Tag Archives: objects

qrocodile: the kid-friendly Sonos system

Post Syndicated from Alex Bate original https://www.raspberrypi.org/blog/qrocodile-kid-friendly-sonos-system/

Chris Campbell’s qrocodile uses a Raspberry Pi, a camera, and QR codes to allow Chris’s children to take full control of the Sonos home sound system. And we love it!

qrocodile

Introducing qrocodile, a kid-friendly system for controlling your Sonos with QR codes. Source code is available at: https://github.com/chrispcampbell/qrocodile Learn more at: http://labonnesoupe.org https://twitter.com/chrscmpbll

Sonos

SONOS is SONOS backwards. It’s also SONOS upside down, and SONOS upside down and backwards. I just learnt that this means SONOS is an ambigram. Hurray for learning!

Sonos (the product, not the ambigram) is a multi-room speaker system controlled by an app. Speakers in different rooms can play different tracks or join forces to play one track for a smooth musical atmosphere throughout your home.

sonos raspberry pi

If you have a Sonos system in your home, I would highly recommend accessing to it from outside your home and set it to play the Imperial March as you walk through the front door. Why wouldn’t you?

qrocodile

One day, Chris’s young children wanted to play an album while eating dinner. By this one request, he was inspired to create qrocodile, a musical jukebox enabling his children to control the songs Sonos plays, and where it plays them, via QR codes.

It all started one night at the dinner table over winter break. The kids wanted to put an album on the turntable (hooked up to the line-in on a Sonos PLAY:5 in the dining room). They’re perfectly capable of putting vinyl on the turntable all by themselves, but using the Sonos app to switch over to play from the line-in is a different story.

The QR codes represent commands (such as Play in the living room, Use the turntable, or Build a song list) and artists (such as my current musical crush Courtney Barnett or the Ramones).

qrocodile raspberry Pi

A camera attached to a Raspberry Pi 3 feeds the Pi the QR code that’s presented, and the Pi runs a script that recognises the code and sends instructions to Sonos accordingly.


Chris used a costum version of the Sonos HTTP API created by Jimmy Shimizu to gain access to Sonos from his Raspberry Pi. To build the QR codes, he wrote a script that utilises the Spotify API via the Spotipy library.

His children are now able to present recognisable album art to the camera in order to play their desired track.

It’s been interesting seeing the kids putting the thing through its paces during their frequent “dance parties”, queuing up their favorite songs and uncovering new ones. I really like that they can use tangible objects to discover music in much the same way I did when I was their age, looking through my parents records, seeing which ones had interesting artwork or reading the song titles on the back, listening and exploring.

Chris has provided all the scripts for the project, along with a tutorial of how to set it up, on his GitHub — have a look if you want to recreate it or learn more about his code. Also check out Chris’ website for more on qrocodile and to see some of his other creations.

The post qrocodile: the kid-friendly Sonos system appeared first on Raspberry Pi.

Build a Multi-Tenant Amazon EMR Cluster with Kerberos, Microsoft Active Directory Integration and EMRFS Authorization

Post Syndicated from Songzhi Liu original https://aws.amazon.com/blogs/big-data/build-a-multi-tenant-amazon-emr-cluster-with-kerberos-microsoft-active-directory-integration-and-emrfs-authorization/

One of the challenges faced by our customers—especially those in highly regulated industries—is balancing the need for security with flexibility. In this post, we cover how to enable multi-tenancy and increase security by using EMRFS (EMR File System) authorization, the Amazon S3 storage-level authorization on Amazon EMR.

Amazon EMR is an easy, fast, and scalable analytics platform enabling large-scale data processing. EMRFS authorization provides Amazon S3 storage-level authorization by configuring EMRFS with multiple IAM roles. With this functionality enabled, different users and groups can share the same cluster and assume their own IAM roles respectively.

Simply put, on Amazon EMR, we can now have an Amazon EC2 role per user assumed at run time instead of one general EC2 role at the cluster level. When the user is trying to access Amazon S3 resources, Amazon EMR evaluates against a predefined mappings list in EMRFS configurations and picks up the right role for the user.

In this post, we will discuss what EMRFS authorization is (Amazon S3 storage-level access control) and show how to configure the role mappings with detailed examples. You will then have the desired permissions in a multi-tenant environment. We also demo Amazon S3 access from HDFS command line, Apache Hive on Hue, and Apache Spark.

EMRFS authorization for Amazon S3

There are two prerequisites for using this feature:

  1. Users must be authenticated, because EMRFS needs to map the current user/group/prefix to a predefined user/group/prefix. There are several authentication options. In this post, we launch a Kerberos-enabled cluster that manages the Key Distribution Center (KDC) on the master node, and enable a one-way trust from the KDC to a Microsoft Active Directory domain.
  2. The application must support accessing Amazon S3 via Applications that have their own S3FileSystem APIs (for example, Presto) are not supported at this time.

EMRFS supports three types of mapping entries: user, group, and Amazon S3 prefix. Let’s use an example to show how this works.

