesis Data Streams has native integrations with other AWS services such as AWS Glue and Amazon EventBridge to build real-time streaming applications on AWS. Refer to Amazon Kinesis Data Streams integrations for additional details.

Modern data streaming architecture with Kinesis Data Streams

A modern streaming data architecture with Kinesis Data Streams can be designed as a stack of five logical layers; each layer is composed of multiple purpose-built components that address specific requirements, as illustrated in the following diagram:

The architecture consists of the following key components:

  • Streaming sources – Your source of streaming data includes data sources like clickstream data, sensors, social media, Internet of Things (IoT) devices, log files generated by using your web and mobile applications, and mobile devices that generate semi-structured and unstructured data as continuous streams at high velocity.
  • Stream ingestion – The stream ingestion layer is responsible for ingesting data into the stream storage layer. It provides the ability to collect data from tens of thousands of data sources and ingest in real time. You can use the Kinesis SDK for ingesting streaming data through APIs, the Kinesis Producer Library for building high-performance and long-running streaming producers, or a Kinesis agent for collecting a set of files and ingesting them into Kinesis Data Streams. In addition, you can use many pre-build integrations such as AWS Database Migration Service (AWS DMS), Amazon DynamoDB, and AWS IoT Core to ingest data in a no-code fashion. You can also ingest data from third-party platforms such as Apache Spark and Apache Kafka Connect
  • Stream storage – Kinesis Data Streams offer two modes to support the data throughput: On-Demand and Provisioned. On-Demand mode, now the default choice, can elastically scale to absorb variable throughputs, so that customers do not need to worry about capacity management and pay by data throughput. The On-Demand mode automatically scales up 2x the stream capacity over its historic maximum data ingestion to provide sufficient capacity for unexpected spikes in data ingestion. Alternatively, customers who want granular control over stream resources can use the Provisioned mode and proactively scale up and down the number of Shards to meet their throughput requirements. Additionally, Kinesis Data Streams can store streaming data up to 24 hours by default, but can extend to 7 days or 365 days depending upon use cases. Multiple applications can consume the same stream.
  • Stream processing – The stream processing layer is responsible for transforming data into a consumable state through data validation, cleanup, normalization, transformation, and enrichment. The streaming records are read in the order they are produced, allowing for real-time analytics, building event-driven applications or streaming ETL (extract, transform, and load). You can use Amazon Managed Service for Apache Flink for complex stream data processing, AWS Lambda for stateless stream data processing, and AWS Glue & Amazon EMR for near-real-time compute. You can also build customized consumer applications with Kinesis Consumer Library, which will take care of many complex tasks associated with distributed computing.
  • Destination – The destination layer is like a purpose-built destination depending on your use case. You can stream data directly to Amazon Redshift for data warehousing and Amazon EventBridge for building event-driven applications. You can also use Amazon Kinesis Data Firehose for streaming integration where you can light stream processing with AWS Lambda, and then deliver processed streaming into destinations like Amazon S3 data lake, OpenSearch Service for operational analytics, a Redshift data warehouse, No-SQL databases like Amazon DynamoDB, and relational databases like Amazon RDS to consume real-time streams into business applications. The destination can be an event-driven application for real-time dashboards, automatic decisions based on processed streaming data, real-time altering, and more.

Real-time analytics architecture for time series

Time series data is a sequence of data points recorded over a time interval for measuring events that change over time. Examples are stock prices over time, webpage clickstreams, and device logs over time. Customers can use time series data to monitor changes over time, so that they can detect anomalies, identify patterns, and analyze how certain variables are influenced over time. Time series data is typically generated from multiple sources in high volumes, and it needs to be cost-effectively collected in near real time.

Typically, there are three primary goals that customers want to achieve in processing time-series data:

  • Gain insights real-time into system performance and detect anomalies
  • Understand end-user behavior to track trends and query/build visualizations from these insights
  • Have a durable storage solution to ingest and store both archival and frequently accessed data.

With Kinesis Data Streams, customers can continuously capture terabytes of time series data from thousands of sources for cleaning, enrichment, storage, analysis, and visualization.

The following architecture pattern illustrates how real time analytics can be achieved for Time Series data with Kinesis Data Streams:

Build a serverless streaming data pipeline for time series data

The workflow steps are as follows:

  1. Data Ingestion & Storage – Kinesis Data Streams can continuously capture and store terabytes of data from thousands of sources.
  2. Stream Processing – An application created with Amazon Managed Service for Apache Flink can read the records from the data stream to detect and clean any errors in the time series data and enrich the data with specific metadata to optimize operational analytics. Using a data stream in the middle provides the advantage of using the time series data in other processes and solutions at the same time. A Lambda function is then invoked with these events, and can perform time series calculations in memory.
  3. Destinations – After cleaning and enrichment, the processed time series data can be streamed to Amazon Timestream database for real-time dashboarding and analysis, or stored in databases such as DynamoDB for end-user query. The raw data can be streamed to Amazon S3 for archiving.
  4. Visualization & Gain insights – Customers can query, visualize, and create alerts using Amazon Managed Service for Grafana. Grafana supports data sources that are storage backends for time series data. To access your data from Timestream, you need to install the Timestream plugin for Grafana. End-users can query data from the DynamoDB table with Amazon API Gateway acting as a proxy.

Refer to Near Real-Time Processing with Amazon Kinesis, Amazon Timestream, and Grafana showcasing a serverless streaming pipeline to process and store device telemetry IoT data into a time series optimized data store such as Amazon Timestream.

Enriching & replaying data in real time for event-sourcing microservices

Microservices are an architectural and organizational approach to software development where software is composed of small independent services that communicate over well-defined APIs. When building event-driven microservices, customers want to achieve 1. high scalability to handle the volume of incoming events and 2. reliability of event processing and maintain system functionality in the face of failures.

Customers utilize microservice architecture patterns to accelerate innovation and time-to-market for new features, because it makes applications easier to scale and faster to develop. However, it is challenging to enrich and replay the data in a network call to another microservice because it can impact the reliability of the application and make it difficult to debug and trace errors. To solve this problem, event-sourcing is an effective design pattern that centralizes historic records of all state changes for enrichment and replay, and decouples read from write workloads. Customers can use Kinesis Data Streams as the centralized event store for event-sourcing microservices, because KDS can 1/ handle gigabytes of data throughput per second per stream and stream the data in milliseconds, to meet the requirement on high scalability and near real-time latency, 2/ integrate with Flink and S3 for data enrichment and achieving while being completely decoupled from the microservices, and 3/ allow retry and asynchronous read in a later time, because KDS retains the data record for a default of 24 hours, and optionally up to 365 days.

The following architectural pattern is a generic illustration of how Kinesis Data Streams can be used for Event-Sourcing Microservices:

The steps in the workflow are as follows:

  1. Data Ingestion and Storage – You can aggregate the input from your microservices to your Kinesis Data Streams for storage.
  2. Stream processing – Apache Flink Stateful Functions simplifies building distributed stateful event-driven applications. It can receive the events from an input Kinesis data stream and route the resulting stream to an output data stream. You can create a stateful functions cluster with Apache Flink based on your application business logic.
  3. State snapshot in Amazon S3 – You can store the state snapshot in Amazon S3 for tracking.
  4. Output streams – The output streams can be consumed through Lambda remote functions through HTTP/gRPC protocol through API Gateway.
  5. Lambda remote functions – Lambda functions can act as microservices for various application and business logic to serve business applications and mobile apps.

To learn how other customers built their event-based microservices with Kinesis Data Streams, refer to the following:

Key considerations and best practices

The following are considerations and best practices to keep in mind:

  • Data discovery should be your first step in building modern data streaming applications. You must define the business value and then identify your streaming data sources and user personas to achieve the desired business outcomes.
  • Choose your streaming data ingestion tool based on your steaming data source. For example, you can use the Kinesis SDK for ingesting streaming data through APIs, the Kinesis Producer Library for building high-performance and long-running streaming producers, a Kinesis agent for collecting a set of files and ingesting them into Kinesis Data Streams, AWS DMS for CDC streaming use cases, and AWS IoT Core for ingesting IoT device data into Kinesis Data Streams. You can ingest streaming data directly into Amazon Redshift to build low-latency streaming applications. You can also use third-party libraries like Apache Spark and Apache Kafka to ingest streaming data into Kinesis Data Streams.
  • You need to choose your streaming data processing services based on your specific use case and business requirements. For example, you can use Amazon Kinesis Managed Service for Apache Flink for advanced streaming use cases with multiple streaming destinations and complex stateful stream processing or if you want to monitor business metrics in real time (such as every hour). Lambda is good for event-based and stateless processing. You can use Amazon EMR for streaming data processing to use your favorite open source big data frameworks. AWS Glue is good for near-real-time streaming data processing for use cases such as streaming ETL.
  • Kinesis Data Streams on-demand mode charges by usage and automatically scales up resource capacity, so it’s good for spiky streaming workloads and hands-free maintenance. Provisioned mode charges by capacity and requires proactive capacity management, so it’s good for predictable streaming workloads.
  • You can use the Kinesis Shared Calculator to calculate the number of shards needed for provisioned mode. You don’t need to be concerned about shards with on-demand mode.
  • When granting permissions, you decide who is getting what permissions to which Kinesis Data Streams resources. You enable specific actions that you want to allow on those resources. Therefore, you should grant only the permissions that are required to perform a task. You can also encrypt the data at rest by using a KMS customer managed key (CMK).
  • You can update the retention period via the Kinesis Data Streams console or by using the IncreaseStreamRetentionPeriod and the DecreaseStreamRetentionPeriod operations based on your specific use cases.
  • Kinesis Data Streams supports resharding. The recommended API for this function is UpdateShardCount, which allows you to modify the number of shards in your stream to adapt to changes in the rate of data flow through the stream. The resharding APIs (Split and Merge) are typically used to handle hot shards.

Conclusion

This post demonstrated various architectural patterns for building low-latency streaming applications with Kinesis Data Streams. You can build your own low-latency steaming applications with Kinesis Data Streams using the information in this post.

For detailed architectural patterns, refer to the following resources:

If you want to build a data vision and strategy, check out the AWS Data-Driven Everything (D2E) program.

About the Authors

Raghavarao Sodabathina is a Principal Solutions Architect at AWS, focusing on Data Analytics, AI/ML, and cloud security. He engages with customers to create innovative solutions that address customer business problems and to accelerate the adoption of AWS services. In his spare time, Raghavarao enjoys spending time with his family, reading books, and watching movies.

Hang Zuo is a Senior Product Manager on the Amazon Kinesis Data Streams team at Amazon Web Services. He is passionate about developing intuitive product experiences that solve complex customer problems and enable customers to achieve their business goals.

Shwetha Radhakrishnan is a Solutions Architect for AWS with a focus in Data Analytics. She has been building solutions that drive cloud adoption and help organizations make data-driven decisions within the public sector. Outside of work, she loves dancing, spending time with friends and family, and traveling.

Brittany Ly is a Solutions Architect at AWS. She is focused on helping enterprise customers with their cloud adoption and modernization journey and has an interest in the security and analytics field. Outside of work, she loves to spend time with her dog and play pickleball.

Run Kinesis Agent on Amazon ECS

Post Syndicated from Buddhike de Silva original https://aws.amazon.com/blogs/big-data/run-kinesis-agent-on-amazon-ecs/

Kinesis Agent is a standalone Java software application that offers a straightforward way to collect and send data to Amazon Kinesis Data Streams and Amazon Kinesis Data Firehose. The agent continuously monitors a set of files and sends new data to the desired destination. The agent handles file rotation, checkpointing, and retry upon failures. It delivers all of your data in a reliable, timely, and simple manner. It also emits Amazon CloudWatch metrics to help you better monitor and troubleshoot the streaming process.

This post describes the steps to send data from a containerized application to Kinesis Data Firehose using Kinesis Agent. More specifically, we show how to run Kinesis Agent as a sidecar container for an application running in Amazon Elastic Container Service (Amazon ECS). After the data is in Kinesis Data Firehose, it can be sent to any supported destination, such as Amazon Simple Storage Service (Amazon S3).

In order to present the key points required for this setup, we assume that you are familiar with Amazon ECS and working with containers. We also avoid the implementation details and packaging process of our test data generation application, referred to as the producer.

Solution overview

As depicted in the following figure, we configure a Kinesis Agent container as a sidecar that can read files created by the producer container. In this instance, the producer and Kinesis Agent containers share data via a bind mount in Amazon ECS.

Solution design diagram

Prerequisites

You should satisfy the following prerequisites for the successful completion of this task:

With these prerequisites in place, you can begin next step to package a Kinesis Agent and your desired agent configuration as a container in your local development machine.

Create a Kinesis Agent configuration file

We use the Kinesis Agent configuration file to configure the source and destination, among other data transfer settings. The following code uses the minimal configuration required to read the contents of files matching /var/log/producer/*.log and publish them to a Kinesis Data Firehose delivery stream called kinesis-agent-demo:

{
    "firehose.endpoint": "firehose.ap-southeast-2.amazonaws.com",
    "flows": [
        {
            "deliveryStream": "kinesis-agent-demo",
            "filePattern": "/var/log/producer/*.log"
        }
    ]
}

Create a container image for Kinesis Agent

To deploy Kinesis Agent as a sidecar in Amazon ECS, you first have to package it as a container image. The container must have Kinesis Agent, which and find binaries, and the Kinesis Agent configuration file that you prepared earlier. Its entry point must be configured using the start-aws-kinesis-agent script. This command is installed when you run the yum install aws-kinesis-agent step. The resulting Dockerfile should look as follows:

FROM amazonlinux

RUN yum install -y aws-kinesis-agent which findutils
COPY agent.json /etc/aws-kinesis/agent.json

CMD ["start-aws-kinesis-agent"]

Run the docker build command to build this container:

docker build -t kinesis-agent .

After the image is built, it should be pushed to a container registry like Amazon ECR so that you can reference it in the next section.

Create an ECS task definition with Kinesis Agent and the application container

Now that you have Kinesis Agent packaged as a container image, you can use it in your ECS task definitions to run as sidecar. To do that, you create an ECS task definition with your application container (called producer) and Kinesis Agent container. All containers in a task definition are scheduled on the same container host and therefore can share resources such as bind mounts.

In the following sample container definition, we use a bind mount called logs_dir to share a directory between the producer container and kinesis-agent container.

You can use the following template as a starting point, but be sure to change taskRoleArn and executionRoleArn to valid IAM roles in your AWS account. In this instance, the IAM role used for taskRoleArn must have write permissions to Kinesis Data Firehose that you specified earlier in the agent.json file. Additionally, make sure that the ECR image paths and awslogs-region are modified as per your AWS account.

{
    "family": "kinesis-agent-demo",
    "taskRoleArn": "arn:aws:iam::111111111:role/kinesis-agent-demo-task-role",
    "executionRoleArn": "arn:aws:iam::111111111:role/kinesis-agent-test",
    "networkMode": "awsvpc",
    "containerDefinitions": [
        {
            "name": "producer",
            "image": "111111111.dkr.ecr.ap-southeast-2.amazonaws.com/producer:latest",
            "cpu": 1024,
            "memory": 2048,
            "essential": true,
            "command": [
                "-output",
                "/var/log/producer/test.log"
            ],
            "mountPoints": [
                {
                    "sourceVolume": "logs_dir",
                    "containerPath": "/var/log/producer",
                    "readOnly": false
                }
            ],
            "logConfiguration": {
                "logDriver": "awslogs",
                "options": {
                    "awslogs-create-group": "true",
                    "awslogs-group": "producer",
                    "awslogs-stream-prefix": "producer",
                    "awslogs-region": "ap-southeast-2"
                }
            }
        },
        {
            "name": "kinesis-agent",
            "image": "111111111.dkr.ecr.ap-southeast-2.amazonaws.com/kinesis-agent:latest",
            "cpu": 1024,
            "memory": 2048,
            "essential": true,
            "mountPoints": [
                {
                    "sourceVolume": "logs_dir",
                    "containerPath": "/var/log/producer",
                    "readOnly": true
                }
            ],
            "logConfiguration": {
                "logDriver": "awslogs",
                "options": {
                    "awslogs-create-group": "true",
                    "awslogs-group": "kinesis-agent",
                    "awslogs-stream-prefix": "kinesis-agent",
                    "awslogs-region": "ap-southeast-2"
                }
            }
        }
    ],
    "volumes": [
        {
            "name": "logs_dir"
        }
    ],
    "requiresCompatibilities": [
        "FARGATE"
    ],
    "cpu": "2048",
    "memory": "4096"
}

Register the task definition with the following command:

aws ecs register-task-definition --cli-input-json file://./task-definition.json

Run a new ECS task

Finally, you can run a new ECS task using the task definition you just created using the aws ecs run-task command. When the task is started, you should be able to see two containers running under that task on the Amazon ECS console.

Amazon ECS console screenshot

Conclusion

This post showed how straightforward it is to run Kinesis Agent in a containerized environment. Although we used Amazon ECS as our container orchestration service in this post, you can use a Kinesis Agent container in other environments such as Amazon Elastic Kubernetes Service (Amazon EKS).

To learn more about using Kinesis Agent, refer to Writing to Amazon Kinesis Data Streams Using Kinesis Agent. For more information about Amazon ECS, refer to the Amazon ECS Developer Guide.

About the Author

Buddhike de Silva is a Senior Specialist Solutions Architect at Amazon Web Services. Buddhike helps customers run large scale streaming analytics workloads on AWS and make the best out of their cloud journey.

Amazon Kinesis Data Streams: celebrating a decade of real-time data innovation

Post Syndicated from Roy Wang original https://aws.amazon.com/blogs/big-data/amazon-kinesis-data-streams-celebrating-a-decade-of-real-time-data-innovation/

Data is a key strategic asset for every organization, and every company is a data business at its core. However, in many organizations, data is typically spread across a number of different systems such as software as a service (SaaS) applications, operational databases, and data warehouses. Such data silos make it difficult to get unified views of the data in an organization and act in real time to derive the most value.

Ten years ago, we launched Amazon Kinesis Data Streams, the first cloud-native serverless streaming data service, to serve as the backbone for companies, to move data across system boundaries, breaking data silos. With data streaming, you can power data lakes running on Amazon Simple Storage Service (Amazon S3), enrich customer experiences via personalization, improve operational efficiency with predictive maintenance of machinery in your factories, and achieve better insights with more accurate machine learning (ML) models. Amazon Kinesis Data Streams is a foundational data strategy pillar for tens of thousands of customers. As streams of raw data come together, they unlock capabilities to continuously transform, enrich, and query data in real time via seamless integration with stream processing engines such as Amazon Managed Service for Apache Flink.

As an example, the National Hockey League (NHL) reimagined the fan experience by streaming live NHL EDGE game data and stats to offer hockey fans valuable insights to keep fans at the edge of their seats. NHL EDGE technology in the puck and players’ sweaters (jerseys) generate thousands of data points every second for the NHL, which can be analyzed by AWS to predict likely outcomes for key events like face-offs. To process and analyze thousands of signals, the NHL built a real-time streaming data foundation with Kinesis Data Streams and Amazon Managed Service for Apache Flink to stream, prepare, and feed data into ML models, helping inform face-off predictions in seconds and expanding new ways to engage viewers.

Building on such streaming data foundations, many customers are currently thinking about how to deliver transformative new products and services with generative AI. Streaming allows companies to connect the data available within data stores to large language models (LLMs) securely and in real time. Although LLMs are capable of working with billions of parameters, in order to deliver an engaging experience that is tailored to a company’s customers, LLMs require personalization data for the company’s users and proprietary knowledge stores within the company’s data stores. A data strategy that incorporates streaming is necessary to deliver personalization and proprietary data that is available for querying in real time.

Customers with real-time streaming data strategy are at the cutting edge of providing innovative products with generative AI. One customer adopted Kinesis Data Streams for their data strategy, and they stream billions of events from their digital products to derive real-time insights. With a combination of low-latency data streaming and analytics, they are able to understand and personalize the user experience via a seamlessly integrated, self-reliant system for experimentation and automated feedback. Earlier this year, building on their already strong data foundation, they launched an innovative digital media generative AI product. The same data foundation built on Kinesis Data Streams is used to continuously analyze how users interact with the generated content and helps the product team fine-tune the application.

“Real-time streaming data technologies are essential for digital transformation. These services help customers bring data to their applications and models, making them smarter. Real-time data gives companies an advantage in data-driven decisions, predictions, and insights by using the data at the very moment it is generated, providing an unparalleled edge in a world where timing is the key to success. Bring the data in once, use it across your organization, and act before the value of that data diminishes.”

