Post Syndicated from Steffen Hausmann original https://aws.amazon.com/blogs/big-data/streaming-etl-with-apache-flink-and-amazon-kinesis-data-analytics/
Most businesses generate data continuously in real time and at ever-increasing volumes. Data is generated as users play mobile games, load balancers log requests, customers shop on your website, and temperature changes on IoT sensors. You can capitalize on time-sensitive events, improve customer experiences, increase efficiency, and drive innovation by analyzing this data quickly. The speed at which you get these insights often depends on how quickly you can load data into data lakes, data stores, and other analytics tools. As data volume and velocity increases, it becomes more important to not only load the incoming data, but also to transform and analyze it in near-real time.
This post looks at how to use Apache Flink as a basis for sophisticated streaming extract-transform-load (ETL) pipelines. Apache Flink is a framework and distributed processing engine for processing data streams. AWS provides a fully managed service for Apache Flink through Amazon Kinesis Data Analytics, which enables you to build and run sophisticated streaming applications quickly, easily, and with low operational overhead.
This post discusses the concepts that are required to implement powerful and flexible streaming ETL pipelines with Apache Flink and Kinesis Data Analytics. It also looks at code examples for different sources and sinks. For more information, see the GitHub repo. The repo also contains an AWS CloudFormation template so you can get started in minutes and explore the example streaming ETL pipeline.
Architecture for streaming ETL with Apache Flink
Apache Flink is a framework and distributed processing engine for stateful computations over unbounded and bounded data streams. It supports a wide range of highly customizable connectors, including connectors for Apache Kafka, Amazon Kinesis Data Streams, Elasticsearch, and Amazon Simple Storage Service (Amazon S3). Moreover, Apache Flink provides a powerful API to transform, aggregate, and enrich events, and supports exactly-once semantics. Apache Flink is therefore a good foundation for the core of your streaming architecture.
To deploy and run the streaming ETL pipeline, the architecture relies on Kinesis Data Analytics. Kinesis Data Analytics enables you to run Flink applications in a fully managed environment. The service provisions and manages the required infrastructure, scales the Flink application in response to changing traffic patterns, and automatically recovers from infrastructure and application failures. You can combine the expressive Flink API for processing streaming data with the advantages of a managed service by using Kinesis Data Analytics to deploy and run Flink applications. It allows you to build robust streaming ETL pipelines and reduces the operational overhead of provisioning and operating infrastructure.
The architecture in this post takes advantage of several capabilities that you can achieve when you run Apache Flink with Kinesis Data Analytics. Specifically, the architecture supports the following:
- Private network connectivity – Connect to resources in your Amazon Virtual Private Cloud (Amazon VPC), in your data center with a VPN connection, or in a remote region with a VPC peering connection
- Multiple sources and sinks – Read and write data from Kinesis data streams, Apache Kafka clusters, and Amazon Managed Streaming for Apache Kafka (Amazon MSK) clusters
- Data partitioning – Determine the partitioning of data that is ingested into Amazon S3 based on information extracted from the event payload
- Multiple Elasticsearch indexes and custom document IDs – Fan out from a single input stream to different Elasticsearch indexes and explicitly control the document ID
- Exactly-once semantics – Avoid duplicates when ingesting and delivering data between Apache Kafka, Amazon S3, and Amazon Elasticsearch Service (Amazon ES)
The following diagram illustrates this architecture.
The remainder of this post discusses how to implement streaming ETL architectures with Apache Flink and Kinesis Data Analytics. The architecture persists streaming data from one or multiple sources to different destinations and is extensible to your needs. This post does not cover additional filtering, enrichment, and aggregation transformations, although that is a natural extension for practical applications.
This post shows how to build, deploy, and operate the Flink application with Kinesis Data Analytics, without further focusing on these operational aspects. It is only relevant to know that you can create a Kinesis Data Analytics application by uploading the compiled Flink application jar file to Amazon S3 and specifying some additional configuration options with the service. You can then execute the Kinesis Data Analytics application in a fully managed environment. For more information, see Build and run streaming applications with Apache Flink and Amazon Kinesis Data Analytics for Java Applications and the Amazon Kinesis Data Analytics developer guide.