Assume that you have the following three identities in your organization, and they are defined in the Active Directory:

To enable all these groups and users to share the EMR cluster, you need to define the following IAM roles:

In this case, you create a separate Amazon EC2 role that doesn’t give any permission to Amazon S3. Let’s call the role the base role (the EC2 role attached to the EMR cluster), which in this example is named EMR_EC2_RestrictedRole. Then, you define all the Amazon S3 permissions for each specific user or group in their own roles. The restricted role serves as the fallback role when the user doesn’t belong to any user/group, nor does the user try to access any listed Amazon S3 prefixes defined on the list.

Important: For all other roles, like emrfs_auth_group_role_data_eng, you need to add the base role (EMR_EC2_RestrictedRole) as the trusted entity so that it can assume other roles. See the following example:

{
  "Version": "2012-10-17",
  "Statement": [
    {
      "Effect": "Allow",
      "Principal": {
        "Service": "ec2.amazonaws.com"
      },
      "Action": "sts:AssumeRole"
    },
    {
      "Effect": "Allow",
      "Principal": {
        "AWS": "arn:aws:iam::511586466501:role/EMR_EC2_RestrictedRole"
      },
      "Action": "sts:AssumeRole"
    }
  ]
}

The following is an example policy for the admin user role (emrfs_auth_user_role_admin_user):

{
    "Version": "2012-10-17",
    "Statement": [
        {
            "Effect": "Allow",
            "Action": "s3:*",
            "Resource": "*"
        }
    ]
}

We are assuming the admin user has access to all buckets in this example.

The following is an example policy for the data science group role (emrfs_auth_group_role_data_sci):

{
    "Version": "2012-10-17",
    "Statement": [
        {
            "Effect": "Allow",
            "Resource": [
                "arn:aws:s3:::emrfs-auth-data-science-bucket-demo/*",
                "arn:aws:s3:::emrfs-auth-data-science-bucket-demo"
            ],
            "Action": [
                "s3:*"
            ]
        }
    ]
}

This role grants all Amazon S3 permissions to the emrfs-auth-data-science-bucket-demo bucket and all the objects in it. Similarly, the policy for the role emrfs_auth_group_role_data_eng is shown below:

{
    "Version": "2012-10-17",
    "Statement": [
        {
            "Effect": "Allow",
            "Resource": [
                "arn:aws:s3:::emrfs-auth-data-engineering-bucket-demo/*",
                "arn:aws:s3:::emrfs-auth-data-engineering-bucket-demo"
            ],
            "Action": [
                "s3:*"
            ]
        }
    ]
}

Example role mappings configuration

To configure EMRFS authorization, you use EMR security configuration. Here is the configuration we use in this post

Consider the following scenario.

First, the admin user admin1 tries to log in and run a command to access Amazon S3 data through EMRFS. The first role emrfs_auth_user_role_admin_user on the mapping list, which is a user role, is mapped and picked up. Then admin1 has access to the Amazon S3 locations that are defined in this role.

Then a user from the data engineer group (grp_data_engineering) tries to access a data bucket to run some jobs. When EMRFS sees that the user is a member of the grp_data_engineering group, the group role emrfs_auth_group_role_data_eng is assumed, and the user has proper access to Amazon S3 that is defined in the emrfs_auth_group_role_data_eng role.

Next, the third user comes, who is not an admin and doesn’t belong to any of the groups. After failing evaluation of the top three entries, EMRFS evaluates whether the user is trying to access a certain Amazon S3 prefix defined in the last mapping entry. This type of mapping entry is called the prefix type. If the user is trying to access s3://emrfs-auth-default-bucket-demo/, then the prefix mapping is in effect, and the prefix role emrfs_auth_prefix_role_default_s3_prefix is assumed.

If the user is not trying to access any of the Amazon S3 paths that are defined on the list—which means it failed the evaluation of all the entries—it only has the permissions defined in the EMR_EC2RestrictedRole. This role is assumed by the EC2 instances in the cluster.

In this process, all the mappings defined are evaluated in the defined order, and the first role that is mapped is assumed, and the rest of the list is skipped.

Setting up an EMR cluster and mapping Active Directory users and groups

Now that we know how EMRFS authorization role mapping works, the next thing we need to think about is how we can use this feature in an easy and manageable way.

Active Directory setup

Many customers manage their users and groups using Microsoft Active Directory or other tools like OpenLDAP. In this post, we create the Active Directory on an Amazon EC2 instance running Windows Server and create the users and groups we will be using in the example below. After setting up Active Directory, we use the Amazon EMR Kerberos auto-join capability to establish a one-way trust from the KDC running on the EMR master node to the Active Directory domain on the EC2 instance. You can use your own directory services as long as it talks to the LDAP (Lightweight Directory Access Protocol).

To create and join Active Directory to Amazon EMR, follow the steps in the blog post Use Kerberos Authentication to Integrate Amazon EMR with Microsoft Active Directory.

After configuring Active Directory, you can create all the users and groups using the Active Directory tools and add users to appropriate groups. In this example, we created users like admin1, dataeng1, datascientist1, grp_data_engineering, and grp_data_science, and then add the users to the right groups.

Join the EMR cluster to an Active Directory domain

For clusters with Kerberos, Amazon EMR now supports automated Active Directory domain joins. You can use the security configuration to configure the one-way trust from the KDC to the Active Directory domain. You also configure the EMRFS role mappings in the same security configuration.