– Mindy Ferguson, VP of AWS Streaming and Messaging.

As we celebrate the tenth anniversary of Kinesis Data Streams, customers have shared four key reasons they continue to value this revolutionary service. They love how they can easily stream data with no underlying servers to provision or manage, operate at a massive scale with consistent performance, achieve high resiliency and durability, and benefit from broad integration with myriad sources and sinks to ingest and process data respectively.

Ease of use

Getting started with Kinesis Data Streams is straightforward: developers can create a data stream with a few clicks on the Kinesis Data Streams console or with a single API call. Changing the size or configuration is also a single API call, and each data stream comes with a default 24-hour data retention period. Developers don’t have to worry about clusters, version upgrades, or storage capacity planning. They just turn on a data stream and start ingesting data.

The needs of our customers have evolved in the past 10 years. As more events get captured and streamed, customers want their data streams to scale elastically without any operational overhead. In response, we launched On-Demand streams in 2021 to provide a simple and automatic scaling experience. With On-Demand streams, you let the service handle scaling up a stream’s capacity proactively, and you’re only charged for the actual data ingested, retrieved, and stored. As our customers continued to ask for more capabilities, we increased the ingestion throughput limit of each On-Demand stream from 200MB/s to 1GB/s in March 2023, and then to 2GB/s in October 2023, to accommodate higher throughput workloads. To continue innovating to be the easiest streaming data service to use, we actively listen to our customer use cases.

Canva is an online design and visual communication platform. As it has rapidly grown from 30 million to 135 million monthly users, it has built a streaming data platform at scale that is effortless to operate for driving product innovations and personalizing the user experience.

“Amazon Kinesis Data Streams and AWS Lambda are used throughout Canva’s logging platform, ingesting and processing over 60 billion log events per day. The combination of Kinesis Data Streams and Lambda has abstracted plenty of work that’s often required in managing a massive data pipeline, such as deploying and managing a fleet of servers, whilst also providing a highly scalable and reliable service. It has allowed us to focus on delivering a world-class product by building highly requested features rather than spending time on operational work.”

– Phoebe Zhou, Software Engineer at Canva.

Operate at massive scale with consistent performance

A fundamental requirement of a streaming data strategy is ingesting and processing large volumes of data with low latency. Kinesis Data Streams processes trillions of records per day across tens of thousands of customers. Customers run more than 3.5 million unique streams and process over 45 PB of data per day. Our largest customers ingest more than 15 GB per second of real-time data with individual streams. That’s equivalent to streaming multiple data points for every person on earth, every second! Even at this scale, all our customers still retrieve data within milliseconds of availability.

Customers also want to process the same data with multiple applications, with each deriving a different value, without worrying about one application impacting the read throughput of another. Enhanced Fan-out offers dedicated read throughput and low latency for each data consumer. This has enabled enterprise platform teams to provide real-time data to more teams and applications.

VMware Carbon Black uses Kinesis Data Streams to ingest petabytes of data every day to secure millions of customer endpoints. The team focuses on its expertise while AWS manages data streaming to meet growing customer traffic and needs in real time.

“When an individual customer’s data increases or decreases, we can use the elasticity of Amazon Kinesis Data Streams to scale compute up or down to process data reliably while effectively managing our cost. This is why Kinesis Data Streams is a good fit. The biggest advantage is the managed nature of our solution on AWS. This has shaped our architecture and helped us shift complexity elsewhere.”

– Stoyan Dimkov, Staff Engineer and Software Architect at VMware Carbon Black.

Learn more about the case study.

Provide resiliency and durability for data streaming

With burgeoning data, customers want more flexibility in processing and reprocessing data. For example, if an application that is consuming data goes offline for a period, teams want to ensure that they resume processing at a later time without data loss. Kinesis Data Streams provides a default 24-hour retention period, enabling you to select a specific timestamp from which to start processing records. With the extended retention feature, you can configure the data retention period to be up to 7 days.

Some industries like financial services and healthcare have stricter compliance requirements, so customers asked for even longer data retention periods to support these requirements. Therefore, we followed up with long-term storage that supports data retention for up to 1 year. Now, thousands of Kinesis Data Streams customers use these features to make their streaming applications more resilient and durable.

Mercado Libre, a leading ecommerce and payments platform in Latin America, relies on Kinesis Data Streams to power its streaming data strategy around payment processing, customer experience, and operations.

“With Amazon Kinesis Data Streams at the core, we process approximately 70 billion daily messages distributed across thousands of data producers. By leveraging Kinesis Data Streams and Amazon DynamoDB Streams, we’ve embraced an event-driven architecture and are able to swiftly respond to data changes.”

– Joaquin Fernandez, Senior Software Expert at Mercado Libre.

Access your data no matter where it lives

Our customers use a wide variety of tools and applications, and an organization’s data often resides in many places. Therefore, the ability to easily integrate data across an organization is crucial to derive timely insights. Developers use the Kinesis Producer Library, Kinesis Client Library, and AWS SDK to quickly build custom data producer and data consumer applications. Customers have expanded their data producers ranging from microservices to smart TVs and even cars. We have over 40 integrations with AWS services and third-party applications like Adobe Experience Platform and Databricks. As detailed in our whitepaper on building a modern data streaming architecture on AWS, Kinesis Data Streams serves as the backbone to serverless and real-time use cases such as personalization, real-time insights, Internet of Things (IoT), and event-driven architecture. Our recent integration with Amazon Redshift enables you to ingest hundreds of megabytes of data from Kinesis Data Streams into data warehouses in seconds. To learn more about how to use this integration to detect fraud in near-real time, refer to Near-real-time fraud detection using Amazon Redshift Streaming Ingestion with Amazon Kinesis Data Streams and Amazon Redshift ML.

Another integration launched in 2023 is with Amazon Monitron to power predictive maintenance management. You can now stream measurement data and the corresponding inference results to Kinesis Data Streams, coordinate predictive maintenance, and build an IoT data lake. For more details, refer to Generate actionable insights for predictive maintenance management with Amazon Monitron and Amazon Kinesis.

Next, let’s go back to the NHL use case where they combine IoT, data streaming, and machine learning.

The NHL Edge IQ powered by AWS is helping bring fans closer to the action with advanced analytics and new ML stats such as Face-off Probability and Opportunity Analysis.

“We use Amazon Kinesis Data Streams to process NHL EDGE data on puck and Player positions, face-off location, and the current game situation to decouple data producers from consuming applications. Amazon Managed Service for Apache Flink is used to run Flink applications and consumes data from Kinesis Data Streams to call the prediction model in Amazon SageMaker to deliver the real-time Face-off Probability metric. The probability results are also stored in Amazon S3 to continuously retrain the model in SageMaker. The success of this project led us to build the next metric, Opportunity Analysis, which delivers over 25 insights into the quality of the scoring opportunity presented by each shot on goal. Kinesis Data Streams and Amazon Managed Service for Apache Flink applications were critical to making live, in-game predictions, enabling the system to perform opportunity analysis calculations for up to 16 live NHL games simultaneously.”

– Eric Schneider, SVP, Software Engineering at National Hockey League.

Learn more about the case study.

The future of data is real time

The fusion of real-time data streaming and generative AI promises to be the cornerstone of our digitally connected world. Generative AI, empowered by a constant influx of real-time information from IoT devices, sensors, social media, and beyond, is becoming ubiquitous. From autonomous vehicles navigating dynamically changing traffic conditions to smart cities optimizing energy consumption based on real-time demand, the combination of AI and real-time data will underpin efficiency and innovation across industries. Ubiquitous, adaptive, and deeply integrated into our lives, these AI-driven applications will enhance convenience and address critical challenges such as climate change, healthcare, and disaster response by using the wealth of real-time insights at their disposal. With Kinesis Data Streams, organizations can build a solid data foundation, positioning you to quickly adopt new technologies and unlock new opportunities sooner—which we anticipate will be enormous.

Learn more about what our customers are doing with data streaming. If you would like a quick exploration of Kinesis Data Streams concepts and use cases, check out our Amazon Kinesis Data Streams 101 playlist. To get started with building your data streams, visit the Amazon Kinesis Data Streams Developer Guide.

About the author

Roy (KDS) Wang is a Senior Product Manager with Amazon Kinesis Data Streams. He is passionate about learning from and collaborating with customers to help organizations run faster and smarter. Outside of work, Roy strives to be a good dad to his new son and builds plastic model kits.

Processing large records with Amazon Kinesis Data Streams

Post Syndicated from Masudur Rahaman Sayem original https://aws.amazon.com/blogs/big-data/processing-large-records-with-amazon-kinesis-data-streams/

In today’s digital era, data is abundant and constantly flowing. Businesses across industries are seeking ways to harness this wealth of information to gain valuable insights and make real-time decisions. To meet this need, AWS offers Amazon Kinesis Data Streams, a powerful and scalable real-time data streaming service. With Kinesis Data Streams, you can effortlessly collect, process, and analyze streaming data in real time at any scale. This service seamlessly integrates into your data architecture, allowing you to tap into the full potential of your data for informed decision-making.

Data streaming technologies like Kinesis Data Streams are designed to efficiently process and manage continuous streams of data in real time at large scale. The individual pieces of data within these streams are often referred to as records. In scenarios like large file processing or performing image, audio, or video analytics, your record may exceed 1 MB. You may struggle to ingest such a large record with Kinesis Data Streams because, as of this writing, the service has a 1 MB upper limit for maximum data record size.

In this post, we show you some different options for handling large records within Kinesis Data Streams and the benefits and disadvantages of each approach. We provide some sample code for each option to help you get started with any of these approaches with your own workloads.

Understanding the default behavior of Kinesis Data Streams

You can send records to Kinesis Data Streams using the PutRecord or PutRecords API calls. These APIs include a mandatory field known as PartitionKey, where you must provide a specific value. This partition key is used by the service to map records with the same partition keys to the same shard to ensure ordering and locality for consumption. Locality means that you want the same consumer to process all records for a given partition key. This helps ensure that data with the same partition key stays together within the same shard, maintaining data order.

Each shard, which holds your data, can handle writing up to 1 MB per second. Let’s consider a scenario where you define a partition key and attempt to send a data record that exceeds 1 MB in size. Based on the explanation so far, the service will reject this request because the record size is over 1 MB. To help you understand better, we experimented by trying to send a record of 1.5 MB to a stream, and the outcome was the following exception message:

import json
import boto3
client = boto3.client('kinesis', region_name='ap-southeast-2')

def lambda_handler(event, context):
    try:
        response = client.put_record(
            StreamName='test',
            Data=b'Sample 1 MB....',
            PartitionKey='string'
            #StreamARN='string'
        )
    
    except Exception as e:
        print (e)


START RequestId: 84b3ab0c-3f30-4267-aec1-549c2d59dfdb Version: $LATEST An error occurred (ValidationException) when calling the PutRecord operation: 1 validation error detected: Value at 'data' failed to satisfy constraint: Member must have length less than or equal to 1048576 END RequestId: 84b3ab0c-3f30-4267-aec1-549c2d59dfdb

Strategies for handling large records

Now that we understand the behavior of the PutRecord and PutRecords APIs, let’s discuss strategies you can use to overcome this situation. One thing to keep in mind is that there is no single best solution; in the following sections, we discuss some of the approaches that you can evaluate based on your use case:

  • Store large records in Amazon Simple Storage Service (Amazon S3) with a reference in Kinesis Data Streams
  • Split one large record into multiple records
  • Compress your large records

Let’s discuss these points one by one.

Store large records in Amazon S3 with a reference in Kinesis Data Streams

A useful approach for storing large records involves utilizing an alternative storage solution while employing a reference within Kinesis Data Streams. In this context, Amazon S3 stands out as an excellent choice due to its exceptional durability and cost-effectiveness. The procedure involves uploading the record as an object to an S3 bucket and subsequently writing a reference entry in Kinesis Data Streams. This entry incorporates an attribute that serves as a pointer, indicating the location of the object within Amazon S3.

With this approach, you can generate a pre-signed URL associated with the S3 object’s location. This link can be shared with the requester, offering them direct access to the object without the need for intermediary server-side data transfers.

The following diagram illustrates the architecture of this solution.

The following is the sample code to write data to Kinesis Data Streams using this approach:

import json
import boto3
import random

def lambda_handler(event, context):
    try:
        s3 = boto3.client('s3', region_name='ap-southeast-2')
        kds = boto3.client('kinesis', region_name='ap-southeast-2')
        expiration=3600
        pk=str(random.randint(100,100000000))
        bucket_name = 'MY_BUCKET'
        object_key = 'air/' + pk + '.txt'
        file_content = b'LARGE OBJECT'
        response = s3.put_object(Bucket=bucket_name, Key=object_key, Body=file_content)
        presigned_url = s3.generate_presigned_url(
            'get_object',
            Params={'Bucket': bucket_name, 'Key': object_key},
            ExpiresIn=expiration
        )
        
        kdata = {'message': presigned_url}
        response = kds.put_record(
            StreamName='test',
            Data=json.dumps(kdata),
            PartitionKey=pk
        )
        print (response)
    except Exception as e:
        print (e)

If you are using an AWS Lambda consumer to process this data, you can now decode the record to get the S3 pre-signed URL to efficiently retrieve the object from Amazon S3. Then you can implement your business logic to effectively process the data. The following is sample code for reference:

import json
import base64
import json

def lambda_handler(event, context):
    item = None
    decoded_record_data = [base64.b64decode(record['kinesis']['data']).decode().replace('\n','') for record in event['Records']]
    deserialized_data = [json.loads(decoded_record) for decoded_record in decoded_record_data]
    
    
    for item in deserialized_data:
        LOB=(item['message'])
        #process LOB implementing your business logic

An inherent benefit of adopting this technique is the capability to store data in Amazon S3, accommodating an extensive range of sizes per individual object. This method helps you reduce the costs of using Kinesis Data Streams because it uses less storage space and requires fewer read and write throughput for item access. This optimization is achieved by storing just the URL within Kinesis Data Streams. However, it’s important to acknowledge that accessing the sizable object necessitates an additional call to Amazon S3, thereby introducing higher latency for clients as they manage the additional request.

Split one large record into multiple records

Splitting large records into smaller ones in Kinesis Data Streams brings advantages like faster processing, improved throughput, efficient resource use, and more straightforward error handling. Let’s say you have a large record that you want to split into smaller chunks before sending them to a Kinesis data stream. First, you need to set up a Kinesis producer. Suppose you have a large record as a string. You can split it into smaller chunks of a predefined size. For this example, let’s say you’re splitting the record into chunks of 100 characters each. After you split that, loop through the record chunks and send each chunk as a separate message to a Kinesis data stream. The following is the sample code:

import boto3
kinesis = boto3.client('kinesis', region_name='ap-southeast-2')  

def split_record(record, chunk_size):
    chunks = [record[i:i + chunk_size] for i in range(0, len(record), chunk_size)]
    return chunks

def send_to_kinesis(stream_name, record):
    response = kinesis.put_record(
        StreamName=stream_name,
        Data=record,
        PartitionKey= '100'
    )
    return response

def main():
    stream_name = 'test'  
    large_record = 'Your large record'  # Replace with your actual record
    chunk_size = 100  

    record_chunks = split_record(large_record, chunk_size)

    for chunk in record_chunks:
        response = send_to_kinesis(stream_name, chunk)
        print(f"Record sent: {response['SequenceNumber']}")

if __name__ == "__main__":
    main()

Ensure that all chunks of a given message are directed to a single partition, thereby guaranteeing the preservation of their order. In the final chunk, include metadata within the header indicating the conclusion of the message during production. This enables consumers to identify the ultimate chunk and facilitates seamless message reconstruction. The drawback of this method is that it adds complexity to the client-side tasks of dividing and putting back together the different parts. Therefore, these functions need thorough testing to prevent any loss of data.

Compress your large records

Applying data compression prior to transmitting it to Kinesis Data Streams has numerous advantages. This approach not only reduces the data’s size, enabling swifter travel and more efficient utilization of network resources, but also leads to cost savings in terms of storage expenses while optimizing overall resource consumption. Additionally, this practice simplifies storage and data retention. By using compression algorithms such as GZIP, Snappy, or LZ4, you can achieve substantial reduction in the size of large records. Compression brings the benefit of simplicity because it’s implemented seamlessly without requiring the caller to make changes to the item or use extra AWS services to support storage. However, compression introduces additional CPU overhead and latency on the producer side, and its impact on the compression ratio and efficiency can vary depending on the data type and format. Also, compression can enhance consumer throughput at the expense of some decompression overhead.

Conclusion

For real-time data streaming use cases, it’s essential to carefully consider the handling of large records when using Kinesis Data Streams. In this post, we discussed the challenges associated with managing large records and explored strategies such as utilizing Amazon S3 references, record splitting, and compression. Each approach has its own set of benefits and drawbacks, so it’s crucial to evaluate the nature of your data and the tasks you need to perform. Select the most suitable approach based on your data’s characteristics and your processing task requirements.

We encourage you to try out the approaches discussed in this post and share your thoughts in the comments section.

About the author

Masudur Rahaman Sayem is a Streaming Data Architect at AWS. He works with AWS customers globally to design and build data streaming architectures to solve real-world business problems. He specializes in optimizing solutions that use streaming data services and NoSQL. Sayem is very passionate about distributed computing.

Non-JSON ingestion using Amazon Kinesis Data Streams, Amazon MSK, and Amazon Redshift Streaming Ingestion

Post Syndicated from M Mehrtens original https://aws.amazon.com/blogs/big-data/non-json-ingestion-using-amazon-kinesis-data-streams-amazon-msk-and-amazon-redshift-streaming-ingestion/

Organizations are grappling with the ever-expanding spectrum of data formats in today’s data-driven landscape. From Avro’s binary serialization to the efficient and compact structure of Protobuf, the landscape of data formats has expanded far beyond the traditional realms of CSV and JSON. As organizations strive to derive insights from these diverse data streams, the challenge lies in seamlessly integrating them into a scalable solution.

In this post, we dive into Amazon Redshift Streaming Ingestion to ingest, process, and analyze non-JSON data formats. Amazon Redshift Streaming Ingestion allows you to connect to Amazon Kinesis Data Streams and Amazon Managed Streaming for Apache Kafka (Amazon MSK) directly through materialized views, in real time and without the complexity associated with staging the data in Amazon Simple Storage Service (Amazon S3) and loading it into the cluster. These materialized views not only provide a landing zone for streaming data, but also offer the flexibility of incorporating SQL transforms and blending into your extract, load, and transform (ELT) pipeline for enhanced processing. For a deeper exploration on configuring and using streaming ingestion in Amazon Redshift, refer to Real-time analytics with Amazon Redshift streaming ingestion.

JSON data in Amazon Redshift

Amazon Redshift enables storage, processing, and analytics on JSON data through the SUPER data type, PartiQL language, materialized views, and data lake queries. The base construct to access streaming data in Amazon Redshift provides metadata from the source stream (attributes like stream timestamp, sequence numbers, refresh timestamp, and more) and the raw binary data from the stream itself. For streams that contain the raw binary data encoded in JSON format, Amazon Redshift provides a variety of tools for parsing and managing the data. For more information about the metadata of each stream format, refer to Getting started with streaming ingestion from Amazon Kinesis Data Streams and Getting started with streaming ingestion from Amazon Managed Streaming for Apache Kafka.

At the most basic level, Amazon Redshift allows parsing the raw data into distinct columns. The JSON_EXTRACT_PATH_TEXT and JSON_EXTRACT_ARRAY_ELEMENT_TEXT functions enable the extraction of specific details from JSON objects and arrays, transforming them into separate columns for analysis. When the structure of the JSON documents and specific reporting requirements are defined, these methods allow for pre-computing a materialized view with the exact structure needed for reporting, with improved compression and sorting for analytics.

In addition to this approach, the Amazon Redshift JSON functions allow storing and analyzing the JSON data in its original state using the adaptable SUPER data type. The function JSON_PARSE allows you to extract the binary data in the stream and convert it into the SUPER data type. With the SUPER data type and PartiQL language, Amazon Redshift extends its capabilities for semi-structured data analysis. It uses the SUPER data type for JSON data storage, offering schema flexibility within a column. For more information on using the SUPER data type, refer to Ingesting and querying semistructured data in Amazon Redshift. This dynamic capability simplifies data ingestion, storage, transformation, and analysis of semi-structured data, enriching insights from diverse sources within the Redshift environment.