Exploring a streaming ETL pipeline in your AWS account
Before you consider the implementation details and operational aspects, you should get a first impression of the streaming ETL pipeline in action. To create the required resources, deploy the following AWS CloudFormation template:
The template creates a Kinesis data stream and an Amazon Elastic Compute Cloud (Amazon EC2) instance to replay a historic data set into the data stream. This post uses data based on the public dataset obtained from the New York City Taxi and Limousine Commission. Each event describes a taxi trip made in New York City and includes timestamps for the start and end of a trip, information on the boroughs the trip started and ended in, and various details on the fare of the trip. A Kinesis Data Analytics application then reads the events and persists them to Amazon S3 in Parquet format and partitioned by event time.
Connect to the instance by following the link next to
ConnectToInstance in the output section of the CloudFromation template that you executed previously. You can then start replaying a set of taxi trips into the data stream with the following code:
You can obtain this command with the correct parameters from the output section of the AWS CloudFormation template. The output section also points you to the S3 bucket to which events are persisted and an Amazon CloudWatch dashboard that lets you monitor the pipeline.
For more information about enabling the remaining combinations of sources and sinks, for example, Apache Kafka and Elasticsearch, see the GitHub repo.
Building a streaming ETL pipeline with Apache Flink
Now that you have seen the pipeline in action, you can dive into the technical aspects of how to implement the functionality with Apache Flink and Kinesis Data Analytics.
Kinesis Data Analytics applications can access resources on the public internet and resources in a private subnet that is part of your VPC. By default, a Kinesis Data Analytics application only enables access to resources that you can reach over the public internet. This works well for resources that provide a public endpoint, for example, Kinesis data streams or Amazon Elasticsearch Service.
If your resources are private to your VPC, either for technical or security-related reasons, you can configure VPC connectivity for your Kinesis Data Analytics application. For example, MSK clusters are private; you cannot access them from the public internet. You may run your own Apache Kafka cluster on premises that is not exposed to the public internet and is only accessible from your VPC through a VPN connection. The same is true for other resources that are private to your VPC, such as relational databases or AWS PrivateLink-powered endpoints.
To enable VPC connectivity, configure the Kinesis Data Analytics application to connect to private subnets in your VPC. Kinesis Data Analytics creates elastic network interfaces in one or more of the subnets provided in your VPC configuration for the application, depending on the parallelism of the application. For more information, see Configuring Kinesis Data Analytics for Java Applications to access Resources in an Amazon VPC.
The following screenshot shows an example configuration of a Kinesis Data Analytics application with VPC connectivity:
The application can then access resources that have network connectivity from the configured subnets. This includes resources that are not directly contained in these subnets, which you can reach over a VPN connection or through VPC peering. This configuration also supports endpoints that are available over the public internet if you have a NAT gateway configured for the respective subnets. For more information, see Internet and Service Access for a VPC-Connected Kinesis Data Analytics for Java application.
Configuring Kinesis and Kafka sources
Apache Flink supports various data sources, including Kinesis Data Streams and Apache Kafka. For more information, see Streaming Connectors on the Apache Flink website.
To connect to a Kinesis data stream, first configure the Region and a credentials provider. As a general best practice, choose AUTO as the credentials provider. The application will then use temporary credentials from the role of the respective Kinesis Data Analytics application to read events from the specified data stream. This avoids baking static credentials into the application. In this context, it is also reasonable to increase the time between two read operations from the data stream. When you increase the default of 200 milliseconds to 1 second, the latency increases slightly, but it facilitates multiple consumers reading from the same data stream. See the following code:
This config is passed to the
FlinkKinesisConsumer with the stream name and a
DeserializationSchema. This post uses the
TripEventSchema for deserialization, which specifies how to deserialize a byte array that represents a Kinesis record into a
TripEvent object. See the following code:
Apache Flink is not limited to reading from Kinesis data streams. If you configure the Kinesis Data Analytics application’s VPC settings correctly, Apache Flink can also read events from Apache Kafka and MSK clusters. Specify a comma-separated list of broker and port pairs to use for the initial connection to your cluster. This config is passed to the
FlinkKafkaConsumer with the topic name and a
DeserializationSchema to create a source that reads from the respective topic of the Apache Kafka cluster. See the following code:
TripEvent objects that have been deserialized from the data ingested into the data stream and Kafka topic, respectively. You can then use the data streams in combination with a sink to persist the events into their respective destination.