The following is an example of the EMR security configuration with a trusted Active Directory domain EMRKRB.TEST.COM and the EMRFS role mappings as we discussed earlier:

The EMRFS role mapping configuration is shown in this example:

We will also provide an example AWS CLI command that you can run.

Launching the EMR cluster and running the tests

Now you have configured Kerberos and EMRFS authorization for Amazon S3.

Additionally, you need to configure Hue with Active Directory using the Amazon EMR configuration API in order to log in using the AD users created before. The following is an example of Hue AD configuration.

[
  {
    "Classification":"hue-ini",
    "Properties":{

    },
    "Configurations":[
      {
        "Classification":"desktop",
        "Properties":{

        },
        "Configurations":[
          {
            "Classification":"ldap",
            "Properties":{

            },
            "Configurations":[
              {
                "Classification":"ldap_servers",
                "Properties":{

                },
                "Configurations":[
                  {
                    "Classification":"AWS",
                    "Properties":{
                      "base_dn":"DC=emrkrb,DC=test,DC=com",
                      "ldap_url":"ldap://emrkrb.test.com",
                      "search_bind_authentication":"false",
                      "bind_dn":"CN=adjoiner,CN=users,DC=emrkrb,DC=test,DC=com",
                      "bind_password":"Abc123456",
                      "create_users_on_login":"true",
                      "nt_domain":"emrkrb.test.com"
                    },
                    "Configurations":[

                    ]
                  }
                ]
              }
            ]
          },
          {
            "Classification":"auth",
            "Properties":{
              "backend":"desktop.auth.backend.LdapBackend"
            },
            "Configurations":[

            ]
          }
        ]
      }
    ]
  }

Note: In the preceding configuration JSON file, change the values as required before pasting it into the software setting section in the Amazon EMR console.

Now let’s use this configuration and the security configuration you created before to launch the cluster.

In the Amazon EMR console, choose Create cluster. Then choose Go to advanced options. On the Step1: Software and Steps page, under Edit software settings (optional), paste the configuration in the box.

The rest of the setup is the same as an ordinary cluster setup, except in the Security Options section. In Step 4: Security, under Permissions, choose Custom, and then choose the RestrictedRole that you created before.

Choose the appropriate subnets (these should meet the base requirement in order for a successful Active Directory join—see the Amazon EMR Management Guide for more details), and choose the appropriate security groups to make sure it talks to the Active Directory. Choose a key so that you can log in and configure the cluster.

Most importantly, choose the security configuration that you created earlier to enable Kerberos and EMRFS authorization for Amazon S3.

You can use the following AWS CLI command to create a cluster.

aws emr create-cluster --name "TestEMRFSAuthorization" \ 
--release-label emr-5.10.0 \ --instance-type m3.xlarge \ 
--instance-count 3 \ 
--ec2-attributes InstanceProfile=EMR_EC2_DefaultRole,KeyName=MyEC2KeyPair \ --service-role EMR_DefaultRole \ 
--security-configuration MyKerberosConfig \ 
--configurations file://hue-config.json \
--applications Name=Hadoop Name=Hive Name=Hue Name=Spark \ 
--kerberos-attributes Realm=EC2.INTERNAL, \ KdcAdminPassword=<YourClusterKDCAdminPassword>, \ ADDomainJoinUser=<YourADUserLogonName>,ADDomainJoinPassword=<YourADUserPassword>, \ 
CrossRealmTrustPrincipalPassword=<MatchADTrustPwd>

Note: If you create the cluster using CLI, you need to save the JSON configuration for Hue into a file named hue-config.json and place it on the server where you run the CLI command.

After the cluster gets into the Waiting state, try to connect by using SSH into the cluster using the Active Directory user name and password.

ssh -l [email protected] <EMR IP or DNS name>

Quickly run two commands to show that the Active Directory join is successful:

  1. id [user name] shows the mapped AD users and groups in Linux.
  2. hdfs groups [user name] shows the mapped group in Hadoop.

Both should return the current Active Directory user and group information if the setup is correct.

Now, you can test the user mapping first. Log in with the admin1 user, and run a Hadoop list directory command:

hadoop fs -ls s3://emrfs-auth-data-science-bucket-demo/

Now switch to a user from the data engineer group.

Retry the previous command to access the admin’s bucket. It should throw an Amazon S3 Access Denied exception.

When you try listing the Amazon S3 bucket that a data engineer group member has accessed, it triggers the group mapping.

hadoop fs -ls s3://emrfs-auth-data-engineering-bucket-demo/

It successfully returns the listing results. Next we will test Apache Hive and then Apache Spark.

 

To run jobs successfully, you need to create a home directory for every user in HDFS for staging data under /user/<username>. Users can configure a step to create a home directory at cluster launch time for every user who has access to the cluster. In this example, you use Hue since Hue will create the home directory in HDFS for the user at the first login. Here Hue also needs to be integrated with the same Active Directory as explained in the example configuration described earlier.

First, log in to Hue as a data engineer user, and open a Hive Notebook in Hue. Then run a query to create a new table pointing to the data engineer bucket, s3://emrfs-auth-data-engineering-bucket-demo/table1_data_eng/.

You can see that the table was created successfully. Now try to create another table pointing to the data science group’s bucket, where the data engineer group doesn’t have access.