Streaming data formats

Organizations using alternative serialization formats must explore different deserialization methods. In the next section, we dive into the optimal approach for deserialization. In this section, we take a closer look at the diverse formats and strategies organizations use to effectively manage their data. This understanding is key in determining the data parsing approach in Amazon Redshift.

Many organizations use a format other than JSON for their streaming use cases. JSON is a self-describing serialization format, where the schema of the data is stored alongside the actual data itself. This makes JSON flexible for applications, but this approach can lead to increased data transmission between applications due to the additional data contained in the JSON keys and syntax. Organizations seeking to optimize their serialization and deserialization performance, and their network communication between applications, may opt to use a format like Avro, Protobuf, or even a custom proprietary format to serialize application data into binary format in an optimized way. This provides the advantage of an efficient serialization where only the message values are packed into a binary message. However, this requires the consumer of the data to know what schema and protocol was used to serialize the data to deserialize the message. There are several ways that organizations can solve this problem, as illustrated in the following figure.

Visualization of different binary message serialization approaches

Embedded schema

In an embedded schema approach, the data format itself contains the schema information alongside the actual data. This means that when a message is serialized, it includes both the schema definition and the data values. This allows anyone receiving the message to directly interpret and understand its structure without needing to refer to an external source for schema information. Formats like JSON, MessagePack, and YAML are examples of embedded schema formats. When you receive a message in this format, you can immediately parse it and access the data with no additional steps.

Assumed schema

In an assumed schema approach, the message serialization contains only the data values, and there is no schema information included. To interpret the data correctly, the receiving application needs to have prior knowledge of the schema that was used to serialize the message. This is typically achieved by associating the schema with some identifier or context, like a stream name. When the receiving application reads a message, it uses this context to retrieve the corresponding schema and then decodes the binary data accordingly. This approach requires an additional step of schema retrieval and decoding based on context. This generally requires setting up a mapping in-code or in an external database so that consumers can dynamically retrieve the schemas based on stream metadata (such as the AWS Glue Schema Registry).

One drawback of this approach is in tracking schema versions. Although consumers can identify the relevant schema from the stream name, they can’t identify the particular version of the schema that was used. Producers need to ensure that they are making backward-compatible changes to schemas to ensure consumers aren’t disrupted when using a different schema version.

Embedded schema ID

In this case, the producer continues to serialize the data in binary format (like Avro or Protobuf), similar to the assumed schema approach. However, an additional step is involved: the producer adds a schema ID at the beginning of the message header. When a consumer processes the message, it starts by extracting the schema ID from the header. With this schema ID, the consumer then fetches the corresponding schema from a registry. Using the retrieved schema, the consumer can effectively parse the rest of the message. For example, the AWS Glue Schema Registry provides Java SDK SerDe libraries, which can natively serialize and deserialize messages in a stream using embedded schema IDs. Refer to How the schema registry works for more information about using the registry.

The usage of an external schema registry is common in streaming applications because it provides a number of benefits to consumers and developers. This registry contains all the message schemas for the applications and associates them with a unique identifier to facilitate schema retrieval. In addition, the registry may provide other functionalities like schema version change handling and documentation to facilitate application development.

The embedded schema ID in the message payload can contain version information, ensuring publishers and consumers are always using the same schema version to manage data. When schema version information isn’t available, schema registries can help enforce producers making backward-compatible changes to avoid causing issues in consumers. This helps decouple producers and consumers, provides schema validation at both the publisher and consumer stage, and allows for more flexibility in stream usage to allow for a variety of application requirements. Messages can be published with one schema per stream, or with multiple schemas inside a single stream, allowing consumers to dynamically interpret messages as they arrive.

For a deeper dive into the benefits of a schema registry, refer to Validate streaming data over Amazon MSK using schemas in cross-account AWS Glue Schema Registry.

Schema in file

For batch processing use cases, applications may embed the schema used to serialize the data into the data file itself to facilitate data consumption. This is an extension of the embedded schema approach but is less costly because the data file is generally larger, so the schema accounts for a proportionally smaller amount of the overall data. In this case, the consumers can process the data directly without additional logic. Amazon Redshift supports loading Avro data that has been serialized in this manner using the COPY command.

Convert non-JSON data to JSON

Organizations aiming to use non-JSON serialization formats need to develop an external method for parsing their messages outside of Amazon Redshift. We recommend using an AWS Lambda-based external user-defined function (UDF) for this process. Using an external Lambda UDF allows organizations to define arbitrary deserialization logic to support any message format, including embedded schema, assumed schema, and embedded schema ID approaches. Although Amazon Redshift supports defining Python UDFs natively, which may be a viable alternative for some use cases, we demonstrate the Lambda UDF approach in this post to cover more complex scenarios. For examples of Amazon Redshift UDFs, refer to AWS Samples on GitHub.

The basic architecture for this solution is as follows.

See the following code:

-- Step 1
CREATE OR REPLACE EXTERNAL FUNCTION fn_lambda_decode_avro_binary(varchar)
RETURNS varchar IMMUTABLE LAMBDA 'redshift-avro-udf';

-- Step 2
CREATE EXTERNAL SCHEMA kds FROM KINESIS

-- Step 3
CREATE MATERIALIZED VIEW {name} AUTO REFRESH YES AS
SELECT
    -- Step 4
   t.kinesis_data AS binary_avro,
   to_hex(binary_avro) AS hex_avro,
   -- Step 5
   fn_lambda_decode_avro_binary('{stream-name}', hex_avro) AS json_string,
   -- Step 6
   JSON_PARSE(json_string) AS super_data,
   t.sequence_number,
   t.refresh_time,
   t.approximate_arrival_timestamp,
   t.shard_id
FROM kds.{stream_name} AS t

Let’s explore each step in more detail.

Create the Lambda UDF

The overall goal is to develop a method that can accept the raw data as input and produce JSON-encoded data as an output. This aligns with the Amazon Redshift ability to natively process JSON into the SUPER data type. The specifics of the function depend on the serialization and streaming approach. For example, using the assumed schema approach with Avro format, your Lambda function may complete the following steps:

  1. Take in the stream name and hexadecimal-encoded data as inputs.
  2. Use the stream name to perform a lookup to identify the schema for the given stream name.
  3. Decode the hexadecimal data into binary format.
  4. Use the schema to deserialize the binary data into readable format.
  5. Re-serialize the data into JSON format.

The f_glue_schema_registry_avro_to_json AWS samples example illustrates the process of decoding Avro using the assumed schema approach using the AWS Glue Schema Registry in a Lambda UDF to retrieve and use Avro schemas by stream name. For other approaches (such as embedded schema ID), you should author your Lambda function to handle deserialization as defined by your serialization process and schema registry implementation. If your application depends on an external schema registry or table lookup to process the message schema, we recommend that you implement caching for schema lookups to help reduce the load on the external systems and reduce the average Lambda function invocation duration.

When creating the Lambda function, make sure you accommodate the Amazon Redshift input event format and ensure compliance with the expected Amazon Redshift event output format. For details, refer to Creating a scalar Lambda UDF.

After you create and test the Lambda function, you can define it as a UDF in Amazon Redshift. For effective integration within Amazon Redshift, designate this Lambda function UDF as IMMUTABLE. This classification supports incremental materialized view updates. This treats the Lambda function as idempotent and minimizes the Lambda function costs for the solution, because a message doesn’t need to be processed if it has been processed before.

Configure the baseline Kinesis data stream

Regardless of your messaging format or approach (embedded schema, assumed schema, and embedded schema ID), you begin with setting up the external schema for streaming ingestion from your messaging source into Amazon Redshift. For more information, refer to Streaming ingestion.

CREATE EXTERNAL SCHEMA kds FROM KINESIS

IAM_ROLE 'arn:aws:iam::0123456789:role/redshift-streaming-role';

Create the raw materialized view

Next, you define your raw materialized view. This view contains the raw message data from the streaming source in Amazon Redshift VARBYTE format.

Convert the VARBYTE data to VARCHAR format

External Lambda function UDFs don’t support VARBYTE as an input data type. Therefore, you must convert the raw VARBYTE data from the stream into VARCHAR format to pass to the Lambda function. The best way to do this in Amazon Redshift is using the TO_HEX built-in method. This converts the binary data into hexadecimal-encoded character data, which can be sent to the Lambda UDF.

Invoke the Lambda function to retrieve JSON data

After the UDF has been defined, we can invoke the UDF to convert our hexadecimal-encoded data into JSON-encoded VARCHAR data.

Use the JSON_PARSE method to convert the JSON data to SUPER data type

Finally, we can use the Amazon Redshift native JSON parsing methods like JSON_PARSE, JSON_EXTRACT_PATH_TEXT, and more to parse the JSON data into a format that we can use for analytics.

Considerations

Consider the following when using this strategy:

  • Cost – Amazon Redshift invokes the Lambda function in batches to improve scalability and reduce the overall number of Lambda invocations. The cost of this solution depends on the number of messages in your stream, the frequency of the refresh, and the invocation time required to process the messages in a batch from Amazon Redshift. Using the IMMUTABLE UDF type in Amazon Redshift can also help minimize costs by utilizing the incremental refresh strategy for the materialized view.
  • Permissions and network access – The AWS Identity and Access Management (IAM) role used for the Amazon Redshift UDF must have permissions to invoke the Lambda function, and you must deploy the Lambda function such that it has access to invoke its external dependencies (for example, you may need to deploy it in a VPC to access private resources like a schema registry).
  • Monitoring – Use Lambda function logging and metrics to identify errors in deserialization, connection to the schema registry, and data processing. For details on monitoring the UDF Lambda function, refer to Embedding metrics within logs and Monitoring and troubleshooting Lambda functions.

Conclusion

In this post, we dove into different data formats and ingestion methods for a streaming use case. By exploring strategies for handling non-JSON data formats, we examined the use of Amazon Redshift streaming to seamlessly ingest, process, and analyze these formats in near-real time using materialized views.

Furthermore, we navigated through schema-per-stream, embedded schema, assumed schema, and embedded schema ID approaches, highlighting their merits and considerations. To bridge the gap between non-JSON formats and Amazon Redshift, we explored the creation of Lambda UDFs for data parsing and conversion. This approach offers a comprehensive means to integrate diverse data streams into Amazon Redshift for subsequent analysis.

As you navigate the ever-evolving landscape of data formats and analytics, we hope this exploration provides valuable guidance to derive meaningful insights from your data streams. We welcome any thoughts or questions in the comments section.

About the Authors

M Mehrtens has been working in distributed systems engineering throughout their career, working as a Software Engineer, Architect, and Data Engineer. In the past, M has supported and built systems to process terrabytes of streaming data at low latency, run enterprise Machine Learning pipelines, and created systems to share data across teams seamlessly with varying data toolsets and software stacks. At AWS, they are a Sr. Solutions Architect supporting US Federal Financial customers.

Sindhu Achuthan is a Sr. Solutions Architect with Federal Financials at AWS. She works with customers to provide architectural guidance on analytics solutions using AWS Glue, Amazon EMR, Amazon Kinesis, and other services. Outside of work, she loves DIYs, to go on long trails, and yoga.

How Chime Financial uses AWS to build a serverless stream analytics platform and defeat fraudsters

Post Syndicated from Khandu Shinde original https://aws.amazon.com/blogs/big-data/how-chime-financial-uses-aws-to-build-a-serverless-stream-analytics-platform-and-defeat-fraudsters/

This is a guest post by Khandu Shinde, Staff Software Engineer and Edward Paget, Senior Software Engineering at Chime Financial.

Chime is a financial technology company founded on the premise that basic banking services should be helpful, easy, and free. Chime partners with national banks to design member first financial products. This creates a more competitive market with better, lower-cost options for everyday Americans who aren’t being served well by traditional banks. We help drive innovation, inclusion, and access across the industry.

Chime has a responsibility to protect our members against unauthorized transactions on their accounts. Chime’s Risk Analysis team constantly monitors trends in our data to find patterns that indicate fraudulent transactions.

This post discusses how Chime utilizes AWS Glue, Amazon Kinesis, Amazon DynamoDB, and Amazon SageMaker to build an online, serverless fraud detection solution — the Chime Streaming 2.0 system.

Problem statement

In order to keep up with the rapid movement of fraudsters, our decision platform must continuously monitor user events and respond in real-time. However, our legacy data warehouse-based solution was not equipped for this challenge. It was designed to manage complex queries and business intelligence (BI) use cases on a large scale. However, with a minimum data freshness of 10 minutes, this architecture inherently didn’t align with the near real-time fraud detection use case.

To make high-quality decisions, we need to collect user event data from various sources and update risk profiles in real time. We also need to be able to add new fields and metrics to the risk profiles as our team identifies new attacks, without needing engineering intervention or complex deployments.

We decided to explore streaming analytics solutions where we can capture, transform, and store event streams at scale, and serve rule-based fraud detection models and machine learning (ML) models with milliseconds latency.

Solution overview

The following diagram illustrates the design of the Chime Streaming 2.0 system.

The design included the following key components:

  1. We have Amazon Kinesis Data Streams as our streaming data service to capture and store event streams at scale. Our stream pipelines capture various event types, including user enrollment events, user login events, card swipe events, peer-to-peer payments, and application screen actions.
  2. Amazon DynamoDB is another data source for our Streaming 2.0 system. It acts as the application backend and stores data such as blocked devices list and device-user mapping. We mainly use it as lookup tables in our pipeline.
  3. AWS Glue jobs form the backbone of our Streaming 2.0 system. The simple AWS Glue icon in the diagram represents thousands of AWS Glue jobs performing different transformations. To achieve the 5-15 seconds end-to-end data freshness service level agreement (SLA) for the Steaming 2.0 pipeline, we use streaming ETL jobs in AWS Glue to consume data from Kinesis Data Streams and apply near-real-time transformation. We choose AWS Glue mainly due to its serverless nature, which simplifies infrastructure management with automatic provisioning and worker management, and the ability to perform complex data transformations at scale.
  4. The AWS Glue streaming jobs generate derived fields and risk profiles that get stored in Amazon DynamoDB. We use Amazon DynamoDB as our online feature store due to its millisecond performance and scalability.
  5. Our applications call Amazon SageMaker Inference endpoints for fraud detections. The Amazon DynamoDB online feature store supports real-time inference with single digit millisecond query latency.
  6. We use Amazon Simple Storage Service (Amazon S3) as our offline feature store. It contains historical user activities and other derived ML features.
  7. Our data scientist team can access the dataset and perform ML model training and batch inferencing using Amazon SageMaker.

AWS Glue pipeline implementation deep dive

There are several key design principles for our AWS Glue Pipeline and the Streaming 2.0 project.

  • We want to democratize our data platform and make the data pipeline accessible to all Chime developers.
  • We want to implement cloud financial backend services and achieve cost efficiency.

To achieve data democratization, we needed to enable different personas in the organization to use the platform and define transformation jobs quickly, without worrying about the actual implementation details of the pipelines. The data infrastructure team built an abstraction layer on top of Spark and integrated services. This layer contained API wrappers over integrated services, job tags, scheduling configurations and debug tooling, hiding Spark and other lower-level complexities from end users. As a result, end users were able to define jobs with declarative YAML configurations and define transformation logic with SQL. This simplified the onboarding process and accelerated the implementation phase.

To achieve cost efficiency, our team built a cost attribution dashboard based on AWS cost allocation tags. We enforced tagging with the above abstraction layer and had clear cost attribution for all AWS Glue jobs down to the team level. This enabled us to track down less optimized jobs and work with job owners to implement best practices with impact-based priority. One common misconfiguration we found was sizing of AWS Glue jobs. With data democratization, many users lacked the knowledge to right-size their AWS Glue jobs. The AWS team introduced AWS Glue auto scaling to us as a solution. With AWS Glue Auto Scaling, we no longer needed to plan AWS Glue Spark cluster capacity in advance. We could just set the maximum number of workers and run the jobs. AWS Glue monitors the Spark application execution, and allocates more worker nodes to the cluster in near-real time after Spark requests more executors based on our workload requirements. We noticed a 30–45% cost saving across our AWS Glue Jobs once we turned on Auto Scaling.

Conclusion

In this post, we showed you how Chime’s Streaming 2.0 system allows us to ingest events and make them available to the decision platform just seconds after they are emitted from other services. This enables us to write better risk policies, provide fresher data for our machine learning models, and protect our members from unauthorized transactions on their accounts.

Over 500 developers in Chime are using this streaming pipeline and we ingest more than 1 million events per second. We follow the sizing and scaling process from the AWS Glue streaming ETL jobs best practices blog and land on a 1:1 mapping between Kinesis Shard and vCPU core. The end-to-end latency is less than 15 seconds, and it improves the model score calculation speed by 1200% compared to legacy implementation. This system has proven to be reliable, performant, and cost-effective at scale.

We hope this post will inspire your organization to build a real-time analytics platform using serverless technologies to accelerate your business goals.

About the Authors

Khandu Shinde Khandu Shinde is a Staff Engineer focused on Big Data Platforms and Solutions for Chime. He helps to make the platform scalable for Chime’s business needs with architectural direction and vision. He’s based in San Francisco where he plays cricket and watches movies.

Edward Paget Edward Paget is a Software Engineer working on building Chime’s capabilities to mitigate risk to ensure our members’ financial peace of mind. He enjoys being at the intersection of big data and programming language theory. He’s based in Chicago where he spends his time running along the lake shore.

Dylan Qu is a Specialist Solutions Architect focused on Big Data & Analytics with Amazon Web Services. He helps customers architect and build highly scalable, performant, and secure cloud-based solutions on AWS.

Perform Amazon Kinesis load testing with Locust

Post Syndicated from Luis Morales original https://aws.amazon.com/blogs/big-data/perform-amazon-kinesis-load-testing-with-locust/

Building a streaming data solution requires thorough testing at the scale it will operate in a production environment. Streaming applications operating at scale often handle large volumes of up to GBs per second, and it’s challenging for developers to simulate high-traffic Amazon Kinesis-based applications to generate such load easily.

Amazon Kinesis Data Streams and Amazon Kinesis Data Firehose are capable of capturing and storing terabytes of data per hour from numerous sources. Creating Kinesis data streams or Firehose delivery streams is straightforward through the AWS Management Console, AWS Command Line Interface (AWS CLI), or Kinesis API. However, generating a continuous stream of test data requires a custom process or script to run continuously. Although the Amazon Kinesis Data Generator (KDG) provides a user-friendly UI for this purpose, it has some limitations, such as bandwidth constraints and increased round trip latency. (For more information on the KDG, refer to Test Your Streaming Data Solution with the New Amazon Kinesis Data Generator.)

To overcome these limitations, this post describes how to use Locust, a modern load testing framework, to conduct large-scale load testing for a more comprehensive evaluation of the streaming data solution.

Overview

This project emits temperature sensor readings via Locust to Kinesis. We set up the Amazon Elastic Compute Cloud (Amazon EC2) Locust instance via the AWS Cloud Development Kit (AWS CDK) to load test Kinesis-based applications. You can access the Locust dashboard to perform and observe the load test and connect via Session Manager, a capability of AWS Systems Manager, for configuration changes. The following diagram illustrates this architecture.

Architecture overview

In our testing with the largest recommended instance (c7g.16xlarge), the setup was capable of emitting over 1 million events per second to Kinesis data streams in on-demand capacity mode, with a batch size (simulated users per Locust user) of 500. You can find more details on what this means and how to configure the load test later in this post.

Locust overview

Locust is an open-source, scriptable, and scalable performance testing tool that allows you to define user behavior using Python code. It offers an easy-to-use interface, making it developer-friendly and highly expandable. With its distributed and scalable design, Locust can simulate millions of simultaneous users to mimic real user behavior during a performance test.

Each Locust user represents a scenario or a specific set of actions that a real user might perform on your system. When you run a performance test with Locust, you can specify the number of concurrent Locust users you want to simulate, and Locust will create an instance for each user, allowing you to assess the performance and behavior of your system under different user loads.

For more information on Locust, refer to the Locust documentation.

Prerequisites

To get started, clone or download the code from the GitHub repository.

Test locally

To test Locust out locally first before deploying it to the cloud, you have to install the necessary Python dependencies. If you’re new to Python, refer the README for more information on getting started.

Navigate to the load-test directory and run the following code:

pip install -r requirements.txt

To send events to a Kinesis data stream from your local machine, you will need to have AWS credentials. For more information, refer to Configuration and credential file settings.