Persisting data in Amazon S3 with data partitioning
When you persist streaming data to Amazon S3, you may want to partition the data. You can substantially improve query performance of analytic tools by partitioning data because partitions that cannot contribute to a query can be pruned and therefore do not need to be read. For example, the right partitioning strategy can improve Amazon Athena query performance and cost by reducing the amount of data read for a query. You should choose to partition your data by the same attributes used in your application logic and query patterns. Furthermore, it is common when processing streaming data to include the time of an event, or event time, in your partitioning strategy. This contrasts with using the ingestion time or some other service-side timestamp that does not reflect the time an event occurred as accurately as event time.
For more information about taking data partitioned by ingestion time and repartitioning it by event time with Athena, see Analyze your Amazon CloudFront access logs at scale. However, you can directly partition the incoming data based on event time with Apache Flink by using the payload of events to determine the partitioning, which avoids an additional post-processing step. This capability is called data partitioning and is not limited to partition by time.
You can realize data partitioning with Apache Flink’s
BucketAssigner. For more information, see Streaming File Sink on the Apache Flink website.
When given a specific event, the
BucketAssigner determines the corresponding partition prefix in the form of a string. See the following code:
The sink takes an argument for the S3 bucket as a destination path and a function that converts the
TripEvent Java objects into a string. See the following code:
You can further customize the size of the objects you write to Amazon S3 and the frequency of the object creation with a rolling policy. You can configure your policy to have more events aggregated into fewer objects at the cost of increased latency, or vice versa. This can help avoid many small objects on Amazon S3, which can be a desirable trade-off for increased latency. A high number of objects can negatively impact the query performance of consumers reading the data from Amazon S3. For more information, see DefaultRollingPolicy on the Apache Flink website.
The number of output files that arrive in your S3 bucket per your rolling policy also depends on the parallelism of the
StreamingFileSink and how you distribute events between Flink application operators. In the previous example, the Flink internal
DataStream is partitioned by pickup location ID with the keyBy operator. The location ID is also used in the
BucketAssigner as part of the prefix for objects that are written to Amazon S3. Therefore, the same node aggregates and persists all events with the same prefix, which results in particularly large objects on Amazon S3.
Apache Flink uses multipart uploads under the hood when writing to Amazon S3 with the
StreamingFileSink. In case of failures, Apache Flink may not be able to clean up incomplete multipart uploads. To avoid unnecessary storage fees, set up the automatic cleanup of incomplete multipart uploads by configuring appropriate lifecycle rules on the S3 bucket. For more information, see Important Considerations for S3 on the Apache Flink website and Example 8: Lifecycle Configuration to Abort Multipart Uploads.
Converting output to Apache Parquet
In addition to partitioning data before delivery to Amazon S3, you may want to compress it with a columnar storage format. Apache Parquet is a popular columnar format, which is well supported in the AWS ecosystem. It reduces the storage footprint and can substantially increase query performance and reduce cost.
StreamingFileSink supports Apache Parquet and other bulk-encoded formats through a built-in
BulkWriter factory. See the following code:
For more information, see Bulk-encoded Formats on the Apache Flink website.
Persisting events works a bit differently when you use the Parquet conversion. When you enable Parquet conversion, you can only configure the
StreamingFileSink with the
OnCheckpointRollingPolicy, which commits completed part files to Amazon S3 only when a checkpoint is triggered. You need to enable Apache Flink checkpoints in your Kinesis Data Analytics application to persist data to Amazon S3. It only becomes visible for consumers when a checkpoint is triggered, so your delivery latency depends on how often your application is checkpointing.
Moreover, you previously just needed to generate a string representation of the data to write to Amazon S3. In contrast, the
ParquetAvroWriters expects an Apache Avro schema for events. For more information, see the GitHub repo. You can use and extend the schema on the repo if you are want an example.
In general, it is highly desirable to convert data into Parquet if you want to work with and query the persisted data effectively. Although it requires some additional effort, the benefits of the conversion outweigh these additional complexities compared to storing raw data.