It failed and threw an Amazon S3 Access Denied error.

Now insert one line of data into the successfully create table.

Next, log out, switch to a data science group user, and create another table, test2_datasci_tb.

The creation is successful.

The last task is to test Spark (it requires the user directory, but Hue created one in the previous step).

Now let’s come back to the command line and run some Spark commands.

Login to the master node using the datascientist1 user:

Start the SparkSQL interactive shell by typing spark-sql, and run the show tables command. It should list the tables that you created using Hive.

As a data science group user, try select on both tables. You will find that you can only select the table defined in the location that your group has access to.

Conclusion

EMRFS authorization for Amazon S3 enables you to have multiple roles on the same cluster, providing flexibility to configure a shared cluster for different teams to achieve better efficiency. The Active Directory integration and group mapping make it much easier for you to manage your users and groups, and provides better auditability in a multi-tenant environment.


Additional Reading

If you found this post useful, be sure to check out Use Kerberos Authentication to Integrate Amazon EMR with Microsoft Active Directory and Launching and Running an Amazon EMR Cluster inside a VPC.


About the Authors

Songzhi Liu is a Big Data Consultant with AWS Professional Services. He works closely with AWS customers to provide them Big Data & Machine Learning solutions and best practices on the Amazon cloud.

 

 

 

 

Reactive Microservices Architecture on AWS

Post Syndicated from Sascha Moellering original https://aws.amazon.com/blogs/architecture/reactive-microservices-architecture-on-aws/

Microservice-application requirements have changed dramatically in recent years. These days, applications operate with petabytes of data, need almost 100% uptime, and end users expect sub-second response times. Typical N-tier applications can’t deliver on these requirements.

Reactive Manifesto, published in 2014, describes the essential characteristics of reactive systems including: responsiveness, resiliency, elasticity, and being message driven.

Being message driven is perhaps the most important characteristic of reactive systems. Asynchronous messaging helps in the design of loosely coupled systems, which is a key factor for scalability. In order to build a highly decoupled system, it is important to isolate services from each other. As already described, isolation is an important aspect of the microservices pattern. Indeed, reactive systems and microservices are a natural fit.

Implemented Use Case
This reference architecture illustrates a typical ad-tracking implementation.

Many ad-tracking companies collect massive amounts of data in near-real-time. In many cases, these workloads are very spiky and heavily depend on the success of the ad-tech companies’ customers. Typically, an ad-tracking-data use case can be separated into a real-time part and a non-real-time part. In the real-time part, it is important to collect data as fast as possible and ask several questions including:,  “Is this a valid combination of parameters?,””Does this program exist?,” “Is this program still valid?”

Because response time has a huge impact on conversion rate in advertising, it is important for advertisers to respond as fast as possible. This information should be kept in memory to reduce communication overhead with the caching infrastructure. The tracking application itself should be as lightweight and scalable as possible. For example, the application shouldn’t have any shared mutable state and it should use reactive paradigms. In our implementation, one main application is responsible for this real-time part. It collects and validates data, responds to the client as fast as possible, and asynchronously sends events to backend systems.

The non-real-time part of the application consumes the generated events and persists them in a NoSQL database. In a typical tracking implementation, clicks, cookie information, and transactions are matched asynchronously and persisted in a data store. The matching part is not implemented in this reference architecture. Many ad-tech architectures use frameworks like Hadoop for the matching implementation.

The system can be logically divided into the data collection partand the core data updatepart. The data collection part is responsible for collecting, validating, and persisting the data. In the core data update part, the data that is used for validation gets updated and all subscribers are notified of new data.

Components and Services

Main Application
The main application is implemented using Java 8 and uses Vert.x as the main framework. Vert.x is an event-driven, reactive, non-blocking, polyglot framework to implement microservices. It runs on the Java virtual machine (JVM) by using the low-level IO library Netty. You can write applications in Java, JavaScript, Groovy, Ruby, Kotlin, Scala, and Ceylon. The framework offers a simple and scalable actor-like concurrency model. Vert.x calls handlers by using a thread known as an event loop. To use this model, you have to write code known as “verticles.” Verticles share certain similarities with actors in the actor model. To use them, you have to implement the verticle interface. Verticles communicate with each other by generating messages in  a single event bus. Those messages are sent on the event bus to a specific address, and verticles can register to this address by using handlers.

With only a few exceptions, none of the APIs in Vert.x block the calling thread. Similar to Node.js, Vert.x uses the reactor pattern. However, in contrast to Node.js, Vert.x uses several event loops. Unfortunately, not all APIs in the Java ecosystem are written asynchronously, for example, the JDBC API. Vert.x offers a possibility to run this, blocking APIs without blocking the event loop. These special verticles are called worker verticles. You don’t execute worker verticles by using the standard Vert.x event loops, but by using a dedicated thread from a worker pool. This way, the worker verticles don’t block the event loop.

Our application consists of five different verticles covering different aspects of the business logic. The main entry point for our application is the HttpVerticle, which exposes an HTTP-endpoint to consume HTTP-requests and for proper health checking. Data from HTTP requests such as parameters and user-agent information are collected and transformed into a JSON message. In order to validate the input data (to ensure that the program exists and is still valid), the message is sent to the CacheVerticle.