To perform the test locally, stay in the load-test directory and run the following code:

locust -f locust-load-test.py

You can now access the Locust dashboard via http://0.0.0.0:8089/. Enter the number of Locust users, the spawn rate (users added per second), and the target Amazon Kinesis data stream name for Host. By default, it deploys the Kinesis data stream DemoStream that you can use for testing.

Locust Dashboard - Enter details

To see the generated events logged, run the following command, which filters only Locust and root logs (for example, no Botocore logs):

locust -f locust-load-test.py --loglevel DEBUG 2&gt;&amp;1 | grep -E "(locust|root)"

Set up resources with the AWS CDK

The GitHub repository contains the AWS CDK code to create all the necessary resources for the load test. This removes opportunities for manual error, increases efficiency, and ensures consistent configurations over time. To deploy the resources, complete the following steps:

  1. If not already downloaded, clone the GitHub repository to your local computer using the following command:
git clone https://github.com/aws-samples/amazon-kinesis-load-testing-with-locust
  1. Download and install the latest Node.js.
  2. Navigate to the root folder of the project and run the following command to install the latest version of AWS CDK:
npm install -g aws-cdk
  1. Install the necessary dependencies:
npm install
  1. Run cdk bootstrap to initialize the AWS CDK environment in your AWS account. Replace your AWS account ID and Region before running the following command:
cdk bootstrap

To learn more about the bootstrapping process, refer to Bootstrapping.

  1. After the dependencies are installed, you can run the following command to deploy the stack of the AWS CDK template, which sets up the infrastructure within 5 minutes:
cdk deploy

The template sets up the Locust EC2 test instance, which is by default a c7g.xlarge instance, which at the time of publishing costs approximately $0.145 per hour in us-east-1. To find the most accurate pricing information, see Amazon EC2 On-Demand Pricing. You can find more details on how to change your instance size according to your scale of load testing later in this post.

It’s crucial to consider that the expenses incurred during load testing are not solely attributed to EC2 instance costs, but also heavily influenced by data transfer costs.

Accessing the Locust dashboard

You can access the dashboard by using the AWS CDK output KinesisLocustLoadTestingStack.locustdashboardurl to open the dashboard, for example http://1.2.3.4:8089.

The Locust dashboard is password protected. By default, it’s set to user name locust-user and password locust-dashboard-pwd.

With the default configuration, you can achieve up to 15,000 emitted events per second. Enter the number of Locust users (times the batch size), the spawn rate (users added per second), and the target Kinesis data stream name for Host.

Locust Dashboard - Enter details

After you have started the load test, you can look at the load test on the Charts tab.

Locust Dashboard - Charts

You can also monitor the load test on the Kinesis Data Streams console by navigating to the stream that you are load testing. If you used the default settings, navigate to DemoStream. On the detail page, choose the Monitoring tab to see the ingested load.

Kinesis Data Streams - Monitoring

Adapt workloads

By default, this project generates random temperature sensor readings for every sensor with the following format:

{
    "sensorId": "bfbae19c-2f0f-41c2-952b-5d5bc6e001f1_1",
    "temperature": 147.24,
    "status": "OK",
    "timestamp": 1675686126310
}

The project comes packaged with Faker, which you can use to adapt the payload to your needs. You just have to update the generate_sensor_reading function in the locust-load-test.py file:

class SensorAPIUser(KinesisBotoUser):
    # ...

    def generate_sensor_reading(self, sensor_id, sensor_reading):
        current_temperature = round(10 + random.random() * 170, 2)

        if current_temperature > 160:
            status = "ERROR"
        elif current_temperature > 140 or random.randrange(1, 100) > 80:
            status = random.choice(["WARNING", "ERROR"])
        else:
            status = "OK"

        return {
            'sensorId': f"{sensor_id}_{sensor_reading}",
            'temperature': current_temperature,
            'status': status,
            'timestamp': round(time.time()*1000)
        }

    # ...

Change configurations

After the initial deployment of the load testing tool, you can change configuration in two ways:

  1. Connect to the EC2 instance, make any configuration and code changes, and restart the Locust process
  2. Change the configuration and load testing code locally and redeploy it via cdk deploy

The first option helps you iterate more quickly on the remote instance without a need to redeploy. The latter uses the infrastructure as code (IaC) approach and makes sure that your configuration changes can be committed to your source control system. For a fast development cycle, it’s recommended to test your load test configuration locally first, connect to your instance to apply the changes, and after successful implementation, codify it as part of your IaC repository and then redeploy.

Locust is created on the EC2 instance as a systemd service and can therefore be controlled with systemctl. If you want to change the configuration of Locust as needed without redeploying the stack, you can connect to the instance via Systems Manager, navigate to the project directory on /usr/local/load-test, change the locust.env file, and restart the service by running sudo systemctl restart locust.

Large-scale load testing

This setup is capable of emitting over 1 million events per second to Kinesis data stream, with a batch size of 500 and 64 secondaries on a c7g.16xlarge.

To achieve peak performance with Locust and Kinesis, keep the following in mind:

  • Instance size – Your performance is bound by the underlying EC2 instance, so refer to EC2 instance type for more information about scaling. To set the correct instance size, you can configure the instance size in the file kinesis-locust-load-testing.ts.
  • Number of secondaries – Locust benefits from a distributed setup. Therefore, the setup spins up a primary, which does the coordination, and multiple secondaries, which do the actual work. To fully take advantage of the cores, you should specify one secondary per core. You can configure the number in the locust.env file.
  • Batch size – The amount of Kinesis data stream events you can send per Locust user is limited due to the resource overhead of switching Locust users and threads. To overcome this, you can configure a batch size to define how much users are simulated per Locust user. These are sent as a Kinesis data stream put_records call. You can configure the number in the locust.env file.

This setup is capable of emitting over 1 million events per second to the Kinesis data stream, with a batch size of 500 and 64 secondaries on a c7g.16xlarge instance.

Locust Dashboard - Large Scale Load Test Charts

You can observe this on the Monitoring tab for the Kinesis data stream as well.

Kinesis Data Stream - Large Scale Load Test Monitoring

Clean up

In order to not incur any unnecessary costs, delete the stack by running the following code:

cdk destroy

Summary

Kinesis is already popular for its ease of use among users building streaming applications. With this load testing capability using Locust, you can now test your workloads in a more straightforward and faster way. Visit the GitHub repo to embark on your testing journey.

The project is licensed under the Apache 2.0 license, providing the freedom to clone and modify it according to your needs. Furthermore, you can contribute to the project by submitting issues or pull requests via GitHub, fostering collaboration and improvement in the testing ecosystem.

About the author

Luis Morales works as Senior Solutions Architect with digital native businesses to support them in constantly reinventing themselves in the cloud. He is passionate about software engineering, cloud-native distributed systems, test-driven development, and all things code and security

Create an Apache Hudi-based near-real-time transactional data lake using AWS DMS, Amazon Kinesis, AWS Glue streaming ETL, and data visualization using Amazon QuickSight

Post Syndicated from Raj Ramasubbu original https://aws.amazon.com/blogs/big-data/create-an-apache-hudi-based-near-real-time-transactional-data-lake-using-aws-dms-amazon-kinesis-aws-glue-streaming-etl-and-data-visualization-using-amazon-quicksight/

With the rapid growth of technology, more and more data volume is coming in many different formats—structured, semi-structured, and unstructured. Data analytics on operational data at near-real time is becoming a common need. Due to the exponential growth of data volume, it has become common practice to replace read replicas with data lakes to have better scalability and performance. In most real-world use cases, it’s important to replicate the data from the relational database source to the target in real time. Change data capture (CDC) is one of the most common design patterns to capture the changes made in the source database and reflect them to other data stores.

We recently announced support for streaming extract, transform, and load (ETL) jobs in AWS Glue version 4.0, a new version of AWS Glue that accelerates data integration workloads in AWS. AWS Glue streaming ETL jobs continuously consume data from streaming sources, clean and transform the data in-flight, and make it available for analysis in seconds. AWS also offers a broad selection of services to support your needs. A database replication service such as AWS Database Migration Service (AWS DMS) can replicate the data from your source systems to Amazon Simple Storage Service (Amazon S3), which commonly hosts the storage layer of the data lake. Although it’s straightforward to apply updates on a relational database management system (RDBMS) that backs an online source application, it’s difficult to apply this CDC process on your data lakes. Apache Hudi, an open-source data management framework used to simplify incremental data processing and data pipeline development, is a good option to solve this problem.

This post demonstrates how to apply CDC changes from Amazon Relational Database Service (Amazon RDS) or other relational databases to an S3 data lake, with flexibility to denormalize, transform, and enrich the data in near-real time.

Solution overview

We use an AWS DMS task to capture near-real-time changes in the source RDS instance, and use Amazon Kinesis Data Streams as a destination of the AWS DMS task CDC replication. An AWS Glue streaming job reads and enriches changed records from Kinesis Data Streams and performs an upsert into the S3 data lake in Apache Hudi format. Then we can query the data with Amazon Athena visualize it in Amazon QuickSight. AWS Glue natively supports continuous write operations for streaming data to Apache Hudi-based tables.

The following diagram illustrates the architecture used for this post, which is deployed through an AWS CloudFormation template.

Prerequisites

Before you get started, make sure you have the following prerequisites:

Source data overview

To illustrate our use case, we assume a data analyst persona who is interested in analyzing near-real-time data for sport events using the table ticket_activity. An example of this table is shown in the following screenshot.

Apache Hudi connector for AWS Glue

For this post, we use AWS Glue 4.0, which already has native support for the Hudi framework. Hudi, an open-source data lake framework, simplifies incremental data processing in data lakes built on Amazon S3. It enables capabilities including time travel queries, ACID (Atomicity, Consistency, Isolation, Durability) transactions, streaming ingestion, CDC, upserts, and deletes.

Set up resources with AWS CloudFormation

This post includes a CloudFormation template for a quick setup. You can review and customize it to suit your needs.

The CloudFormation template generates the following resources:

  • An RDS database instance (source).
  • An AWS DMS replication instance, used to replicate the data from the source table to Kinesis Data Streams.
  • A Kinesis data stream.
  • Four AWS Glue Python shell jobs:
    • rds-ingest-rds-setup-<CloudFormation Stack name> – creates one source table called ticket_activity on Amazon RDS.
    • rds-ingest-data-initial-<CloudFormation Stack name> – Sample data is automatically generated at random by the Faker library and loaded to the ticket_activity table.
    • rds-ingest-data-incremental-<CloudFormation Stack name> – Ingests new ticket activity data into the source table ticket_activity continuously. This job simulates customer activity.
    • rds-upsert-data-<CloudFormation Stack name> – Upserts specific records in the source table ticket_activity. This job simulates administrator activity.
  • AWS Identity and Access Management (IAM) users and policies.
  • An Amazon VPC, a public subnet, two private subnets, internet gateway, NAT gateway, and route tables.
    • We use private subnets for the RDS database instance and AWS DMS replication instance.
    • We use the NAT gateway to have reachability to pypi.org to use the MySQL connector for Python from the AWS Glue Python shell jobs. It also provides reachability to Kinesis Data Streams and an Amazon S3 API endpoint

To set up these resources, you must have the following prerequisites:

The following diagram illustrates the architecture of our provisioned resources.

To launch the CloudFormation stack, complete the following steps:

  1. Sign in to the AWS CloudFormation console.
  2. Choose Launch Stack
  3. Choose Next.
  4. For S3BucketName, enter the name of your new S3 bucket.
  5. For VPCCIDR, enter a CIDR IP address range that doesn’t conflict with your existing networks.
  6. For PublicSubnetCIDR, enter the CIDR IP address range within the CIDR you gave for VPCCIDR.
  7. For PrivateSubnetACIDR and PrivateSubnetBCIDR, enter the CIDR IP address range within the CIDR you gave for VPCCIDR.
  8. For SubnetAzA and SubnetAzB, choose the subnets you want to use.
  9. For DatabaseUserName, enter your database user name.
  10. For DatabaseUserPassword, enter your database user password.
  11. Choose Next.
  12. On the next page, choose Next.
  13. Review the details on the final page and select I acknowledge that AWS CloudFormation might create IAM resources with custom names.
  14. Choose Create stack.

Stack creation can take about 20 minutes.

Set up an initial source table

The AWS Glue job rds-ingest-rds-setup-<CloudFormation stack name> creates a source table called event on the RDS database instance. To set up the initial source table in Amazon RDS, complete the following steps:

  1. On the AWS Glue console, choose Jobs in the navigation pane.
  2. Choose rds-ingest-rds-setup-<CloudFormation stack name> to open the job.
  3. Choose Run.
  4. Navigate to the Runs tab and wait for Run status to show as SUCCEEDED.

This job will only create the one table, ticket_activity, in the MySQL instance (DDL). See the following code:

CREATE TABLE ticket_activity (
ticketactivity_id INT NOT NULL AUTO_INCREMENT PRIMARY KEY,
sport_type VARCHAR(256) NOT NULL,
start_date DATETIME NOT NULL,
location VARCHAR(256) NOT NULL,
seat_level VARCHAR(256) NOT NULL,
seat_location VARCHAR(256) NOT NULL,
ticket_price INT NOT NULL,
customer_name VARCHAR(256) NOT NULL,
email_address VARCHAR(256) NOT NULL,
created_at DATETIME NOT NULL,
updated_at DATETIME NOT NULL )

Ingest new records

In this section, we detail the steps to ingest new records. Implement following steps to star the execution of the jobs.

Start data ingestion to Kinesis Data Streams using AWS DMS

To start data ingestion from Amazon RDS to Kinesis Data Streams, complete the following steps:

  1. On the AWS DMS console, choose Database migration tasks in the navigation pane.
  2. Select the task rds-to-kinesis-<CloudFormation stack name>.
  3. On the Actions menu, choose Restart/Resume.
  4. Wait for the status to show as Load complete and Replication ongoing.

The AWS DMS replication task ingests data from Amazon RDS to Kinesis Data Streams continuously.

Start data ingestion to Amazon S3

Next, to start data ingestion from Kinesis Data Streams to Amazon S3, complete the following steps:

  1. On the AWS Glue console, choose Jobs in the navigation pane.
  2. Choose streaming-cdc-kinesis2hudi-<CloudFormation stack name> to open the job.
  3. Choose Run.

Do not stop this job; you can check the run status on the Runs tab and wait for it to show as Running.

Start the data load to the source table on Amazon RDS

To start data ingestion to the source table on Amazon RDS, complete the following steps:

  1. On the AWS Glue console, choose Jobs in the navigation pane.
  2. Choose rds-ingest-data-initial-<CloudFormation stack name> to open the job.
  3. Choose Run.
  4. Navigate to the Runs tab and wait for Run status to show as SUCCEEDED.

Validate the ingested data

After about 2 minutes from starting the job, the data should be ingested into the Amazon S3. To validate the ingested data in the Athena, complete the following steps:

  1. On the Athena console, complete the following steps if you’re running an Athena query for the first time:
    • On the Settings tab, choose Manage.
    • Specify the stage directory and the S3 path where Athena saves the query results.
    • Choose Save.

  1. On the Editor tab, run the following query against the table to check the data:
SELECT * FROM "database_<account_number>_hudi_cdc_demo"."ticket_activity" limit 10;

Note that AWS Cloud Formation will create the database with the account number as database_<your-account-number>_hudi_cdc_demo.

Update existing records

Before you update the existing records, note down the ticketactivity_id value of a record from the ticket_activity table. Run the following SQL using Athena. For this post, we use ticketactivity_id = 46 as an example:

SELECT * FROM "database_<account_number>_hudi_cdc_demo"."ticket_activity" limit 10;

To simulate a real-time use case, update the data in the source table ticket_activity on the RDS database instance to see that the updated records are replicated to Amazon S3. Complete the following steps:

  1. On the AWS Glue console, choose Jobs in the navigation pane.
  2. Choose rds-ingest-data-incremental-<CloudFormation stack name> to open the job.
  3. Choose Run.
  4. Choose the Runs tab and wait for Run status to show as SUCCEEDED.

To upsert the records in the source table, complete the following steps:

  1. On the AWS Glue console, choose Jobs in the navigation pane.
  2. Choose the job rds-upsert-data-<CloudFormation stack name>.
  3. On the Job details tab, under Advanced properties, for Job parameters, update the following parameters:
    • For Key, enter --ticketactivity_id.
    • For Value, replace 1 with one of the ticket IDs you noted above (for this post, 46).

  1. Choose Save.
  2. Choose Run and wait for the Run status to show as SUCCEEDED.

This AWS Glue Python shell job simulates a customer activity to buy a ticket. It updates a record in the source table ticket_activity on the RDS database instance using the ticket ID passed in the job argument --ticketactivity_id. It will update ticket_price=500 and updated_at with the current timestamp.

To validate the ingested data in Amazon s3, run the same query from Athena and check the ticket_activity value you noted earlier to observe the ticket_price and updated_at fields:

SELECT * FROM "database_<account_number>_hudi_cdc_demo"."ticket_activity" where ticketactivity_id = 46 ;

Visualize the data in QuickSight

After you have the output file generated by the AWS Glue streaming job in the S3 bucket, you can use QuickSight to visualize the Hudi data files. QuickSight is a scalable, serverless, embeddable, ML-powered business intelligence (BI) service built for the cloud. QuickSight lets you easily create and publish interactive BI dashboards that include ML-powered insights. QuickSight dashboards can be accessed from any device and seamlessly embedded into your applications, portals, and websites.

Build a QuickSight dashboard

To build a QuickSight dashboard, complete the following steps:

  1. Open the QuickSight console.

You’re presented with the QuickSight welcome page. If you haven’t signed up for QuickSight, you may have to complete the signup wizard. For more information, refer to Signing up for an Amazon QuickSight subscription.

After you have signed up, QuickSight presents a “Welcome wizard.” You can view the short tutorial, or you can close it.

  1. On the QuickSight console, choose your user name and choose Manage QuickSight.
  2. Choose Security & permissions, then choose Manage.
  3. Select Amazon S3 and select the buckets that you created earlier with AWS CloudFormation.
  4. Select Amazon Athena.
  5. Choose Save.
  6. If you changed your Region during the first step of this process, change it back to the Region that you used earlier during the AWS Glue jobs.

Create a dataset

Now that you have QuickSight up and running, you can create your dataset. Complete the following steps:

  1. On the QuickSight console, choose Datasets in the navigation pane.
  2. Choose New dataset.
  3. Choose Athena.
  4. For Data source name, enter a name (for example, hudi-blog).
  5. Choose Validate.
  6. After the validation is successful, choose Create data source.
  7. For Database, choose database_<your-account-number>_hudi_cdc_demo.
  8. For Tables, select ticket_activity.
  9. Choose Select.
  10. Choose Visualize.
  11. Choose hour and then ticket_activity_id to get the count of ticket_activity_id by hour.

Clean up

To clean up your resources, complete the following steps:

  1. Stop the AWS DMS replication task rds-to-kinesis-<CloudFormation stack name>.
  2. Navigate to the RDS database and choose Modify.
  3. Deselect Enable deletion protection, then choose Continue.
  4. Stop the AWS Glue streaming job streaming-cdc-kinesis2redshift-<CloudFormation stack name>.
  5. Delete the CloudFormation stack.
  6. On the QuickSight dashboard, choose your user name, then choose Manage QuickSight.
  7. Choose Account settings, then choose Delete account.
  8. Choose Delete account to confirm.
  9. Enter confirm and choose Delete account.

Conclusion

In this post, we demonstrated how you can stream data—not only new records, but also updated records from relational databases—to Amazon S3 using an AWS Glue streaming job to create an Apache Hudi-based near-real-time transactional data lake. With this approach, you can easily achieve upsert use cases on Amazon S3. We also showcased how to visualize the Apache Hudi table using QuickSight and Athena. As a next step, refer to the Apache Hudi performance tuning guide for a high-volume dataset. To learn more about authoring dashboards in QuickSight, check out the QuickSight Author Workshop.

About the Authors

Raj Ramasubbu is a Sr. Analytics Specialist Solutions Architect focused on big data and analytics and AI/ML with Amazon Web Services. He helps customers architect and build highly scalable, performant, and secure cloud-based solutions on AWS. Raj provided technical expertise and leadership in building data engineering, big data analytics, business intelligence, and data science solutions for over 18 years prior to joining AWS. He helped customers in various industry verticals like healthcare, medical devices, life science, retail, asset management, car insurance, residential REIT, agriculture, title insurance, supply chain, document management, and real estate.

Rahul Sonawane is a Principal Analytics Solutions Architect at AWS with AI/ML and Analytics as his area of specialty.