Fanning out to multiple Elasticsearch indexes and custom document IDs
Amazon ES is a fully managed service that makes it easy for you to deploy, secure, and run Elasticsearch clusters. A popular use case is to stream application and network log data into Amazon S3. These logs are documents in Elasticsearch parlance, and you can create one for every event and store it in an Elasticsearch index.
The Elasticsearch sink that Apache Flink provides is flexible and extensible. You can specify an index based on the payload of each event. This is useful when the stream contains different event types and you want to store the respective documents in different Elasticsearch indexes. With this capability, you can use a single sink and application, respectively, to write into multiple indexes. With newer Elasticsearch versions, a single index cannot contain multiple types. See the following code:
You can also explicitly set the document ID when you send documents to Elasticsearch. If an event with the same ID is ingested into Elasticsearch multiple times, it is overwritten rather than creating duplicates. This enables your writes to Elasticsearch to be idempotent. In this way, you can obtain exactly-once semantics of the entire architecture, even if your data sources only provide at-least-once semantics.
AmazonElasticsearchSink used above is an extension of the Elasticsearch sink that is comes with Apache Flink. The sink adds support to sign requests with IAM credentials so you can use the strong IAM-based authentication and authorization that is available from the service. To this end, the sink picks up temporary credentials from the Kinesis Data Analytics environment in which the application is running. It uses the Signature Version 4 method to add authentication information to the request that is sent to the Elasticsearch endpoint.
Leveraging exactly-once semantics
You can obtain exactly-once semantics by combining an idempotent sink with at-least-once semantics, but that is not always feasible. For instance, if you want to replicate data from one Apache Kafka cluster to another or persist transactional CDC data from Apache Kafka to Amazon S3, you may not be able to tolerate duplicates in the destination, but both of these sinks are not idempotent.
Apache Flink natively supports exactly-once semantics. Kinesis Data Analytics implicitly enables exactly-once mode for checkpoints. To obtain end-to-end exactly-once semantics, you need to enable checkpoints for the Kinesis Data Analytics application and choose a connector that supports exactly-once semantics, such as the
StreamingFileSink. For more information, see Fault Tolerance Guarantees of Data Sources and Sinks on the Apache Flink website.
There are some side effects to using exactly-once semantics. For example, end-to-end latency increases for several reasons. First, you can only commit the output when a checkpoint is triggered. This is the same as the latency increases that occurred when you turned on Parquet conversion. The default checkpoint interval is 1 minute, which you can decrease. However, obtaining sub-second delivery latencies are difficult with this approach.
Also, the details of end-to-end exactly-once semantics are subtle. Although the Flink application may read in an exactly-once fashion from a data stream, duplicates may already be part of the stream, so you can only obtain at-least-once semantics of the entire application. For Apache Kafka as a source and sink, different caveats apply. For more information, see Caveats on the Apache Flink website.
Be sure that you understand all the details of the entire application stack before you take a hard dependency on exactly-once semantics. In general, if your application can tolerate at-least-once semantics, it’s a good idea to use that semantic instead of relying on stronger semantics that you don’t need.
Using multiple sources and sinks
One Flink application can read data from multiple sources and persist data to multiple destinations. This is interesting for several reasons. First, you can persist the data or different subsets of the data to different destinations. For example, you can use the same application to replicate all events from your on-premises Apache Kafka cluster to an MSK cluster. At the same time, you can deliver specific, valuable events to an Elasticsearch cluster.
Second, you can use multiple sinks to increase the robustness of your application. For example, your application that applies filters and enriches streaming data can also archive the raw data stream. If something goes wrong with your more complex application logics, Amazon S3 still has the raw data, which you can use to backfill the sink.
However, there are some trade-offs. When you bundle many functionalities in a single application, you increase the blast radius of failures. If a single component of the application fails, the entire application fails and you need to recover it from the last checkpoint. This causes some downtime and increased delivery latency to all delivery destinations in the application. Also, a single large application is often harder to maintain and to change. You should strike a balance between adding functionality or creating additional Kinesis Data Analytics applications.
When you run the architecture in production, you set out to execute a single Flink application continuously and indefinitely. It is crucial to implement monitoring and proper alarming to make sure that the pipeline is working as expected and the processing can keep up with the incoming data. Ideally, the pipeline should adapt to changing throughput conditions and cause notifications if it fails to deliver data from the sources to the destinations.