This verticle implements an LRU-cache with a TTL of 10 minutes and a capacity of 100,000 entries. Instead of adding additional functionality to a standard JDK map implementation, we use Google Guava, which has all the features we need. If the data is not in the L1 cache, the message is sent to the RedisVerticle. This verticle is responsible for data residing in Amazon ElastiCache and uses the Vert.x-redis-client to read data from Redis. In our example, Redis is the central data store. However, in a typical production implementation, Redis would just be the L2 cache with a central data store like Amazon DynamoDB. One of the most important paradigms of a reactive system is to switch from a pull- to a push-based model. To achieve this and reduce network overhead, we’ll use Redis pub/sub to push core data changes to our main application.

Vert.x also supports direct Redis pub/sub-integration, the following code shows our subscriber-implementation:

vertx.eventBus().<JsonObject>consumer(REDIS_PUBSUB_CHANNEL_VERTX, received -> {

JsonObject value = received.body().getJsonObject("value");

String message = value.getString("message");

JsonObject jsonObject = new JsonObject(message);

eb.send(CACHE_REDIS_EVENTBUS_ADDRESS, jsonObject);

});

redis.subscribe(Constants.REDIS_PUBSUB_CHANNEL, res -> {

if (res.succeeded()) {

LOGGER.info("Subscribed to " + Constants.REDIS_PUBSUB_CHANNEL);

} else {

LOGGER.info(res.cause());

}

});

The verticle subscribes to the appropriate Redis pub/sub-channel. If a message is sent over this channel, the payload is extracted and forwarded to the cache-verticle that stores the data in the L1-cache. After storing and enriching data, a response is sent back to the HttpVerticle, which responds to the HTTP request that initially hit this verticle. In addition, the message is converted to ByteBuffer, wrapped in protocol buffers, and send to an Amazon Kinesis Data Stream.

The following example shows a stripped-down version of the KinesisVerticle:

public class KinesisVerticle extends AbstractVerticle {

private static final Logger LOGGER = LoggerFactory.getLogger(KinesisVerticle.class);

private AmazonKinesisAsync kinesisAsyncClient;

private String eventStream = "EventStream";

@Override

public void start() throws Exception {

EventBus eb = vertx.eventBus();

kinesisAsyncClient = createClient();

eventStream = System.getenv(STREAM_NAME) == null ? "EventStream" : System.getenv(STREAM_NAME);

eb.consumer(Constants.KINESIS_EVENTBUS_ADDRESS, message -> {

try {

TrackingMessage trackingMessage = Json.decodeValue((String)message.body(), TrackingMessage.class);

String partitionKey = trackingMessage.getMessageId();

byte [] byteMessage = createMessage(trackingMessage);

ByteBuffer buf = ByteBuffer.wrap(byteMessage);

sendMessageToKinesis(buf, partitionKey);

message.reply("OK");

}

catch (KinesisException exc) {

LOGGER.error(exc);

}

});

}

Kinesis Consumer
This AWS Lambda function consumes data from an Amazon Kinesis Data Stream and persists the data in an Amazon DynamoDB table. In order to improve testability, the invocation code is separated from the business logic. The invocation code is implemented in the class KinesisConsumerHandler and iterates over the Kinesis events pulled from the Kinesis stream by AWS Lambda. Each Kinesis event is unwrapped and transformed from ByteBuffer to protocol buffers and converted into a Java object. Those Java objects are passed to the business logic, which persists the data in a DynamoDB table. In order to improve duration of successive Lambda calls, the DynamoDB-client is instantiated lazily and reused if possible.

Redis Updater
From time to time, it is necessary to update core data in Redis. A very efficient implementation for this requirement is using AWS Lambda and Amazon Kinesis. New core data is sent over the AWS Kinesis stream using JSON as data format and consumed by a Lambda function. This function iterates over the Kinesis events pulled from the Kinesis stream by AWS Lambda. Each Kinesis event is unwrapped and transformed from ByteBuffer to String and converted into a Java object. The Java object is passed to the business logic and stored in Redis. In addition, the new core data is also sent to the main application using Redis pub/sub in order to reduce network overhead and converting from a pull- to a push-based model.

The following example shows the source code to store data in Redis and notify all subscribers:

public void updateRedisData(final TrackingMessage trackingMessage, final Jedis jedis, final LambdaLogger logger) {

try {

ObjectMapper mapper = new ObjectMapper();

String jsonString = mapper.writeValueAsString(trackingMessage);

Map<String, String> map = marshal(jsonString);

String statusCode = jedis.hmset(trackingMessage.getProgramId(), map);

}

catch (Exception exc) {

if (null == logger)

exc.printStackTrace();

else

logger.log(exc.getMessage());

}

}

public void notifySubscribers(final TrackingMessage trackingMessage, final Jedis jedis, final LambdaLogger logger) {

try {

ObjectMapper mapper = new ObjectMapper();

String jsonString = mapper.writeValueAsString(trackingMessage);

jedis.publish(Constants.REDIS_PUBSUB_CHANNEL, jsonString);

}

catch (final IOException e) {

log(e.getMessage(), logger);

}

}

Similarly to our Kinesis Consumer, the Redis-client is instantiated somewhat lazily.