Sundeep Kumar is a Sr. Data Architect, Data Lake at AWS, helping customers build data lake and analytics platform and solutions. When not building and designing data lakes, Sundeep enjoys listening music and playing guitar.

Amazon Kinesis Data Streams on-demand capacity mode now scales up to 1 GB/second ingest capacity

Post Syndicated from Nihar Sheth original https://aws.amazon.com/blogs/big-data/amazon-kinesis-data-streams-on-demand-capacity-mode-now-scales-up-to-1-gb-second-ingest-capacity/

Amazon Kinesis Data Streams is a serverless data streaming service that makes it easy to capture, process, and store streaming data at any scale. As customers collect and stream more types of data, they have asked for simpler, elastic data streams that can handle variable and unpredictable data traffic. In November 2021, Amazon Web Services launched the on-demand capacity mode for Kinesis Data Streams, which is capable of serving gigabytes of write and read throughput per minute and helps reduce the operational pain point of manually updating data stream capacity. You can create a new on-demand data stream or convert an existing data stream to on-demand mode with a single click and never have to provision and manage servers, storage, or throughput. By default, on-demand capacity mode can automatically scale up to 200 MB/s of write throughput.

We were encouraged by customers’ adoption of on-demand capacity mode, but as customers scaled their workloads, some ran into the 200 MB/s data ingestion limit and asked for a solution. The team worked backward from customer feedback to raise that limit. As of March 2023, Kinesis Data Streams supports an increased on-demand write throughput limit to 1 GB/s, a five-times increase from the current limit of 200 MB/s. It’s like having a truly serverless and elastic data streaming service that works for all your use cases. If you require an increase in capacity, you can contact AWS Support to enable on-demand streams to scale up to 1 GB/s write throughput for each requested account. You pay for throughput consumed rather than for provisioned resources, making it easier to balance costs and performance. Overall, if your data volume can spike unpredictably or you don’t want to manage the number of shards, use on-demand streams.

In this post, we explore how to use Kinesis Data Streams on-demand scaling and best practices to build an efficient data-streaming solution. We discuss different scenarios to avoid write throughput exceptions and scale ingest capacity of Kinesis Data Streams to 1 GB/s in on-demand capacity mode.

Kinesis Data Streams on-demand scaling

A shard serves as a base throughput unit of Kinesis Data Streams. A shard supports 1 MB/s and 1,000 records/s for writes and 2 MB/s for reads. The shard limits ensure predictable performance, making it easy to design and operate a highly reliable data streaming workflow. In on-demand capacity mode, scaling happens at the individual shard level. When the average ingest shard utilization reaches 50% (0.5 MB/s or 500 records/s) in 1 minute, then a shard is split into two shards. If you use random values as a partition key, all shards of the stream will have even traffic, and they will be scaled at the same time. If you use a business-specific key as a partition key, the shards will have uneven traffic. In that scenario, only the shards exceeding an average of 50% utilization will be scaled. Depending upon the number of shards being scaled, it will take up to 15 minutes to split the shards.

When we create a new Kinesis data stream in on-demand capacity mode, by default, Kinesis Data Streams provisions four shards, which provides 4 MB/s write and 8 MB/s read throughput. As the workload ramps up, Kinesis Data Streams increases the number of shards in the stream by monitoring ingest throughput at the shard level. The 4 MB/s default ingest throughput and scaling at shard level in on-demand capacity mode works for most use cases. However, in some specific scenarios, producers may face WriteThroughputExceeded and Rate Exceeded errors, even in on-demand capacity mode. We discuss a few of these scenarios in the following sections and strategies to avoid these errors.

You can create and save record templates and easily send data to Kinesis Data Streams using the Amazon Kinesis Data Generator (KDG) to test the streaming data solution. Alternatively, you can also use the modern load testing framework Locust to run large-scale Kinesis Data Streams load testing. For this post, we use the Locust tool to produce and ingest messages in Kinesis Data Streams for our different use cases.

Scenario 1: A baseline ingest throughput greater than 4 MB/s is needed

To simulate this scenario, run the following AWS Command Line Interface (AWS CLI) command to create the kds-od-default-shards data stream in on-demand capacity mode:

aws kinesis create-stream --stream-name kds-od-default-shards --stream-mode-details StreamMode=ON_DEMAND --region us-east-1

When the kds-od-default-shards data stream is active, run following AWS CLI command to check the number of shards in the data stream:

aws kinesis describe-stream-summary --stream-name kds-od-default-shards --region us-east-1

You can observe that the OpenShardCount value is 4, which means the kds-od-default-shards data stream has an ingest capacity of 4 MB/s.

Next, we use the Locust tool to set the baseline to approximately 25 MB/s records. As displayed in the following Amazon CloudWatch metrics graph, records are getting throttled for the first couple of minutes. Then the kds-od-default-shards data stream scales the number of shards to support 25 MB/s ingest throughput, and records stop getting throttled. You can also rerun the describe-stream-summary AWS CLI command to check the increased number of shards in the data stream.

BDB-3047-scenario-1-incoming-data

BDB-3047-scenario-1-record-throttle

In a scenario where we know our ingest throughput baseline (25 MB/s) ahead of the time and we don’t want to observe any write throttles, we can create a stream in provisioned mode by specifying the number of shards (30), as shown in the following AWS CLI command (make sure to delete kds-od-default-shards manually from the Kinesis Data Streams console before running the following command):

aws kinesis create-stream --stream-name kds-od-default-shards --stream-mode-details StreamMode=PROVISIONED --shard-count 30 --region us-east-1

When the kds-od-default-shards data stream is active, run the following AWS CLI command to convert the data stream’s capacity mode to on-demand:

aws kinesis update-stream-mode --stream-arn arn:aws:kinesis:us-east-1:<AccountId>:stream/kds-od-default-shards --stream-mode-details StreamMode=ON_DEMAND --region us-east-1

Next, we send 25 MB/s records to the kds-od-default-shards data stream. As displayed in the following CloudWatch metrics graph, we can observe no write throttles, and the kds-od-default-shards data stream scales the number of shards to handle the increase in ingest volume.

BDB-3047-scenario-1-incoming-data1

BDB-3047-scenario-1-record-throttle1

After we send 25 MB/s traffic to the data stream for some time, we can run following AWS CLI command to see that the OpenShardCount value is increased to more than 30 now:

aws kinesis describe-stream-summary --stream-name kds-od-default-shards --region us-east-1

Scenario 2: A significant ingestion spike is expected, which needs ingest throughput greater than the number of shards in the stream

To simulate the scenario, run the following AWS CLI command to create the kds-od-significant-spike data stream in on-demand capacity mode:

aws kinesis create-stream --stream-name kds-od-significant-spike --stream-mode-details StreamMode=ON_DEMAND --region us-east-1

As mentioned earlier, by default, the kds-od-significant-spike data stream will have four shards initially because this stream is created in on-demand mode. When the data stream is active, we send 4 MB/s ingest throughput initially and grow the ingest throughput by 30–50% every 5–10 minutes. As displayed in the following CloudWatch metrics graph, the kds-od-significant-spike data stream scales the number of shards to handle the increase in ingest volume.

After approximately 15 minutes, run the following AWS CLI command to find the OpenShardCount value (x) of the kds-od-significant-spike data stream. Then send (x * 2) MB/s ingest throughput in the data stream for 2–3 minutes and reduced ingest throughput to the prior level:

aws kinesis describe-stream-summary --stream-name kds-od-significant-spike --region us-east-1

As displayed in the following CloudWatch metrics graph, the records are getting throttled for a few minutes, and then the throttling goes away.

BDB-3047-scenario-2-incoming-data

BDB-3047-scenario-2-record-throttle

Typically, we face a significant spike scenario when running planned events, such as shopping holidays and product launches. To handle such scenarios, we can proactively change capacity mode from on-demand to provisioned. We can configure the number of shards and pick the ingest capacity we anticipate. After we successfully scale the number of shards to our desired peak capacity in provisioned capacity mode, we can change the capacity mode back to on-demand mode.

Scenario 3: A single partition key starts pushing more than 1 MB/s

Partition keys are used to segregate and route records to different shards of a stream. A partition key is specified by the data producer while adding data to the data stream. For example, let’s assume we have a stream with two shards (shard 1 and shard 2). We can configure the data producer to use two partition keys (key A and key B) so that all records with key A are added to shard 1 and all records with key B are added to shard 2. Choosing a partition key is a very important decision, and we should carefully pick the partition key to ensure equal distribution of records across all the shards of the stream. Messages tied to a single partition key A will be sent to a single shard (shard 1), and at any given instance, messages tied to a single partition key A cannot be distributed across different shards. As mentioned earlier, by default, one shard supports 1 MB/s and 1,000 records/s for writes, and we may end up with an edge case scenario where we are trying to push more than 1 MB/s for a specific partition key. In this scenario, producers will continue to experience throttles and keep retrying indefinitely.

To simulate the scenario, run the following AWS CLI command to create the kds-od-partition-key-throttle data stream in on-demand capacity mode:

aws kinesis create-stream --stream-name kds-od-partition-key-throttle --stream-mode-details StreamMode=ON_DEMAND --region us-east-1

As mentioned earlier, by default, the data stream will have four shards initially because this stream is created in on-demand mode. When the data stream is active, we send 1.5 MB/s ingest throughput continuously for the specific partition key A. As displayed in the following CloudWatch metrics graph, we can observe that throttling continues from a single shard even if we are sending 1.5 MB/s ingest throughput, and the kds-od-partition-key-throttle data stream has an overall ingest capacity of 4 MB/s.

BDB-3047-scenario-3-incoming-data

BDB-3047-scenario-3-record-throttle

To avoid this scenario, we should carefully pick our partition key and ensure that this specific partition key won’t be continuously sending more than 1 MB/s ingest throughput in the data stream.

Scale the ingest capacity of Kinesis Data Streams to 1 GB/s in on-demand capacity mode

To test, we start with approximately 100 MB/s baseline ingest throughput to Kinesis Data Streams in on-demand capacity mode, then we increase ingest throughput rate by 30–50% every 5–10 minutes using Locust load testing tool.

To set up the scenario, first create the kds-od-1gb-stream data stream in provisioned capacity mode and provide a value of 120 for the provisioned shards field:

aws kinesis create-stream --stream-name kds-od-1gb-stream --stream-mode-details StreamMode=PROVISIONED --shard-count 120 --region us-east-1

When the kds-od-1gb-stream data stream is active, switch its capacity mode to on-demand, as shown in the following code. When we change capacity mode from provisioned to on-demand, the shard count (120) remains the same for the data stream even in on-demand capacity mode.

aws kinesis update-stream-mode --stream-arn arn:aws:kinesis:us-east-1:<AccountId>:stream/kds-od-1gb-stream --stream-mode-details StreamMode=ON_DEMAND --region us-east-1

When the kds-od-1gb-stream data stream is in on-demand mode, start the experiment. We send approximately 100 MB/s baseline ingest throughput using the Locust tool and increase 30–50% ingest throughput every 5–10 minutes. As displayed in the following CloudWatch metrics graph, the kds-od-1gb-stream data stream seamlessly scaled to 1 GB/s in on-demand capacity mode. We can also observe that the producers didn’t encounter any write throttles while the data stream was scaling in on-demand capacity mode.

BDB-3047-scale-to-1-GB

Clean up

To avoid ongoing costs, delete all the data streams that you created as part of this post using the Kinesis Data Streams console.

Conclusion

This post demonstrated the on-demand scaling policy of Kinesis Data Streams with a few scenarios using best practices and showed how to scale ingest capacity to 1 GB/s in on-demand capacity mode. You can have an on-demand write throughput limit that is five times larger than the previous limit of 200 MB/s. Choose on-demand mode if you create new data streams with unknown workloads, have unpredictable application traffic, or prefer not to manage capacity. You can switch between on-demand and provisioned capacity modes two times per 24-hour rolling period. Please leave any feedback in the comments section.

About the Authors

Nihar Sheth is a Senior Product Manager on the Amazon Kinesis Data Streams team at Amazon Web Services. He is passionate about developing intuitive product experiences that solve complex customer problems and enable customers to achieve their business goals.

Pratik Patel is Sr. Technical Account Manager and streaming analytics specialist. He works with AWS customers and provides ongoing support and technical guidance to help plan and build solutions using best practices and proactively keep customers’ AWS environments operationally healthy.

Nisha Dekhtawala is a Partner Solutions Architect and data analytics specialist. She works with global consulting partners as their trusted advisor, providing technical guidance and support in building Well-Architected innovative industry solutions.

A side-by-side comparison of Apache Spark and Apache Flink for common streaming use cases

Post Syndicated from Deepthi Mohan original https://aws.amazon.com/blogs/big-data/a-side-by-side-comparison-of-apache-spark-and-apache-flink-for-common-streaming-use-cases/

Apache Flink and Apache Spark are both open-source, distributed data processing frameworks used widely for big data processing and analytics. Spark is known for its ease of use, high-level APIs, and the ability to process large amounts of data. Flink shines in its ability to handle processing of data streams in real-time and low-latency stateful computations. Both support a variety of programming languages, scalable solutions for handling large amounts of data, and a wide range of connectors. Historically, Spark started out as a batch-first framework and Flink began as a streaming-first framework.

In this post, we share a comparative study of streaming patterns that are commonly used to build stream processing applications, how they can be solved using Spark (primarily Spark Structured Streaming) and Flink, and the minor variations in their approach. Examples cover code snippets in Python and SQL for both frameworks across three major themes: data preparation, data processing, and data enrichment. If you are a Spark user looking to solve your stream processing use cases using Flink, this post is for you. We do not intend to cover the choice of technology between Spark and Flink because it’s important to evaluate both frameworks for your specific workload and how the choice fits in your architecture; rather, this post highlights key differences for use cases that both these technologies are commonly considered for.

Apache Flink offers layered APIs that offer different levels of expressiveness and control and are designed to target different types of use cases. The three layers of API are Process Functions (also known as the Stateful Stream Processing API), DataStream, and Table and SQL. The Stateful Stream Processing API requires writing verbose code but offers the most control over time and state, which are core concepts in stateful stream processing. The DataStream API supports Java, Scala, and Python and offers primitives for many common stream processing operations, as well as a balance between code verbosity or expressiveness and control. The Table and SQL APIs are relational APIs that offer support for Java, Scala, Python, and SQL. They offer the highest abstraction and intuitive, SQL-like declarative control over data streams. Flink also allows seamless transition and switching across these APIs. To learn more about Flink’s layered APIs, refer to layered APIs.

Apache Spark Structured Streaming offers the Dataset and DataFrames APIs, which provide high-level declarative streaming APIs to represent static, bounded data as well as streaming, unbounded data. Operations are supported in Scala, Java, Python, and R. Spark has a rich function set and syntax with simple constructs for selection, aggregation, windowing, joins, and more. You can also use the Streaming Table API to read tables as streaming DataFrames as an extension to the DataFrames API. Although it’s hard to draw direct parallels between Flink and Spark across all stream processing constructs, at a very high level, we could say Spark Structured Streaming APIs are equivalent to Flink’s Table and SQL APIs. Spark Structured Streaming, however, does not yet (at the time of this writing) offer an equivalent to the lower-level APIs in Flink that offer granular control of time and state.

Both Flink and Spark Structured Streaming (referenced as Spark henceforth) are evolving projects. The following table provides a simple comparison of Flink and Spark capabilities for common streaming primitives (as of this writing).

. Flink Spark
Row-based processing Yes Yes
User-defined functions Yes Yes
Fine-grained access to state Yes, via DataStream and low-level APIs No
Control when state eviction occurs Yes, via DataStream and low-level APIs No
Flexible data structures for state storage and querying Yes, via DataStream and low-level APIs No
Timers for processing and stateful operations Yes, via low level APIs No

In the following sections, we cover the greatest common factors so that we can showcase how Spark users can relate to Flink and vice versa. To learn more about Flink’s low-level APIs, refer to Process Function. For the sake of simplicity, we cover the four use cases in this post using the Flink Table API. We use a combination of Python and SQL for an apples-to-apples comparison with Spark.

Data preparation

In this section, we compare data preparation methods for Spark and Flink.

Reading data

We first look at the simplest ways to read data from a data stream. The following sections assume the following schema for messages:

symbol: string,
price: int,
timestamp: timestamp,
company_info:
{
    name: string,
    employees_count: int
}

Reading data from a source in Spark Structured Streaming

In Spark Structured Streaming, we use a streaming DataFrame in Python that directly reads the data in JSON format:

spark = ...  # spark session

# specify schema
stock_ticker_schema = ...

# Create a streaming DataFrame
df = spark.readStream \
    .format("kafka") \
    .option("kafka.bootstrap.servers", "mybroker1:port") \
    .option("topic", "stock_ticker") \
    .load()
    .select(from_json(col("value"), stock_ticker_schema).alias("ticker_data")) \
    .select(col("ticker_data.*"))

Note that we have to supply a schema object that captures our stock ticker schema (stock_ticker_schema). Compare this to the approach for Flink in the next section.

Reading data from a source using Flink Table API

For Flink, we use the SQL DDL statement CREATE TABLE. You can specify the schema of the stream just like you would any SQL table. The WITH clause allows us to specify the connector to the data stream (Kafka in this case), the associated properties for the connector, and data format specifications. See the following code:

# Create table using DDL

CREATE TABLE stock_ticker (
  symbol string,
  price INT,
  timestamp TIMESTAMP(3),
  company_info STRING,
  WATERMARK FOR timestamp AS timestamp - INTERVAL '3' MINUTE
) WITH (
 'connector' = 'kafka',
 'topic' = 'stock_ticker',
 'properties.bootstrap.servers' = 'mybroker1:port',
 'properties.group.id' = 'testGroup',
 'format' = 'json',
 'json.fail-on-missing-field' = 'false',
 'json.ignore-parse-errors' = 'true'
)

JSON flattening

JSON flattening is the process of converting a nested or hierarchical JSON object into a flat, single-level structure. This converts multiple levels of nesting into an object where all the keys and values are at the same level. Keys are combined using a delimiter such as a period (.) or underscore (_) to denote the original hierarchy. JSON flattening is useful when you need to work with a more simplified format. In both Spark and Flink, nested JSONs can be complicated to work with and may need additional processing or user-defined functions to manipulate. Flattened JSONs can simplify processing and improve performance due to reduced computational overhead, especially with operations like complex joins, aggregations, and windowing. In addition, flattened JSONs can help in easier debugging and troubleshooting data processing pipelines because there are fewer levels of nesting to navigate.

JSON flattening in Spark Structured Streaming

JSON flattening in Spark Structured Streaming requires you to use the select method and specify the schema that you need flattened. JSON flattening in Spark Structured Streaming involves specifying the nested field name that you’d like surfaced to the top-level list of fields. In the following example, company_info is a nested field and within company_info, there’s a field called company_name. With the following query, we’re flattening company_info.name to company_name:

stock_ticker_df = ...  # Streaming DataFrame w/ schema shown above

stock_ticker_df.select("symbol", "timestamp", "price", "company_info.name" as "company_name")

JSON flattening in Flink

In Flink SQL, you can use the JSON_VALUE function. Note that you can use this function only in Flink versions equal to or greater than 1.14. See the following code:

SELECT
   symbol,
   timestamp,
   price,
   JSON_VALUE(company_info, 'lax $.name' DEFAULT NULL ON EMPTY) AS company_name
FROM
   stock_ticker

The term lax in the preceding query has to do with JSON path expression handling in Flink SQL. For more information, refer to System (Built-in) Functions.

Data processing

Now that you have read the data, we can look at a few common data processing patterns.

Deduplication

Data deduplication in stream processing is crucial for maintaining data quality and ensuring consistency. It enhances efficiency by reducing the strain on the processing from duplicate data and helps with cost savings on storage and processing.

Spark Streaming deduplication query

The following code snippet is related to a Spark Streaming DataFrame named stock_ticker. The code performs an operation to drop duplicate rows based on the symbol column. The dropDuplicates method is used to eliminate duplicate rows in a DataFrame based on one or more columns.

stock_ticker = ...  # Streaming DataFrame w/ schema shown above

stock_ticker.dropDuplicates("symbol")

Flink deduplication query

The following code shows the Flink SQL equivalent to deduplicate data based on the symbol column. The query retrieves the first row for each distinct value in the symbol column from the stock_ticker stream, based on the ascending order of proctime:

SELECT symbol, timestamp, price
FROM (
  SELECT *,
    ROW_NUMBER() OVER (PARTITION BY symbol ORDER BY proctime ASC) AS row_num
  FROM stock_ticker)
WHERE row_num = 1

Windowing

Windowing in streaming data is a fundamental construct to process data within specifications. Windows commonly have time bounds, number of records, or other criteria. These time bounds bucketize continuous unbounded data streams into manageable chunks called windows for processing. Windows help in analyzing data and gaining insights in real time while maintaining processing efficiency. Analyses or operations are performed on constantly updating streaming data within a window.