Some aspects require specific attention from an operational perspective. The following section provides some ideas and further references on how you can increase the robustness of your streaming ETL pipelines.
Monitoring and scaling sources
The data stream and the MSK cluster, respectively, are the entry point to the entire architecture. They decouple the data producers from the rest of the architecture. To avoid any impact to data producers, which you often cannot control directly, you need to scale the input stream of the architecture appropriately and make sure that it can ingest messages at any time.
Kinesis Data Streams uses a throughput provisioning model based on shards. Each shard provides a certain read and write capacity. From the number of provisioned shards, you can derive the maximum throughput of the stream in terms of ingested and emitted events and data volume per second. For more information, see Kinesis Data Streams Quotas.
Kinesis Data Streams exposes metrics through CloudWatch that report on these characteristics and indicate whether the stream is over- or under-provisioned. You can use the
IncomingRecords metrics to scale the stream proactively, or you can use the
WriteProvisionedThroughputExceeded metrics to scale the stream reactively. Similar metrics exist for data egress, which you should also monitor. For more information, see Monitoring the Amazon Kinesis Data Streams with Amazon CloudWatch.
The following graph shows some of these metrics for the data stream of the example architecture. On average the Kinesis data stream receives 2.8 million events and 1.1 GB of data every minute.
You can use the
IncomingRecords metrics to scale the stream proactively whereas you can use the
WriteProvisionedThroughputExceeded metrics to scale the stream reactively. You can even automate the scaling of your Kinesis Data Streams. For more information, see Scale Your Amazon Kinesis Stream Capacity with UpdateShardCount.
Apache Kafka and Amazon MSK use a node-based provisioning model. Amazon MSK also exposes metrics through CloudWatch, including metrics that indicate how much data and how many events are ingested into the cluster. For more information, see Amazon MSK Metrics for Monitoring with CloudWatch.
In addition, you can also enable open monitoring with Prometheus for MSK clusters. It is a bit harder to know the total capacity of the cluster, and you often need benchmarking to know when you should scale. For more information about important metrics to monitor, see Monitoring Kafka on the Confluent website.
Monitoring and scaling the Kinesis Data Analytics application
The Flink application is the central core of the architecture. Kinesis Data Analytics executes it in a managed environment, and you want to make sure that it continuously reads data from the sources and persists data in the data sinks without falling behind or getting stuck.
When the application falls behind, it often is an indicator that it is not scaled appropriately. Two important metrics to track the progress of the application are
millisBehindLastest (when the application is reading from a Kinesis data stream) and
records-lag-max (when it is reading from Apache Kafka and Amazon MSK). These metrics not only indicate that data is read from the sources, but they also tell if data is read fast enough. If the values of these metrics are continuously growing, the application is continuously falling behind, which may indicate that you need to scale up the Kinesis Data Analytics application. For more information, see Kinesis Data Streams Connector Metrics and Application Metrics.
The following graph shows the metrics for the example application in this post. During checkpointing, the maximum
millisBehindLatest metric occasionally spikes up to 7 seconds. However, because the reported average of the metric is less than 1 second and the application immediately catches up to the tip of the stream again, it is not a concern for this architecture.
Although the lag of the application is one of the most important metrics to monitor, there are other relevant metrics that Apache Flink and Kinesis Data Analytics expose. For more information, see Monitoring Apache Flink Applications 101 on the Apache Flink website.
To verify that sinks are receiving data and, depending on the sink type, do not run out of storage, you need to monitor sinks closely.
You can enable detailed metrics for your S3 buckets that track the number of requests and data uploaded into the bucket with 1-minute granularity. For more information, see Monitoring Metrics with Amazon CloudWatch. The following graph shows these metrics for the S3 bucket of the example architecture:
When the architecture persists data into a Kinesis data stream or a Kafka topic, it acts as a producer, so the same recommendations as for monitoring and scaling sources apply. For more information about operating and monitoring the service in production environments, see Amazon Elasticsearch Service Best Practices.
“Failures are a given and everything eventually fails over time”, so you should expect the application to fail at some point. For example, an underlying node of the infrastructure that Kinesis Data Analytics manages might fail, or intermittent timeouts on the network can prevent the application from reading from sources or writing to sinks. When this happens, Kinesis Data Analytics restarts the application and resumes processing by recovering from the latest checkpoint. Because the raw events have been persisted in a data stream or Kafka topic, the application can reread the events that have been persisted in the stream between the last checkpoint and when it recovered and continue standard processing.