Infrastructure as Code
As already outlined, latency and response time are a very critical part of any ad-tracking solution because response time has a huge impact on conversion rate. In order to reduce latency for customers world-wide, it is common practice to roll out the infrastructure in different AWS Regions in the world to be as close to the end customer as possible. AWS CloudFormation can help you model and set up your AWS resources so that you can spend less time managing those resources and more time focusing on your applications that run in AWS.

You create a template that describes all the AWS resources that you want (for example, Amazon EC2 instances or Amazon RDS DB instances), and AWS CloudFormation takes care of provisioning and configuring those resources for you. Our reference architecture can be rolled out in different Regions using an AWS CloudFormation template, which sets up the complete infrastructure (for example, Amazon Virtual Private Cloud (Amazon VPC), Amazon Elastic Container Service (Amazon ECS) cluster, Lambda functions, DynamoDB table, Amazon ElastiCache cluster, etc.).

Conclusion
In this blog post we described reactive principles and an example architecture with a common use case. We leveraged the capabilities of different frameworks in combination with several AWS services in order to implement reactive principles—not only at the application-level but also at the system-level. I hope I’ve given you ideas for creating your own reactive applications and systems on AWS.

About the Author

Sascha Moellering is a Senior Solution Architect. Sascha is primarily interested in automation, infrastructure as code, distributed computing, containers and JVM. He can be reached at [email protected]

 

 

Top 8 Best Practices for High-Performance ETL Processing Using Amazon Redshift

Post Syndicated from Thiyagarajan Arumugam original https://aws.amazon.com/blogs/big-data/top-8-best-practices-for-high-performance-etl-processing-using-amazon-redshift/

An ETL (Extract, Transform, Load) process enables you to load data from source systems into your data warehouse. This is typically executed as a batch or near-real-time ingest process to keep the data warehouse current and provide up-to-date analytical data to end users.

Amazon Redshift is a fast, petabyte-scale data warehouse that enables you easily to make data-driven decisions. With Amazon Redshift, you can get insights into your big data in a cost-effective fashion using standard SQL. You can set up any type of data model, from star and snowflake schemas, to simple de-normalized tables for running any analytical queries.

To operate a robust ETL platform and deliver data to Amazon Redshift in a timely manner, design your ETL processes to take account of Amazon Redshift’s architecture. When migrating from a legacy data warehouse to Amazon Redshift, it is tempting to adopt a lift-and-shift approach, but this can result in performance and scale issues long term. This post guides you through the following best practices for ensuring optimal, consistent runtimes for your ETL processes:

  • COPY data from multiple, evenly sized files.
  • Use workload management to improve ETL runtimes.
  • Perform table maintenance regularly.
  • Perform multiple steps in a single transaction.
  • Loading data in bulk.
  • Use UNLOAD to extract large result sets.
  • Use Amazon Redshift Spectrum for ad hoc ETL processing.
  • Monitor daily ETL health using diagnostic queries.

1. COPY data from multiple, evenly sized files

Amazon Redshift is an MPP (massively parallel processing) database, where all the compute nodes divide and parallelize the work of ingesting data. Each node is further subdivided into slices, with each slice having one or more dedicated cores, equally dividing the processing capacity. The number of slices per node depends on the node type of the cluster. For example, each DS2.XLARGE compute node has two slices, whereas each DS2.8XLARGE compute node has 16 slices.

When you load data into Amazon Redshift, you should aim to have each slice do an equal amount of work. When you load the data from a single large file or from files split into uneven sizes, some slices do more work than others. As a result, the process runs only as fast as the slowest, or most heavily loaded, slice. In the example shown below, a single large file is loaded into a two-node cluster, resulting in only one of the nodes, “Compute-0”, performing all the data ingestion:

When splitting your data files, ensure that they are of approximately equal size – between 1 MB and 1 GB after compression. The number of files should be a multiple of the number of slices in your cluster. Also, I strongly recommend that you individually compress the load files using gzip, lzop, or bzip2 to efficiently load large datasets.

When loading multiple files into a single table, use a single COPY command for the table, rather than multiple COPY commands. Amazon Redshift automatically parallelizes the data ingestion. Using a single COPY command to bulk load data into a table ensures optimal use of cluster resources, and quickest possible throughput.

2. Use workload management to improve ETL runtimes

Use Amazon Redshift’s workload management (WLM) to define multiple queues dedicated to different workloads (for example, ETL versus reporting) and to manage the runtimes of queries. As you migrate more workloads into Amazon Redshift, your ETL runtimes can become inconsistent if WLM is not appropriately set up.

I recommend limiting the overall concurrency of WLM across all queues to around 15 or less. This WLM guide helps you organize and monitor the different queues for your Amazon Redshift cluster.

When managing different workloads on your Amazon Redshift cluster, consider the following for the queue setup:

  • Create a queue dedicated to your ETL processes. Configure this queue with a small number of slots (5 or fewer). Amazon Redshift is designed for analytics queries, rather than transaction processing. The cost of COMMIT is relatively high, and excessive use of COMMIT can result in queries waiting for access to the commit queue. Because ETL is a commit-intensive process, having a separate queue with a small number of slots helps mitigate this issue.
  • Claim extra memory available in a queue. When executing an ETL query, you can take advantage of the wlm_query_slot_count to claim the extra memory available in a particular queue. For example, a typical ETL process might involve COPYing raw data into a staging table so that downstream ETL jobs can run transformations that calculate daily, weekly, and monthly aggregates. To speed up the COPY process (so that the downstream tasks can start in parallel sooner), the wlm_query_slot_count can be increased for this step.
  • Create a separate queue for reporting queries. Configure query monitoring rules on this queue to further manage long-running and expensive queries.
  • Take advantage of the dynamic memory parameters. They swap the memory from your ETL to your reporting queue after the ETL job has completed.