There are two common time-based windows used both in Spark Streaming and Flink that we will detail in this post: tumbling and sliding windows. A tumbling window is a time-based window that is a fixed size and doesn’t have any overlapping intervals. A sliding window is a time-based window that is a fixed size and moves forward in fixed intervals that can be overlapping.

Spark Streaming tumbling window query

The following is a Spark Streaming tumbling window query with a window size of 10 minutes:

stock_ticker = ...  # Streaming DataFrame w/ schema shown above

# Get max stock price in tumbling window
# of size 10 minutes
visitsByWindowAndUser = visits
   .withWatermark("timestamp", "3 minutes")
   .groupBy(
      window(stock_ticker.timestamp, "10 minutes"),
      stock_ticker.symbol)
   .max(stock_ticker.price)

Flink Streaming tumbling window query

The following is an equivalent tumbling window query in Flink with a window size of 10 minutes:

SELECT symbol, MAX(price)
  FROM TABLE(
    TUMBLE(TABLE stock_ticker, DESCRIPTOR(timestamp), INTERVAL '10' MINUTES))
  GROUP BY ticker;

Spark Streaming sliding window query

The following is a Spark Streaming sliding window query with a window size of 10 minutes and slide interval of 5 minutes:

stock_ticker = ...  # Streaming DataFrame w/ schema shown above

# Get max stock price in sliding window
# of size 10 minutes and slide interval of size
# 5 minutes

visitsByWindowAndUser = visits
   .withWatermark("timestamp", "3 minutes")
   .groupBy(
      window(stock_ticker.timestamp, "10 minutes", "5 minutes"),
      stock_ticker.symbol)
   .max(stock_ticker.price)

Flink Streaming sliding window query

The following is a Flink sliding window query with a window size of 10 minutes and slide interval of 5 minutes:

SELECT symbol, MAX(price)
  FROM TABLE(
    HOP(TABLE stock_ticker, DESCRIPTOR(timestamp), INTERVAL '5' MINUTES, INTERVAL '10' MINUTES))
  GROUP BY ticker;

Handling late data

Both Spark Structured Streaming and Flink support event time processing, where a field within the payload can be used for defining time windows as distinct from the wall clock time of the machines doing the processing. Both Flink and Spark use watermarking for this purpose.

Watermarking is used in stream processing engines to handle delays. A watermark is like a timer that sets how long the system can wait for late events. If an event arrives and is within the set time (watermark), the system will use it to update a request. If it’s later than the watermark, the system will ignore it.

In the preceding windowing queries, you specify the lateness threshold in Spark using the following code:

.withWatermark("timestamp", "3 minutes")

This means that any records that are 3 minutes late as tracked by the event time clock will be discarded.

In contrast, with the Flink Table API, you can specify an analogous lateness threshold directly in the DDL:

WATERMARK FOR timestamp AS timestamp - INTERVAL '3' MINUTE

Note that Flink provides additional constructs for specifying lateness across its various APIs.

Data enrichment

In this section, we compare data enrichment methods with Spark and Flink.

Calling an external API

Calling external APIs from user-defined functions (UDFs) is similar in Spark and Flink. Note that your UDF will be called for every record processed, which can result in the API getting called at a very high request rate. In addition, in production scenarios, your UDF code often gets run in parallel across multiple nodes, further amplifying the request rate.

For the following code snippets, let’s assume that the external API call entails calling the function:

response = my_external_api(request)

External API call in Spark UDF

The following code uses Spark:

class Predict(ScalarFunction):
def open(self, function_context):

with open("resources.zip/resources/model.pkl", "rb") as f:
self.model = pickle.load(f)

def eval(self, x):
return self.model.predict(x)

External API call in Flink UDF

For Flink, assume we define the UDF callExternalAPIUDF, which takes as input the ticker symbol symbol and returns enriched information about the symbol via a REST endpoint. We can then register and call the UDF as follows:

callExternalAPIUDF = udf(callExternalAPIUDF(), result_type=DataTypes.STRING())

SELECT
    symbol, 
    callExternalAPIUDF(symbol) as enriched_symbol
FROM stock_ticker;

Flink UDFs provide an initialization method that gets run one time (as opposed to one time per record processed).

Note that you should use UDFs judiciously as an improperly implemented UDF can cause your job to slow down, cause backpressure, and eventually stall your stream processing application. It’s advisable to use UDFs asynchronously to maintain high throughput, especially for I/O-bound use cases or when dealing with external resources like databases or REST APIs. To learn more about how you can use asynchronous I/O with Apache Flink, refer to Enrich your data stream asynchronously using Amazon Kinesis Data Analytics for Apache Flink.

Conclusion

Apache Flink and Apache Spark are both rapidly evolving projects and provide a fast and efficient way to process big data. This post focused on the top use cases we commonly encountered when customers wanted to see parallels between the two technologies for building real-time stream processing applications. We’ve included samples that were most frequently requested at the time of this writing. Let us know if you’d like more examples in the comments section.

About the author

Deepthi Mohan is a Principal Product Manager on the Amazon Kinesis Data Analytics team.

Karthi Thyagarajan was a Principal Solutions Architect on the Amazon Kinesis team.

Near-real-time analytics using Amazon Redshift streaming ingestion with Amazon Kinesis Data Streams and Amazon DynamoDB

Post Syndicated from Poulomi Dasgupta original https://aws.amazon.com/blogs/big-data/near-real-time-analytics-using-amazon-redshift-streaming-ingestion-with-amazon-kinesis-data-streams-and-amazon-dynamodb/

Amazon Redshift is a fully managed, scalable cloud data warehouse that accelerates your time to insights with fast, easy, and secure analytics at scale. Tens of thousands of customers rely on Amazon Redshift to analyze exabytes of data and run complex analytical queries, making it the widely used cloud data warehouse. You can run and scale analytics in seconds on all your data without having to manage your data warehouse infrastructure.

You can use the Amazon Redshift streaming ingestion capability to update your analytics databases in near-real time. Amazon Redshift streaming ingestion simplifies data pipelines by letting you create materialized views directly on top of data streams. With this capability in Amazon Redshift, you can use SQL (Structured Query Language) to connect to and directly ingest data from data streams, such as Amazon Kinesis Data Streams or Amazon Managed Streaming for Apache Kafka (Amazon MSK) data streams, and pull data directly to Amazon Redshift.

In this post, we discuss a solution that uses Amazon Redshift streaming ingestion to provide near-real-time analytics.

Overview of solution

We walk through an example pipeline to ingest data from an Amazon DynamoDB source table in near-real time using Kinesis Data Streams in combination with Amazon Redshift streaming ingestion. We also walk through using PartiQL in Amazon Redshift to unnest nested JSON documents and build fact and dimension tables that are used in your data warehouse refresh. The solution uses Kinesis Data Streams to capture item-level changes from an application DynamoDB table.

As shown in the following reference architecture, DynamoDB table data changes are streamed into Amazon Redshift through Kinesis Data Streams and Amazon Redshift streaming ingestion for near-real-time analytics dashboard visualization using Amazon QuickSight.

The process flow includes the following steps:

  1. Create a Kinesis data stream and turn on the data stream from DynamoDB to capture item-level changes in your DynamoDB table.
  2. Create a streaming materialized view in your Amazon Redshift cluster to consume live streaming data from the data stream.
  3. The streaming data gets ingested into a JSON payload. Use a combination of a PartiQL statement and dot notation to unnest the JSON document into data columns of a staging table in Amazon Redshift.
  4. Create fact and dimension tables in the Amazon Redshift cluster and keep loading the latest data at regular intervals from the staging table using transformation logic.
  5. Establish connectivity between a QuickSight dashboard and Amazon Redshift to deliver visualization and insights.

Prerequisites

You must have the following:

Set up a Kinesis data stream

To configure your Kinesis data stream, complete the following steps:

  1. Create a Kinesis data stream called demo-data-stream. For instructions, refer to Step 1 in Set up streaming ETL pipelines.

Configure the stream to capture changes from the DynamoDB table.

  1. On the DynamoDB console, choose Tables in the navigation pane.
  2. Open your table.
  3. On the Exports and streams tab, choose Turn on under Amazon Kinesis data stream details.

  1. For Destination Kinesis data stream, choose demo-data-stream.
  2. Choose Turn on stream.

Item-level changes in the DynamoDB table should now be flowing to the Kinesis data stream.

  1. To verify if the data is entering the stream, on the Kinesis Data Streams console, open demo-data-stream.
  2. On the Monitoring tab, find the PutRecord success – average (Percent) and PutRecord – sum (Bytes) metrics to validate record ingestion.

Set up streaming ingestion

To set up streaming ingestion, complete the following steps:

  1. Set up the AWS Identity and Access Management (IAM) role and trust policy required for streaming ingestion. For instructions, refer to Steps 1 and 2 in Getting started with streaming ingestion from Amazon Kinesis Data Streams.
  2. Launch the Query Editor v2 from the Amazon Redshift console or use your preferred SQL client to connect to your Amazon Redshift cluster for the next steps.
  3. Create an external schema:
CREATE EXTERNAL SCHEMA demo_schema
FROM KINESIS
IAM_ROLE { default | 'iam-role-arn' };
  1. To use case-sensitive identifiers, set enable_case_sensitive_identifier to true at either the session or cluster level.
  2. Create a materialized view to consume the stream data and store stream records in semi-structured SUPER format:
CREATE MATERIALIZED VIEW demo_stream_vw AS
    SELECT approximate_arrival_timestamp,
    partition_key,
    shard_id,
    sequence_number,
    json_parse(kinesis_data) as payload    
    FROM demo_schema."demo-data-stream";
  1. Refresh the view, which triggers Amazon Redshift to read from the stream and load data into the materialized view:
REFRESH MATERIALIZED VIEW demo_stream_vw;

You can also set your streaming materialized view to use auto refresh capabilities. This will automatically refresh your materialized view as data arrives in the stream. See CREATE MATERIALIZED VIEW for instructions on how to create a materialized view with auto refresh.

Unnest the JSON document

The following is a sample of a JSON document that was ingested from the Kinesis data stream to the payload column of the streaming materialized view demo_stream_vw:

{
  "awsRegion": "us-east-1",
  "eventID": "6d24680a-6d12-49e2-8a6b-86ffdc7306c1",
  "eventName": "INSERT",
  "userIdentity": null,
  "recordFormat": "application/json",
  "tableName": "sample-dynamoDB",
  "dynamodb": {
    "ApproximateCreationDateTime": 1657294923614,
    "Keys": {
      "pk": {
        "S": "CUSTOMER#CUST_123"
      },
      "sk": {
        "S": "TRANSACTION#2022-07-08T23:59:59Z#CUST_345"
      }
    },
    "NewImage": {
      "completionDateTime": {
        "S": "2022-07-08T23:59:59Z"
      },
      "OutofPockPercent": {
        "N": 50.00
      },
      "calculationRequirements": {
        "M": {
          "dependentIds": {
            "L": [
              {
                "M": {
                  "sk": {
                    "S": "CUSTOMER#2022-07-08T23:59:59Z#CUST_567"
                  },
                  "pk": {
                    "S": "CUSTOMER#CUST_123"
                  }
                }
              },
              {
                "M": {
                  "sk": {
                    "S": "CUSTOMER#2022-07-08T23:59:59Z#CUST_890"
                  },
                  "pk": {
                    "S": "CUSTOMER#CUST_123"
                  }
                }
              }
            ]
          }
        }
      },
      "Event": {
        "S": "SAMPLE"
      },
      "Provider": {
        "S": "PV-123"
      },
      "OutofPockAmount": {
        "N": 1000
      },
      "lastCalculationDateTime": {
        "S": "2022-07-08T00:00:00Z"
      },
      "sk": {
        "S": "CUSTOMER#2022-07-08T23:59:59Z#CUST_567"
      },
      "OutofPockMax": {
        "N": 2000
      },
      "pk": {
        "S": "CUSTOMER#CUST_123"
      }
    },
    "SizeBytes": 694
  },
  "eventSource": "aws:dynamodb"
}

We can use dot notation to unnest the JSON document. But in addition to that, we should use a PartiQL statement to handle arrays if applicable. For example, in the preceding JSON document, there is an array under the element:

"dynamodb"."NewImage"."calculationRequirements"."M"."dependentIds"."L".

The following SQL query uses a combination of dot notation and a PartiQL statement to unnest the JSON document:

select 
substring(a."payload"."dynamodb"."Keys"."pk"."S"::varchar, position('#' in "payload"."dynamodb"."Keys"."pk"."S"::varchar)+1) as Customer_ID,
substring(a."payload"."dynamodb"."Keys"."sk"."S"::varchar, position('#TRANSACTION' in "payload"."dynamodb"."Keys"."sk"."S"::varchar)+1) as Transaction_ID,
substring(b."M"."sk"."S"::varchar, position('#CUSTOMER' in b."M"."sk"."S"::varchar)+1) Dependent_ID,
a."payload"."dynamodb"."NewImage"."OutofPockMax"."N"::int as OutofPocket_Max,
a."payload"."dynamodb"."NewImage"."OutofPockPercent"."N"::decimal(5,2) as OutofPocket_Percent,
a."payload"."dynamodb"."NewImage"."OutofPockAmount"."N"::int as OutofPock_Amount,
a."payload"."dynamodb"."NewImage"."Provider"."S"::varchar as Provider,
a."payload"."dynamodb"."NewImage"."completionDateTime"."S"::timestamptz as Completion_DateTime,
a."payload"."eventName"::varchar Event_Name,
a.approximate_arrival_timestamp
from demo_stream_vw a
left outer join a."payload"."dynamodb"."NewImage"."calculationRequirements"."M"."dependentIds"."L" b on true;

The query unnests the JSON document to the following result set.

Precompute the result set using a materialized view

Optionally, to precompute and store the unnested result set from the preceding query, you can create a materialized view and schedule it to refresh at regular intervals. In this post, we maintain the preceding unnested data in a materialized view called mv_demo_super_unnest, which will be refreshed at regular intervals and used for further processing.

To capture the latest data from the DynamoDB table, the Amazon Redshift streaming materialized view needs to be refreshed at regular intervals, and then the incremental data should be transformed and loaded into the final fact and dimension table. To avoid reprocessing the same data, a metadata table can be maintained at Amazon Redshift to keep track of each ELT process with status, start time, and end time, as explained in the following section.

Maintain an audit table in Amazon Redshift

The following is a sample DDL of a metadata table that is maintained for each process or job:

create table MetaData_ETL
(
JobName varchar(100),
StartDate timestamp, 
EndDate timestamp, 
Status varchar(50)
);

The following is a sample initial entry of the metadata audit table that can be maintained at job level. The insert statement is the initial entry for the ELT process to load the Customer_Transaction_Fact table:

insert into MetaData_ETL 
values
('Customer_Transaction_Fact_Load', current_timestamp, current_timestamp,'Ready' );

Build a fact table with the latest data

In this post, we demonstrate the loading of a fact table using specific transformation logic. We are skipping the dimension table load, which uses similar logic.

As a prerequisite, create the fact and dimension tables in a preferred schema. In following example, we create the fact table Customer_Transaction_Fact in Amazon Redshift:

CREATE TABLE public.Customer_Transaction_Fact (
Transaction_ID character varying(500),
Customer_ID character varying(500),
OutofPocket_Percent numeric(5,2),
OutofPock_Amount integer,
OutofPocket_Max integer,
Provider character varying(500),
completion_datetime timestamp
);

Transform data using a stored procedure

We load this fact table from the unnested data using a stored procedure. For more information, refer to Creating stored procedures in Amazon Redshift.

Note that in this sample use case, we are using transformation logic to identify and load the latest value of each column for a customer transaction.

The stored procedure contains the following components:

  • In the first step of the stored procedure, the job entry in the MetaData_ETL table is updated to change the status to Running and StartDate as the current timestamp, which indicates that the fact load process is starting.
  • Refresh the materialized view mv_demo_super_unnest, which contains the unnested data.
  • In the following example, we load the fact table Customer_Transaction_Fact using the latest data from the streaming materialized view based on the column approximate_arrival_timestamp, which is available as a system column in the streaming materialized view. The value of approximate_arrival_timestamp is set when a Kinesis data stream successfully receives and stores a record.
  • The following logic in the stored procedure checks if the approximate_arrival_timestamp in mv_demo_super_unnest is greater than the EndDate timestamp in the MetaData_ETL audit table, so that it can only process the incremental data.
  • Additionally, while loading the fact table, we identify the latest non-null value of each column for every Transaction_ID depending on the order of the approximate_arrival_timestamp column using the rank and min
  • The transformed data is loaded into the intermediate staging table
  • The impacted records with the same Transaction_ID values are deleted and reloaded into the Customer_Transaction_Fact table from the staging table
  • In the last step of the stored procedure, the job entry in the MetaData_ETL table is updated to change the status to Complete and EndDate as the current timestamp, which indicates that the fact load process has completed successfully.

See the following code:

CREATE OR REPLACE PROCEDURE SP_Customer_Transaction_Fact()
AS $$
BEGIN

set enable_case_sensitive_identifier to true;

--Update metadata audit table entry to indicate that the fact load process is running
update MetaData_ETL
set status = 'Running',
StartDate = getdate()
where JobName = 'Customer_Transaction_Fact_Load';

refresh materialized view mv_demo_super_unnest;

drop table if exists Customer_Transaction_Fact_Stg;

--Create latest record by Merging records based on approximate_arrival_timestamp
create table Customer_Transaction_Fact_Stg as
select 
m.Transaction_ID,
min(case when m.rank_Customer_ID =1 then m.Customer_ID end) Customer_ID,
min(case when m.rank_OutofPocket_Percent =1 then m.OutofPocket_Percent end) OutofPocket_Percent,
min(case when m.rank_OutofPock_Amount =1 then m.OutofPock_Amount end) OutofPock_Amount,
min(case when m.rank_OutofPocket_Max =1 then m.OutofPocket_Max end) OutofPocket_Max,
min(case when m.rank_Provider =1 then m.Provider end) Provider,
min(case when m.rank_Completion_DateTime =1 then m.Completion_DateTime end) Completion_DateTime
from
(
select *,
rank() over(partition by Transaction_ID order by case when mqp.Customer_ID is not null then 1 end, approximate_arrival_timestamp desc ) rank_Customer_ID,
rank() over(partition by Transaction_ID order by case when mqp.OutofPocket_Percent is not null then 1 end, approximate_arrival_timestamp desc ) rank_OutofPocket_Percent,
rank() over(partition by Transaction_ID order by case when mqp.OutofPock_Amount is not null then 1 end, approximate_arrival_timestamp  desc )  rank_OutofPock_Amount,
rank() over(partition by Transaction_ID order by case when mqp.OutofPocket_Max is not null then 1 end, approximate_arrival_timestamp desc ) rank_OutofPocket_Max,
rank() over(partition by Transaction_ID order by case when mqp.Provider is not null then 1 end, approximate_arrival_timestamp  desc ) rank_Provider,
rank() over(partition by Transaction_ID order by case when mqp.Completion_DateTime is not null then 1 end, approximate_arrival_timestamp desc )  rank_Completion_DateTime
from mv_demo_super_unnest mqp
where upper(mqp.event_Name) <> 'REMOVE' and mqp.approximate_arrival_timestamp > (select mde.EndDate from MetaData_ETL mde where mde.JobName = 'Customer_Transaction_Fact_Load') 
) m
group by m.Transaction_ID 
order by m.Transaction_ID
;

--Delete only impacted Transaction_ID from Fact table
delete from Customer_Transaction_Fact  
where Transaction_ID in ( select mqp.Transaction_ID from Customer_Transaction_Fact_Stg mqp);

--Insert latest records from staging table to Fact table
insert into Customer_Transaction_Fact
select * from Customer_Transaction_Fact_Stg; 

--Update metadata audit table entry to indicate that the fact load process is completed
update MetaData_ETL
set status = 'Complete',
EndDate = getdate()
where JobName = 'Customer_Transaction_Fact_Load';
END;
$$ LANGUAGE plpgsql;

Additional considerations for implementation

There are several additional capabilities that you could utilize to modify this solution to meet your needs. Many customers utilize multiple AWS accounts, and it’s common that the Kinesis data stream may be in a different AWS account than the Amazon Redshift data warehouse. If this is the case, you can utilize an Amazon Redshift IAM role that assumes a role in the Kinesis data stream AWS account in order to read from the data stream. For more information, refer to Cross-account streaming ingestion for Amazon Redshift.