These kinds of failures are rare and the application can gracefully recover without sacrificing processing semantics, including exactly-once semantics. However, other failure modes need additional attention and mitigation.
When an exception is thrown anywhere in the application code, for example, in the component that contains the logic for parsing events, the entire application crashes. As before, the application eventually recovers, but if the exception is from a bug in your code that a specific event always hits, it results in an infinite loop. After recovering from the failure, the application rereads the event, because it was not processed successfully before, and crashes again. The process starts again and repeats indefinitely, which effectively blocks the application from making any progress.
Therefore, you want to catch and handle exceptions in the application code to avoid crashing the application. If there is a persistent problem that you cannot resolve programmatically, you can use side outputs to redirect the problematic raw events to a secondary data stream, which you can persist to a dead letter queue or an S3 bucket for later inspection. For more information, see Side Outputs on the Apache Flink website.
When the application is stuck and cannot make any progress, it is at least visible in the metrics for application lag. If your streaming ETL pipeline filters or enriches events, failures may be much more subtle, and you may only notice them long after they have been ingested. For instance, due to a bug in the application, you may accidentally drop important events or corrupt their payload in unintended ways. Kinesis data streams stores events for up to 7 days and, though technically possible, Apache Kafka is often not configured to store events indefinitely either. If you don’t identify the corruption quickly enough, you risk losing information when the retention of the raw events expires.
To protect against this scenario, you can persist the raw events to Amazon S3 before you apply any additional transformations or processing to them. You can keep the raw events and reprocess or replay them into the stream if you need to. To integrate the functionality into the application, add a second sink that just writes to Amazon S3. Alternatively, use a separate application that only reads and persists the raw events from the stream, at the cost of running and paying for an additional application.
When to choose what
AWS provides many services that work with streaming data and can perform streaming ETL. Amazon Kinesis Data Firehose can ingest, process, and persist streaming data into a range of supported destinations. There is a significant overlap of the functionality between Kinesis Data Firehose and the solution in this post, but there are different reasons to use one or the other.
As a rule of thumb, use Kinesis Data Firehose whenever it fits your requirements. The service is built with simplicity and ease of use in mind. To use Kinesis Data Firehose, you just need to configure the service. You can use Kinesis Data Firehose for streaming ETL use cases with no code, no servers, and no ongoing administration. Moreover, Kinesis Data Firehose comes with many built-in capabilities, and its pricing model allows you to only pay for the data processed and delivered. If you don’t ingest data into Kinesis Data Firehose, you pay nothing for the service.
In contrast, the solution in this post requires you to create, build, and deploy a Flink application. Moreover, you need to think about monitoring and how to obtain a robust architecture that is not only tolerant against infrastructure failures but also resilient against application failures and bugs. However, this added complexity unlocks many advanced capabilities, which your use case may require. For more information, see Build and run streaming applications with Apache Flink and Amazon Kinesis Data Analytics for Java Applications and the Amazon Kinesis Data Analytics Developer Guide.
This post discussed how to build a streaming ETL pipeline with Apache Flink and Kinesis Data Analytics. It focused on how to build an extensible solution that addresses some advanced use cases for streaming ingest while maintaining low operational overhead. The solution allows you to quickly enrich, transform, and load your streaming data into your data lake, data store, or another analytical tool without the need for an additional ETL step. The post also explored ways to extend the application with monitoring and error handling.
You should now have a good understanding of how to build streaming ETL pipelines on AWS. You can start capitalizing on your time-sensitive events by using a streaming ETL pipeline that makes valuable information quickly accessible to consumers. You can tailor the format and shape of this information to your use case without adding the substantial latency of traditional batch-based ETL processes.
About the Author
Steffen Hausmann is a Specialist Solutions Architect for Analytics at AWS. He works with customers around the globe to design and build streaming architectures so that they can get value from analyzing their streaming data. He holds a doctorate degree in computer science from the University of Munich and in his free time, he tries to lure his daughters into tech with cute stickers he collects at conferences. You can follow his ruthless attempts on Twitter (@sthmmm).