3. Perform table maintenance regularly

Amazon Redshift is a columnar database, which enables fast transformations for aggregating data. Performing regular table maintenance ensures that transformation ETLs are predictable and performant. To get the best performance from your Amazon Redshift database, you must ensure that database tables regularly are VACUUMed and ANALYZEd. The Analyze & Vacuum schema utility helps you automate the table maintenance task and have VACUUM & ANALYZE executed in a regular fashion.

  • Use VACUUM to sort tables and remove deleted blocks

During a typical ETL refresh process, tables receive new incoming records using COPY, and unneeded data (cold data) is removed using DELETE. New rows are added to the unsorted region in a table. Deleted rows are simply marked for deletion.

DELETE does not automatically reclaim the space occupied by the deleted rows. Adding and removing large numbers of rows can therefore cause the unsorted region and the number of deleted blocks to grow. This can degrade the performance of queries executed against these tables.

After an ETL process completes, perform VACUUM to ensure that user queries execute in a consistent manner. The complete list of tables that need VACUUMing can be found using the Amazon Redshift Util’s table_info script.

Use the following approaches to ensure that VACCUM is completed in a timely manner:

  • Use wlm_query_slot_count to claim all the memory allocated in the ETL WLM queue during the VACUUM process.
  • DROP or TRUNCATE intermediate or staging tables, thereby eliminating the need to VACUUM them.
  • If your table has a compound sort key with only one sort column, try to load your data in sort key order. This helps reduce or eliminate the need to VACUUM the table.
  • Consider using time series This helps reduce the amount of data you need to VACUUM.
  • Use ANALYZE to update database statistics

Amazon Redshift uses a cost-based query planner and optimizer using statistics about tables to make good decisions about the query plan for the SQL statements. Regular statistics collection after the ETL completion ensures that user queries run fast, and that daily ETL processes are performant. The Amazon Redshift utility table_info script provides insights into the freshness of the statistics. Keeping the statistics off (pct_stats_off) less than 20% ensures effective query plans for the SQL queries.

4. Perform multiple steps in a single transaction

ETL transformation logic often spans multiple steps. Because commits in Amazon Redshift are expensive, if each ETL step performs a commit, multiple concurrent ETL processes can take a long time to execute.

To minimize the number of commits in a process, the steps in an ETL script should be surrounded by a BEGIN…END statement so that a single commit is performed only after all the transformation logic has been executed. For example, here is an example multi-step ETL script that performs one commit at the end:

Begin
CREATE temporary staging_table;
INSERT INTO staging_table SELECT .. FROM source (transformation logic);
DELETE FROM daily_table WHERE dataset_date =?;
INSERT INTO daily_table SELECT .. FROM staging_table (daily aggregate);
DELETE FROM weekly_table WHERE weekending_date=?;
INSERT INTO weekly_table SELECT .. FROM staging_table(weekly aggregate);
Commit

5. Loading data in bulk

Amazon Redshift is designed to store and query petabyte-scale datasets. Using Amazon S3 you can stage and accumulate data from multiple source systems before executing a bulk COPY operation. The following methods allow efficient and fast transfer of these bulk datasets into Amazon Redshift:

  • Use a manifest file to ingest large datasets that span multiple files. The manifest file is a JSON file that lists all the files to be loaded into Amazon Redshift. Using a manifest file ensures that Amazon Redshift has a consistent view of the data to be loaded from S3, while also ensuring that duplicate files do not result in the same data being loaded more than one time.
  • Use temporary staging tables to hold the data for transformation. These tables are automatically dropped after the ETL session is complete. Temporary tables can be created using the CREATE TEMPORARY TABLE syntax, or by issuing a SELECT … INTO #TEMP_TABLE query. Explicitly specifying the CREATE TEMPORARY TABLE statement allows you to control the DISTRIBUTION KEY, SORT KEY, and compression settings to further improve performance.
  • User ALTER table APPEND to swap data from the staging tables to the target table. Data in the source table is moved to matching columns in the target table. Column order doesn’t matter. After data is successfully appended to the target table, the source table is empty. ALTER TABLE APPEND is much faster than a similar CREATE TABLE AS or INSERT INTO operation because it doesn’t involve copying or moving data.

6. Use UNLOAD to extract large result sets

Fetching a large number of rows using SELECT is expensive and takes a long time. When a large amount of data is fetched from the Amazon Redshift cluster, the leader node has to hold the data temporarily until the fetches are complete. Further, data is streamed out sequentially, which results in longer elapsed time. As a result, the leader node can become hot, which not only affects the SELECT that is being executed, but also throttles resources for creating execution plans and managing the overall cluster resources. Here is an example of a large SELECT statement. Notice that the leader node is doing most of the work to stream out the rows:

Use UNLOAD to extract large results sets directly to S3. After it’s in S3, the data can be shared with multiple downstream systems. By default, UNLOAD writes data in parallel to multiple files according to the number of slices in the cluster. All the compute nodes participate to quickly offload the data into S3.