Another common use case is that you need to schedule the refresh of your Amazon Redshift data warehouse jobs so that the data warehouse’s data is continuously updated. To do this, you can utilize Amazon EventBridge to schedule the jobs in your data warehouse to run on a regular basis. For more information, refer to Creating an Amazon EventBridge rule that runs on a schedule. Another option is to use Amazon Redshift Query Editor v2 to schedule the refresh. For details, refer to Scheduling a query with query editor v2.

If you have a requirement to load data from a DynamoDB table with existing data, refer to Loading data from DynamoDB into Amazon Redshift.

For more information on Amazon Redshift streaming ingestion capabilities, refer to Real-time analytics with Amazon Redshift streaming ingestion.

Clean up

To avoid unnecessary charges, clean up any resources that you built as part of this architecture that are no longer in use. This includes dropping the materialized view, stored procedure, external schema, and tables created as part of this post. Additionally, make sure you delete the DynamoDB table and delete the Kinesis data stream.

Conclusion

After following the solution in this post, you’re now able to build near-real-time analytics using Amazon Redshift streaming ingestion. We showed how you can ingest data from a DynamoDB source table using a Kinesis data stream in order to refresh your Amazon Redshift data warehouse. With the capabilities presented in this post, you should be able to increase the refresh rate of your Amazon Redshift data warehouse in order to provide the most up-to-date data in your data warehouse for your use case.

About the authors

Poulomi Dasgupta is a Senior Analytics Solutions Architect with AWS. She is passionate about helping customers build cloud-based analytics solutions to solve their business problems. Outside of work, she likes travelling and spending time with her family.

Matt Nispel is an Enterprise Solutions Architect at AWS. He has more than 10 years of experience building cloud architectures for large enterprise companies. At AWS, Matt helps customers rearchitect their applications to take full advantage of the cloud. Matt lives in Minneapolis, Minnesota, and in his free time enjoys spending time with friends and family.

Dan Dressel is a Senior Analytics Specialist Solutions Architect at AWS. He is passionate about databases, analytics, machine learning, and architecting solutions. In his spare time, he enjoys spending time with family, nature walking, and playing foosball.

Best practices for implementing event-driven architectures in your organization

Post Syndicated from Emanuele Levi original https://aws.amazon.com/blogs/architecture/best-practices-for-implementing-event-driven-architectures-in-your-organization/

Event-driven architectures (EDA) are made up of components that detect business actions and changes in state, and encode this information in event notifications. Event-driven patterns are becoming more widespread in modern architectures because:

  • they are the main invocation mechanism in serverless patterns.
  • they are the preferred pattern for decoupling microservices, where asynchronous communications and event persistence are paramount.
  • they are widely adopted as a loose-coupling mechanism between systems in different business domains, such as third-party or on-premises systems.

Event-driven patterns have the advantage of enabling team independence through the decoupling and decentralization of responsibilities. This decentralization trend in turn, permits companies to move with unprecedented agility, enhancing feature development velocity.

In this blog, we’ll explore the crucial components and architectural decisions you should consider when adopting event-driven patterns, and provide some guidance on organizational structures.

Division of responsibilities

The communications flow in EDA (see What is EDA?) is initiated by the occurrence of an event. Most production-grade event-driven implementations have three main components, as shown in Figure 1: producers, message brokers, and consumers.

Three main components of an event-driven architecture

Figure 1. Three main components of an event-driven architecture

Producers, message brokers, and consumers typically assume the following roles:

Producers

Producers are responsible for publishing the events as they happen. They are the owners of the event schema (data structure) and semantics (meaning of the fields, such as the meaning of the value of an enum field). As this is the only contract (coupling) between producers and the downstream components of the system, the schema and its semantics are crucial in EDA. Producers are responsible for implementing a change management process, which involves both non-breaking and breaking changes. With introduction of breaking changes, consumers are able to negotiate the migration process with producers.

Producers are “consumer agnostic”, as their boundary of responsibility ends when an event is published.

Message brokers

Message brokers are responsible for the durability of the events, and will keep an event available for consumption until it is successfully processed. Message brokers ensure that producers are able to publish events for consumers to consume, and they regulate access and permissions to publish and consume messages.

Message brokers are largely “events agnostic”, and do not generally access or interpret the event content. However, some systems provide a routing mechanism based on the event payload or metadata.

Consumers

Consumers are responsible for consuming events, and own the semantics of the effect of events. Consumers are usually bounded to one business context. This means the same event will have different effect semantics for different consumers. Crucial architectural choices when implementing a consumer involve the handling of unsuccessful message deliveries or duplicate messages. Depending on the business interpretation of the event, when recovering from failure a consumer might permit duplicate events, such as with an idempotent consumer pattern.

Crucially, consumers are “producer agnostic”, and their boundary of responsibility begins when an event is ready for consumption. This allows new consumers to onboard into the system without changing the producer contracts.

Team independence

In order to enforce the division of responsibilities, companies should organize their technical teams by ownership of producers, message brokers, and consumers. Although the ownership of producers and consumers is straightforward in an EDA implementation, the ownership of the message broker may not be. Different approaches can be taken to identify message broker ownership depending on your organizational structure.

Decentralized ownership

Ownership of the message broker in a decentralized ownership organizational structure

Figure 2. Ownership of the message broker in a decentralized ownership organizational structure

In a decentralized ownership organizational structure (see Figure 2), the teams producing events are responsible for managing their own message brokers and the durability and availability of the events for consumers.

The adoption of topic fanout patterns based on Amazon Simple Queue Service (SQS) and Amazon Simple Notification Service (SNS) (see Figure 3), can help companies implement a decentralized ownership pattern. A bus-based pattern using Amazon EventBridge can also be similarly utilized (see Figure 4).

Topic fanout pattern based on Amazon SQS and Amazon SNS

Figure 3. Topic fanout pattern based on Amazon SQS and Amazon SNS

Events bus pattern based on Amazon EventBridge

Figure 4. Events bus pattern based on Amazon EventBridge

The decentralized ownership approach has the advantage of promoting team independence, but it is not a fit for every organization. In order to be implemented effectively, a well-established DevOps culture is necessary. In this scenario, the producing teams are responsible for managing the message broker infrastructure and the non-functional requirements standards.

Centralized ownership

Ownership of the message broker in a centralized ownership organizational structure

Figure 5. Ownership of the message broker in a centralized ownership organizational structure

In a centralized ownership organizational structure, a central team (we’ll call it the platform team) is responsible for the management of the message broker (see Figure 5). Having a specialized platform team offers the advantage of standardized implementation of non-functional requirements, such as reliability, availability, and security. One disadvantage is that the platform team is a single point of failure in both the development and deployment lifecycle. This could become a bottleneck and put team independence and operational efficiency at risk.

Streaming pattern based on Amazon MSK and Kinesis Data Streams

Figure 6. Streaming pattern based on Amazon MSK and Kinesis Data Streams

On top of the implementation patterns mentioned in the previous section, the presence of a dedicated team makes it easier to implement streaming patterns. In this case, a deeper understanding on how the data is partitioned and how the system scales is required. Streaming patterns can be implemented using services such as Amazon Managed Streaming for Apache Kafka (MSK) or Amazon Kinesis Data Streams (see Figure 6).

Best practices for implementing event-driven architectures in your organization

The centralized and decentralized ownership organizational structures enhance team independence or standardization of non-functional requirements respectively. However, they introduce possible limits to the growth of the engineering function in a company. Inspired by the two approaches, you can implement a set of best practices which are aimed at minimizing those limitations.

Best practices for implementing event-driven architectures

Figure 7. Best practices for implementing event-driven architectures

  1. Introduce a cloud center of excellence (CCoE). A CCoE standardizes non-functional implementation across engineering teams. In order to promote a strong DevOps culture, the CCoE should not take the form of an external independent team, but rather be a collection of individual members representing the various engineering teams.
  2. Decentralize team ownership. Decentralize ownership and maintenance of the message broker to producing teams. This will maximize team independence and agility. It empowers the team to use the right tool for the right job, as long as they conform to the CCoE guidelines.
  3. Centralize logging standards and observability strategies. Although it is a best practice to decentralize team ownership of the components of an event-driven architecture, logging standards and observability strategies should be centralized and standardized across the engineering function. This centralization provides for end-to-end tracing of requests and events, which are powerful diagnosis tools in case of any failure.

Conclusion

In this post, we have described the main architectural components of an event-driven architecture, and identified the ownership of the message broker as one of the most important architectural choices you can make. We have described a centralized and decentralized organizational approach, presenting the strengths of the two approaches, as well as the limits they impose on the growth of your engineering organization. We have provided some best practices you can implement in your organization to minimize these limitations.

Further reading:
To start your journey building event-driven architectures in AWS, explore the following:

Temporal data lake architecture for benchmark and indices analytics

Post Syndicated from Krishna Gogineni original https://aws.amazon.com/blogs/architecture/temporal-data-lake-architecture-for-benchmark-and-indices-analytics/

Financial trading houses and stock exchanges generate enormous volumes of data in near real-time, making it difficult to perform bi-temporal calculations that yield accurate results. Achieving this requires a processing architecture that can handle large volumes of data during peak bursts, meet strict latency requirements, and scale according to incoming volumes.

In this post, we’ll describe a scenario for an industry leader in the financial services sector and explain how AWS services are used for bi-temporal processing with state management and scale based on variable workloads during the day, all while meeting strict service-level agreement (SLA) requirements.

Problem statement

To design and implement a fully temporal transactional data lake with the repeatable read isolation level for queries is a challenge, particularly with burst events that need the overall architecture to scale accordingly. The data store in the overall architecture needs to record the value history of data at different times, which is especially important for financial data. Financial data can include corporate actions, annual or quarterly reports, or fixed-income securities, like bonds that have variable rates. It’s crucial to be able to correct data inaccuracies during the reporting period.

The example customer seeks a data processing platform architecture to dynamically scale based on the workloads with a capacity of processing 150 million records under 5 minutes. Their platform should be capable of meeting the end-to-end SLA of 15 minutes, from ingestion to reporting, with lowest total cost of ownership. Additionally, managing bi-temporal data requires a database that has critical features, such as ACID (atomicity, consistency, isolation, durability) compliance, time-travel capability, full-schema evolution, partition layout and evolution, rollback to prior versions, and SQL-like query experience.

Solution overview

The solution architecture key building blocks are Amazon Kinesis Data Streams for streaming data, Amazon Kinesis Data Analytics with Apache Flink as processing engine, Flink’s RocksDB for state management, and Apache Iceberg on Amazon Simple Storage Service (Amazon S3) as the storage engine (Figure 1).

End-to-end data-processing architecture

Figure 1. End-to-end data-processing architecture

Data processing

Here’s how it works:

  • A publisher application receives the data from the source systems and publishes data into Kinesis Data Streams using a well-defined JSON format structure.
  • Kinesis Data Streams holds the data for a duration that is configurable so data is not lost and can auto scale based on the data volume ingested.
  • Kinesis Data Analytics runs an Apache Flink application, with state management (RocksDB), to handle bi-temporal calculations. The Apache Flink application consumes data from Kinesis Data Streams and performs the following computations:
    • Transforms the JSON stream into a row-type record, compatible with a SQL table-like structure, resolving nesting and parent–child relationships present within the stream
    • Checks whether the record has already an existing state in in-memory RocksDB or disk attached to Kinesis Data Analytics computational node to avoid read latency from the database, which is critical for meeting the performance requirements
    • Performs bi-temporal calculations and creates the resultant records in an in-memory data structure before invoking the Apache Iceberg sink operator
    • The Apache Flink application sink operator appends the temporal states, expressed as records into existing Apache Iceberg data store. This will comply with key principles of time series data, which is immutable, and the ability to time-travel along with ACID compliance, schema evolution, and partition evolution
  • Kinesis Data Analytics is resilient and provides a no-data-loss capability, with features like periodic checkpoints and savepoints. They are used to store the state management in a secure Amazon S3 location that can be accessed outside of Kinesis Data Analytics. This savepoints mechanism can be used to programmatically to scale the cluster size based on the workloads using time-driven scheduling and AWS Lambda functions.
  • If the time-to-live feature of RocksDB is implemented, old records are stored in Apache Iceberg on Amazon S3. When performing temporal calculations, if the state is not found in memory, data is read from Apache Iceberg into RocksDB and the processing is completed. However, this step is optional and can be circumvented if the Kinesis Data Analytics cluster is initialized with right number of Kinesis processing units to hold the historical information, as per requirements.
  • Because the data is stored in an Apache Iceberg table format in Amazon S3, data is queried using Trino, which supports Apache Iceberg table format.
  • The end user queries data using any SQL tool that supports the Trino query engine.

Apache Iceberg maintenance jobs, such as data compaction, expire snapshot, delete orphan files, can be launched using Amazon Athena to optimize performance out of Apache Iceberg data store. Details of each processing step performed in Apache Flink application are captured using Amazon CloudWatch, which logs all the events.

Scalability

Amazon EventBridge scheduler invokes a Lambda function to scale the Kinesis Data Analytics. Kinesis Data Analytics has a short outage during rescaling that is proportional to the amount of data stored in RocksDB, which is why a state management strategy is necessary for the proper operation of the system.

Figure 2 shows the scaling process, which depicts:

  • Before peak load: The Kinesis Data Analytics cluster is processing off-peak records with minimum configuration before the peak load. A scheduled event is launched from EventBridge that invokes a Lambda function, which shuts down the cluster using the savepoint mechanism and scales up the Kinesis Data Analytics cluster to required Kinesis processing units.
  • During peak load: When the peak data burst happens, the Kinesis Data Analytics cluster is ready to handle the volume of data from Kinesis Data Stream, and processes it within the SLA of 5 minutes.
  • After peak load: A scheduled event from EventBridge invokes a Lambda function to scale down the Kinesis Data Analytics cluster to the minimum configuration that holds the required state for the entire volume of records.
Cluster scaling before, during, and after peak data volume processing

Figure 2. Cluster scaling before, during, and after peak data volume processing

Performance insights

With the discussed architecture, we want to demonstrate that the we are able to meet the SLAs, in terms of performance and processing times. We have taken a subset of benchmarks and indices data and processed the same with the end-to-end architecture. During the process, we observed some very interesting findings, which we would like to share.

Processing time for Apache Iceberg Upsert vs Append operations: During our tests, we expected Upsert operation to be faster than append. But on the contrary, we noticed that Append operations were faster compared to Upsert even though more computations are performed in the Apache Flink application. In our test with 3,500,000 records, Append operation took 1556 seconds while Upsert took 1675 seconds to process the data (Figure 3).

Processing times for Upsert vs. Append

Figure 3. Processing times for Upsert vs. Append

Compute consumption for Apache Iceberg Upsert vs. Append operations: Comparing the compute consumption for 10,000,000 records, we noticed that Append operation was able to process the data in the same amount of time as Upsert operation but with less compute resources. In our tests, we have noted that Append operation only consumed 64 Kinesis processing units, whereas Upsert consumed 78 Kinesis processing units (Figure 4).

Comparing consumption for Upsert vs. Append

Figure 4. Comparing consumption for Upsert vs. Append

Scalability vs performance: To achieve the desired data processing performance, we need a specific configuration of Kinesis processing units, Kinesis Data Streams, and Iceberg parallelism. In our test with the data that we chose, we started with four Kinesis processing units and four Kinesis data streams for data processing. We observed an 80% performance improvement in data processing with 16 Kinesis data processing units. An additional 6% performance improvement was demonstrated when we scaled to 32 Kinesis processing units. When we increased the Kinesis data streams to 16, we observed an additional 2% performance improvement (Figure 5).

Scalability vs. performance

Figure 5. Scalability vs. performance

Data volume processing times for Upsert vs. Append: For this test, we started with 350,000 records of data. When we increased data volume to 3.5M records, we observed that Append performing better than Upsert, demonstrating a five-fold increase in processing time (Figure 6).

Data volume processing times for Upsert vs. Append

Figure 6. Data volume processing times for Upsert vs. Append

Conclusion

The architecture we explored today scales based on the data-volume requirements of the customer and is capable of meeting the end-to-end SLA of 15 minutes, with a potential lowered total cost of ownership. Additionally, the solution is capable of handling high-volume, bi-temporal computations with ACID compliance, time travel, full-schema evolution, partition layout evolution, rollback to prior versions and SQL-like query experience.

Further reading

Migrate from Amazon Kinesis Data Analytics for SQL Applications to Amazon Kinesis Data Analytics Studio

Post Syndicated from Nicholas Tunney original https://aws.amazon.com/blogs/big-data/migrate-from-amazon-kinesis-data-analytics-for-sql-applications-to-amazon-kinesis-data-analytics-studio/

Amazon Kinesis Data Analytics makes it easy to transform and analyze streaming data in real time.

In this post, we discuss why AWS recommends moving from Kinesis Data Analytics for SQL Applications to Amazon Kinesis Data Analytics for Apache Flink to take advantage of Apache Flink’s advanced streaming capabilities. We also show how to use Kinesis Data Analytics Studio to test and tune your analysis before deploying your migrated applications. If you don’t have any Kinesis Data Analytics for SQL applications, this post still provides a background on many of the use cases you’ll see in your data analytics career and how Amazon Data Analytics services can help you achieve your objectives.

Kinesis Data Analytics for Apache Flink is a fully managed Apache Flink service. You only need to upload your application JAR or executable, and AWS will manage the infrastructure and Flink job orchestration. To make things simpler, Kinesis Data Analytics Studio is a notebook environment that uses Apache Flink and allows you to query data streams and develop SQL queries or proof of concept workloads before scaling your application to production in minutes.

We recommend that you use Kinesis Data Analytics for Apache Flink or Kinesis Data Analytics Studio over Kinesis Data Analytics for SQL. This is because Kinesis Data Analytics for Apache Flink and Kinesis Data Analytics Studio offer advanced data stream processing features, including exactly-once processing semantics, event time windows, extensibility using user-defined functions (UDFs) and custom integrations, imperative language support, durable application state, horizontal scaling, support for multiple data sources, and more. These are critical for ensuring accuracy, completeness, consistency, and reliability of data stream processing and are not available with Kinesis Data Analytics for SQL.

Solution overview

For our use case, we use several AWS services to stream, ingest, transform, and analyze sample automotive sensor data in real time using Kinesis Data Analytics Studio. Kinesis Data Analytics Studio allows us to create a notebook, which is a web-based development environment. With notebooks, you get a simple interactive development experience combined with the advanced capabilities provided by Apache Flink. Kinesis Data Analytics Studio uses Apache Zeppelin as the notebook, and uses Apache Flink as the stream processing engine. Kinesis Data Analytics Studio notebooks seamlessly combine these technologies to make advanced analytics on data streams accessible to developers of all skill sets. Notebooks are provisioned quickly and provide a way for you to instantly view and analyze your streaming data. Apache Zeppelin provides your Studio notebooks with a complete suite of analytics tools, including the following:

  • Data visualization
  • Exporting data to files
  • Controlling the output format for easier analysis
  • Ability to turn the notebook into a scalable, production application

Unlike Kinesis Data Analytics for SQL Applications, Kinesis Data Analytics for Apache Flink adds the following SQL support:

  • Joining stream data between multiple Kinesis data streams, or between a Kinesis data stream and an Amazon Managed Streaming for Apache Kafka (Amazon MSK) topic
  • Real-time visualization of transformed data in a data stream
  • Using Python scripts or Scala programs within the same application
  • Changing offsets of the streaming layer

Another benefit of Kinesis Data Analytics for Apache Flink is the improved scalability of the solution once deployed, because you can scale the underlying resources to meet demand. In Kinesis Data Analytics for SQL Applications, scaling is performed by adding more pumps to persuade the application into adding more resources.

In our solution, we create a notebook to access automotive sensor data, enrich the data, and send the enriched output from the Kinesis Data Analytics Studio notebook to an Amazon Kinesis Data Firehose delivery stream for delivery to an Amazon Simple Storage Service (Amazon S3) data lake. This pipeline could further be used to send data to Amazon OpenSearch Service or other targets for additional processing and visualization.

Kinesis Data Analytics for SQL Applications vs. Kinesis Data Analytics for Apache Flink

In our example, we perform the following actions on the streaming data:

  1. Connect to an Amazon Kinesis Data Streams data stream.
  2. View the stream data.
  3. Transform and enrich the data.
  4. Manipulate the data with Python.
  5. Restream the data to a Firehose delivery stream.

To compare Kinesis Data Analytics for SQL Applications with Kinesis Data Analytics for Apache Flink, let’s first discuss how Kinesis Data Analytics for SQL Applications works.

At the root of a Kinesis Data Analytics for SQL application is the concept of an in-application stream. You can think of the in-application stream as a table that holds the streaming data so you can perform actions on it. The in-application stream is mapped to a streaming source such as a Kinesis data stream. To get data into the in-application stream, first set up a source in the management console for your Kinesis Data Analytics for SQL application. Then, create a pump that reads data from the source stream and places it into the table. The pump query runs continuously and feeds the source data into the in-application stream. You can create multiple pumps from multiple sources to feed the in-application stream. Queries are then run on the in-application stream, and results can be interpreted or sent to other destinations for further processing or storage.

The following SQL demonstrates setting up an in-application stream and pump:

CREATE OR REPLACE STREAM "TEMPSTREAM" ( 
   "column1" BIGINT NOT NULL, 
   "column2" INTEGER, 
   "column3" VARCHAR(64));

CREATE OR REPLACE PUMP "SAMPLEPUMP" AS 
INSERT INTO "TEMPSTREAM" ("column1", 
                          "column2", 
                          "column3") 
SELECT STREAM inputcolumn1, 
      inputcolumn2, 
      inputcolumn3
FROM "INPUTSTREAM";

Data can be read from the in-application stream using a SQL SELECT query:

SELECT *
FROM "TEMPSTREAM"

When creating the same setup in Kinesis Data Analytics Studio, you use the underlying Apache Flink environment to connect to the streaming source, and create the data stream in one statement using a connector. The following example shows connecting to the same source we used before, but using Apache Flink:

CREATE TABLE `MY_TABLE` ( 
   "column1" BIGINT NOT NULL, 
   "column2" INTEGER, 
   "column3" VARCHAR(64)
) WITH (
   'connector' = 'kinesis',
   'stream' = sample-kinesis-stream',
   'aws.region' = 'aws-kinesis-region',
   'scan.stream.initpos' = 'LATEST',
   'format' = 'json'
 );

MY_TABLE is now a data stream that will continually receive the data from our sample Kinesis data stream. It can be queried using a SQL SELECT statement:

SELECT column1, 
       column2, 
       column3
FROM MY_TABLE;

Although Kinesis Data Analytics for SQL Applications use a subset of the SQL:2008 standard with extensions to enable operations on streaming data, Apache Flink’s SQL support is based on Apache Calcite, which implements the SQL standard.

It’s also important to mention that Kinesis Data Analytics Studio supports PyFlink and Scala alongside SQL within the same notebook. This allows you to perform complex, programmatic methods on your streaming data that aren’t possible with SQL.

Prerequisites

During this exercise, we set up various AWS resources and perform analytics queries. To follow along, you need an AWS account with administrator access. If you don’t already have an AWS account with administrator access, create one now. The services outlined in this post may incur charges to your AWS account. Make sure to follow the cleanup instructions at the end of this post.

Configure streaming data

In the streaming domain, we’re often tasked with exploring, transforming, and enriching data coming from Internet of Things (IoT) sensors. To generate the real-time sensor data, we employ the AWS IoT Device Simulator. This simulator runs within your AWS account and provides a web interface that lets users launch fleets of virtually connected devices from a user-defined template and then simulate them to publish data at regular intervals to AWS IoT Core. This means we can build a virtual fleet of devices to generate sample data for this exercise.

We deploy the IoT Device Simulator using the following Amazon CloudFront template. It handles creating all the necessary resources in your account.

  1. On the Specify stack details page, assign a name to your solution stack.
  2. Under Parameters, review the parameters for this solution template and modify them as necessary.
  3. For User email, enter a valid email to receive a link and password to log in to the IoT Device Simulator UI.
  4. Choose Next.
  5. On the Configure stack options page, choose Next.
  6. On the Review page, review and confirm the settings. Select the check boxes acknowledging that the template creates AWS Identity and Access Management (IAM) resources.
  7. Choose Create stack.

The stack takes about 10 minutes to install.

  1. When you receive your invitation email, choose the CloudFront link and log in to the IoT Device Simulator using the credentials provided in the email.

The solution contains a prebuilt automotive demo that we can use to begin delivering sensor data quickly to AWS.

  1. On the Device Type page, choose Create Device Type.
  2. Choose Automotive Demo.
  3. The payload is auto populated. Enter a name for your device, and enter automotive-topic as the topic.
  4. Choose Save.

Now we create a simulation.

  1. On the Simulations page, choose Create Simulation.
  2. For Simulation type, choose Automotive Demo.
  3. For Select a device type, choose the demo device you created.
  4. For Data transmission interval and Data transmission duration, enter your desired values.

You can enter any values you like, but use at least 10 devices transmitting every 10 seconds. You’ll want to set your data transmission duration to a few minutes, or you’ll need to restart your simulation several times during the lab.

  1. Choose Save.

Now we can run the simulation.

  1. On the Simulations page, select the desired simulation, and choose Start simulations.

Alternatively, choose View next to the simulation you want to run, then choose Start to run the simulation.

  1. To view the simulation, choose View next to the simulation you want to view.

If the simulation is running, you can view a map with the locations of the devices, and up to 100 of the most recent messages sent to the IoT topic.

We can now check to ensure our simulator is sending the sensor data to AWS IoT Core.

  1. Navigate to the AWS IoT Core console.

Make sure you’re in the same Region you deployed your IoT Device Simulator.

  1. In the navigation pane, choose MQTT Test Client.
  2. Enter the topic filter automotive-topic and choose Subscribe.

As long as you have your simulation running, the messages being sent to the IoT topic will be displayed.

Finally, we can set a rule to route the IoT messages to a Kinesis data stream. This stream will provide our source data for the Kinesis Data Analytics Studio notebook.

  1. On the AWS IoT Core console, choose Message Routing and Rules.
  2. Enter a name for the rule, such as automotive_route_kinesis, then choose Next.
  3. Provide the following SQL statement. This SQL will select all message columns from the automotive-topic the IoT Device Simulator is publishing:
SELECT timestamp, trip_id, VIN, brake, steeringWheelAngle, torqueAtTransmission, engineSpeed, vehicleSpeed, acceleration, parkingBrakeStatus, brakePedalStatus, transmissionGearPosition, gearLeverPosition, odometer, ignitionStatus, fuelLevel, fuelConsumedSinceRestart, oilTemp, location 
FROM 'automotive-topic' WHERE 1=1
  1. Choose Next.
  2. Under Rule Actions, select Kinesis Stream as the source.
  3. Choose Create New Kinesis Stream.

This opens a new window.

  1. For Data stream name, enter automotive-data.

We use a provisioned stream for this exercise.

  1. Choose Create Data Stream.

You may now close this window and return to the AWS IoT Core console.

  1. Choose the refresh button next to Stream name, and choose the automotive-data stream.
  2. Choose Create new role and name the role automotive-role.
  3. Choose Next.
  4. Review the rule properties, and choose Create.

The rule begins routing data immediately.

Set up Kinesis Data Analytics Studio

Now that we have our data streaming through AWS IoT Core and into a Kinesis data stream, we can create our Kinesis Data Analytics Studio notebook.

  1. On the Amazon Kinesis console, choose Analytics applications in the navigation pane.
  2. On the Studio tab, choose Create Studio notebook.
  3. Leave Quick create with sample code selected.
  4. Name the notebook automotive-data-notebook.
  5. Choose Create to create a new AWS Glue database in a new window.
  6. Choose Add database.
  7. Name the database automotive-notebook-glue.
  8. Choose Create.
  9. Return to the Create Studio notebook section.
  10. Choose refresh and choose your new AWS Glue database.
  11. Choose Create Studio notebook.
  12. To start the Studio notebook, choose Run and confirm.
  13. Once the notebook is running, choose the notebook and choose Open in Apache Zeppelin.
  14. Choose Import note.
  15. Choose Add from URL.
  16. Enter the following URL: https://aws-blogs-artifacts-public.s3.amazonaws.com/artifacts/BDB-2461/auto-notebook.ipynb.
  17. Choose Import Note.
  18. Open the new note.

Perform stream analysis

In a Kinesis Data Analytics for SQL application, we add our streaming course via the management console, and then define an in-application stream and pump to stream data from our Kinesis data stream. The in-application stream functions as a table to hold the data and make it available for us to query. The pump takes the data from our source and streams it to our in-application stream. Queries may then be run against the in-application stream using SQL, just as we’d query any SQL table. See the following code:

CREATE OR REPLACE STREAM "AUTOSTREAM" ( 
    `trip_id` CHAR(36),
    `VIN` CHAR(17),
    `brake` FLOAT,
    `steeringWheelAngle` FLOAT,
    `torqueAtTransmission` FLOAT,
    `engineSpeed` FLOAT,
    `vehicleSpeed` FLOAT,
    `acceleration` FLOAT,
    `parkingBrakeStatus` BOOLEAN,
    `brakePedalStatus` BOOLEAN,
    `transmissionGearPosition` VARCHAR(10),
    `gearLeverPosition` VARCHAR(10),
    `odometer` FLOAT,
    `ignitionStatus` VARCHAR(4),
    `fuelLevel` FLOAT,
    `fuelConsumedSinceRestart` FLOAT,
    `oilTemp` FLOAT,
    `location` VARCHAR(100),
    `timestamp` TIMESTAMP(3));

CREATE OR REPLACE PUMP "MYPUMP" AS 
INSERT INTO "AUTOSTREAM" ("trip_id",
    "VIN",
    "brake",
    "steeringWheelAngle",
    "torqueAtTransmission",
    "engineSpeed",
    "vehicleSpeed",
    "acceleration",
    "parkingBrakeStatus",
    "brakePedalStatus",
    "transmissionGearPosition",
    "gearLeverPosition",
    "odometer",
    "ignitionStatus",
    "fuelLevel",
    "fuelConsumedSinceRestart",
    "oilTemp",
    "location",
    "timestamp")
SELECT VIN,
    brake,
    steeringWheelAngle,
    torqueAtTransmission,
    engineSpeed,
    vehicleSpeed,
    acceleration,
    parkingBrakeStatus,
    brakePedalStatus,
    transmissionGearPosition,
    gearLeverPosition,
    odometer,
    ignitionStatus,
    fuelLevel,
    fuelConsumedSinceRestart,
    oilTemp,
    location,
    timestamp
FROM "INPUT_STREAM"

To migrate an in-application stream and pump from our Kinesis Data Analytics for SQL application to Kinesis Data Analytics Studio, we convert this into a single CREATE statement by removing the pump definition and defining a kinesis connector. The first paragraph in the Zeppelin notebook sets up a connector that is presented as a table. We can define columns for all items in the incoming message, or a subset.

Run the statement, and a success result is output in your notebook. We can now query this table using SQL, or we can perform programmatic operations with this data using PyFlink or Scala.

Before performing real-time analytics on the streaming data, let’s look at how the data is currently formatted. To do this, we run a simple Flink SQL query on the table we just created. The SQL used in our streaming application is identical to what is used in a SQL application.

Note that if you don’t see records after a few seconds, make sure that your IoT Device Simulator is still running.

If you’re also running the Kinesis Data Analytics for SQL code, you may see a slightly different result set. This is another key differentiator in Kinesis Data Analytics for Apache Flink, because the latter has the concept of exactly once delivery. If this application is deployed to production and is restarted or if scaling actions occur, Kinesis Data Analytics for Apache Flink ensures you only receive each message once, whereas in a Kinesis Data Analytics for SQL application, you need to further process the incoming stream to ensure you ignore repeat messages that could affect your results.

You can stop the current paragraph by choosing the pause icon. You may see an error displayed in your notebook when you stop the query, but it can be ignored. It’s just letting you know that the process was canceled.

Flink SQL implements the SQL standard, and provides an easy way to perform calculations on the stream data just like you would when querying a database table. A common task while enriching data is to create a new field to store a calculation or conversion (such as from Fahrenheit to Celsius), or create new data to provide simpler queries or improved visualizations downstream. Run the next paragraph to see how we can add a Boolean value named accelerating, which we can easily use in our sink to know if an automobile was currently accelerating at the time the sensor was read. The process here doesn’t differ between Kinesis Data Analytics for SQL and Kinesis Data Analytics for Apache Flink.

You can stop the paragraph from running when you have inspected the new column, comparing our new Boolean value to the FLOAT acceleration column.

Data being sent from a sensor is usually compact to improve latency and performance. Being able to enrich the data stream with external data and enrich the stream, such as additional vehicle information or current weather data, can be very useful. In this example, let’s assume we want to bring in data currently stored in a CSV in Amazon S3, and add a column named color that reflects the current engine speed band.

Apache Flink SQL provides several source connectors for AWS services and other sources. Creating a new table like we did in our first paragraph but instead using the filesystem connector permits Flink to directly connect to Amazon S3 and read our source data. Previously in Kinesis Data Analytics for SQL Applications, you couldn’t add new references inline. Instead, you defined S3 reference data and added it to your application configuration, which you could then use as a reference in a SQL JOIN.

NOTE: If you are not using the us-east-1 region, you can download the csv and place the object your own S3 bucket.  Reference the csv file as s3a://<bucket-name>/<key-name>

Building on the last query, the next paragraph performs a SQL JOIN on our current data and the new lookup source table we created.

Now that we have an enriched data stream, we restream this data. In a real-world scenario, we have many choices on what to do with our data, such as sending the data to an S3 data lake, another Kinesis data stream for further analysis, or storing the data in OpenSearch Service for visualization. For simplicity, we send the data to Kinesis Data Firehose, which streams the data into an S3 bucket acting as our data lake.

Kinesis Data Firehose can stream data to Amazon S3, OpenSearch Service, Amazon Redshift data warehouses, and Splunk in just a few clicks.

Create the Kinesis Data Firehose delivery stream

To create our delivery stream, complete the following steps:

  1. On the Kinesis Data Firehose console, choose Create delivery stream.
  2. Choose Direct PUT for the stream source and Amazon S3 as the target.
  3. Name your delivery stream automotive-firehose.
  4. Under Destination settings, create a new bucket or use an existing bucket.
  5. Take note of the S3 bucket URL.
  6. Choose Create delivery stream.

The stream takes a few seconds to create.

  1. Return to the Kinesis Data Analytics console and choose Streaming applications.
  2. On the Studio tab, and choose your Studio notebook.
  3. Choose the link under IAM role.
  4. In the IAM window, choose Add permissions and Attach policies.
  5. Search for and select AmazonKinesisFullAccess and CloudWatchFullAccess, then choose Attach policy.
  6. You may return to your Zeppelin notebook.

Stream data into Kinesis Data Firehose

As of Apache Flink v1.15, creating the connector to the Firehose delivery stream works similar to creating a connector to any Kinesis data stream. Note that there are two differences: the connector is firehose, and the stream attribute becomes delivery-stream.

After the connector is created, we can write to the connector like any SQL table.

To validate that we’re getting data through the delivery stream, open the Amazon S3 console and confirm you see files being created. Open the file to inspect the new data.

In Kinesis Data Analytics for SQL Applications, we would have created a new destination in the SQL application dashboard. To migrate an existing destination, you add a SQL statement to your notebook that defines the new destination right in the code. You can continue to write to the new destination as you would have with an INSERT while referencing the new table name.

Time data

Another common operation you can perform in Kinesis Data Analytics Studio notebooks is aggregation over a window of time. This sort of data can be used to send to another Kinesis data stream to identify anomalies, send alerts, or be stored for further processing. The next paragraph contains a SQL query that uses a tumbling window and aggregates total fuel consumed for the automotive fleet for 30-second periods. Like our last example, we could connect to another data stream and insert this data for further analysis.

Scala and PyFlink

There are times when a function you’d perform on your data stream is better written in a programming language instead of SQL, for both simplicity and maintenance. Some examples include complex calculations that SQL functions don’t support natively, certain string manipulations, the splitting of data into multiple streams, and interacting with other AWS services (such as text translation or sentiment analysis). Kinesis Data Analytics for Apache Flink has the ability to use multiple Flink interpreters within the Zeppelin notebook, which is not available in Kinesis Data Analytics for SQL Applications.

If you have been paying close attention to our data, you’ll see that the location field is a JSON string. In Kinesis Data Analytics for SQL, we could use string functions and define a SQL function and break apart the JSON string. This is a fragile approach depending on the stability of the message data, but this could be improved with several SQL functions. The syntax for creating a function in Kinesis Data Analytics for SQL follows this pattern:

CREATE FUNCTION ''<function_name>'' ( ''<parameter_list>'' )
    RETURNS ''<data type>''
    LANGUAGE SQL
    [ SPECIFIC ''<specific_function_name>''  | [NOT] DETERMINISTIC ]
    CONTAINS SQL
    [ READS SQL DATA ]
    [ MODIFIES SQL DATA ]
    [ RETURNS NULL ON NULL INPUT | CALLED ON NULL INPUT ]  
  RETURN ''<SQL-defined function body>''

In Kinesis Data Analytics for Apache Flink, AWS recently upgraded the Apache Flink environment to v1.15, which extends Apache Flink SQL’s table SQL to add JSON functions that are similar to JSON Path syntax. This allows us to query the JSON string directly in our SQL. See the following code:

%flink.ssql(type=update)
SELECT JSON_STRING(location, ‘$.latitude) AS latitude,
JSON_STRING(location, ‘$.longitude) AS longitude
FROM my_table

Alternatively, and required prior to Apache Flink v1.15, we can use Scala or PyFlink in our notebook to convert the field and restream the data. Both languages provide robust JSON string handling.

The following PyFlink code defines two user-defined functions, which extract the latitude and longitude from the location field of our message. These UDFs can then be invoked from using Flink SQL. We reference the environment variable st_env. PyFlink creates six variables for you in your Zeppelin notebook. Zeppelin also exposes a context for you as the variable z.

Errors can also happen when messages contain unexpected data. Kinesis Data Analytics for SQL Applications provides an in-application error stream. These errors can then be processed separately and restreamed or dropped. With PyFlink in Kinesis Data Analytics Streaming applications, you can write complex error-handling strategies and immediately recover and continue processing the data. When the JSON string is passed into the UDF, it may be malformed, incomplete, or empty. By catching the error in the UDF, Python will always return a value even if an error would have occurred.

The following sample code shows another PyFlink snippet that performs a division calculation on two fields. If a division-by-zero error is encountered, it provides a default value so the stream can continue processing the message.

%flink.pyflink
@udf(input_types=[DataTypes.BIGINT()], result_type=DataTypes.BIGINT())
def DivideByZero(price):    
	try:        
		price / 0        
	except:        
		return -1
st_env.register_function("DivideByZero", DivideByZero)

Next steps

Building a pipeline as we’ve done in this post gives us the base for testing additional services in AWS. I encourage you to continue your streaming analytics learning before tearing down the streams you created. Consider the following:

Clean up

To clean up the services created in this exercise, complete the following steps:

  1. Navigate to the CloudFormation Console and delete the IoT Device Simulator stack.
  2. On the AWS IoT Core console, choose Message Routing and Rules, and delete the rule automotive_route_kinesis.
  3. Delete the Kinesis data stream automotive-data in the Kinesis Data Stream console.
  4. Remove the IAM role automotive-role in the IAM Console.
  5. In the AWS Glue console, delete the automotive-notebook-glue database.
  6. Delete the Kinesis Data Analytics Studio notebook automotive-data-notebook.
  7. Delete the Firehose delivery stream automotive-firehose.

Conclusion

Thanks for following along with this tutorial on Kinesis Data Analytics Studio. If you’re currently using a legacy Kinesis Data Analytics Studio SQL application, I recommend you reach out to your AWS technical account manager or Solutions Architect and discuss migrating to Kinesis Data Analytics Studio. You can continue your learning path in our Amazon Kinesis Data Streams Developer Guide, and access our code samples on GitHub.

About the Author

Nicholas Tunney is a Partner Solutions Architect for Worldwide Public Sector at AWS. He works with global SI partners to develop architectures on AWS for clients in the government, nonprofit healthcare, utility, and education sectors.