If you are extracting data for use with Amazon Redshift Spectrum, you should make use of the MAXFILESIZE parameter to and keep files are 150 MB. Similar to item 1 above, having many evenly sized files ensures that Redshift Spectrum can do the maximum amount of work in parallel.

7. Use Redshift Spectrum for ad hoc ETL processing

Events such as data backfill, promotional activity, and special calendar days can trigger additional data volumes that affect the data refresh times in your Amazon Redshift cluster. To help address these spikes in data volumes and throughput, I recommend staging data in S3. After data is organized in S3, Redshift Spectrum enables you to query it directly using standard SQL. In this way, you gain the benefits of additional capacity without having to resize your cluster.

For tips on getting started with and optimizing the use of Redshift Spectrum, see the previous post, 10 Best Practices for Amazon Redshift Spectrum.

8. Monitor daily ETL health using diagnostic queries

Monitoring the health of your ETL processes on a regular basis helps identify the early onset of performance issues before they have a significant impact on your cluster. The following monitoring scripts can be used to provide insights into the health of your ETL processes:

Script Use when… Solution
commit_stats.sql – Commit queue statistics from past days, showing largest queue length and queue time first DML statements such as INSERT/UPDATE/COPY/DELETE operations take several times longer to execute when multiple of these operations are in progress Set up separate WLM queues for the ETL process and limit the concurrency to < 5.
copy_performance.sql –  Copy command statistics for the past days Daily COPY operations take longer to execute • Follow the best practices for the COPY command.
• Analyze data growth with the incoming datasets and consider cluster resize to meet the expected SLA.
table_info.sql – Table skew and unsorted statistics along with storage and key information Transformation steps take longer to execute • Set up regular VACCUM jobs to address unsorted rows and claim the deleted blocks so that transformation SQL execute optimally.
• Consider a table redesign to avoid data skewness.
v_check_transaction_locks.sql – Monitor transaction locks INSERT/UPDATE/COPY/DELETE operations on particular tables do not respond back in timely manner, compared to when run after the ETL Multiple DML statements are operating on the same target table at the same moment from different transactions. Set up ETL job dependency so that they execute serially for the same target table.
v_get_schema_priv_by_user.sql – Get the schema that the user has access to Reporting users can view intermediate tables Set up separate database groups for reporting and ETL users, and grants access to objects using GRANT.
v_generate_tbl_ddl.sql – Get the table DDL You need to create an empty table with same structure as target table for data backfill Generate DDL using this script for data backfill.
v_space_used_per_tbl.sql – monitor space used by individual tables Amazon Redshift data warehouse space growth is trending upwards more than normal

Analyze the individual tables that are growing at higher rate than normal. Consider data archival using UNLOAD to S3 and Redshift Spectrum for later analysis.

Use unscanned_table_summary.sql to find unused table and archive or drop them.

top_queries.sql – Return the top 50 time consuming statements aggregated by its text ETL transformations are taking longer to execute Analyze the top transformation SQL and use EXPLAIN to find opportunities for tuning the query plan.

There are several other useful scripts available in the amazon-redshift-utils repository. The AWS Lambda Utility Runner runs a subset of these scripts on a scheduled basis, allowing you to automate much of monitoring of your ETL processes.

Example ETL process

The following ETL process reinforces some of the best practices discussed in this post. Consider the following four-step daily ETL workflow where data from an RDBMS source system is staged in S3 and then loaded into Amazon Redshift. Amazon Redshift is used to calculate daily, weekly, and monthly aggregations, which are then unloaded to S3, where they can be further processed and made available for end-user reporting using a number of different tools, including Redshift Spectrum and Amazon Athena.

Step 1:  Extract from the RDBMS source to a S3 bucket

In this ETL process, the data extract job fetches change data every 1 hour and it is staged into multiple hourly files. For example, the staged S3 folder looks like the following:

 [[email protected] ~]$ aws s3 ls s3://<<S3 Bucket>>/batch/2017/07/02/
2017-07-02 01:59:58   81900220 20170702T01.export.gz
2017-07-02 02:59:56   84926844 20170702T02.export.gz
2017-07-02 03:59:54   78990356 20170702T03.export.gz
…
2017-07-02 22:00:03   75966745 20170702T21.export.gz
2017-07-02 23:00:02   89199874 20170702T22.export.gz
2017-07-02 00:59:59   71161715 20170702T23.export.gz

Organizing the data into multiple, evenly sized files enables the COPY command to ingest this data using all available resources in the Amazon Redshift cluster. Further, the files are compressed (gzipped) to further reduce COPY times.

Step 2: Stage data to the Amazon Redshift table for cleansing

Ingesting the data can be accomplished using a JSON-based manifest file. Using the manifest file ensures that S3 eventual consistency issues can be eliminated and also provides an opportunity to dedupe any files if needed. A sample manifest20170702.json file looks like the following:

{
  "entries": [
    {"url":" s3://<<S3 Bucket>>/batch/2017/07/02/20170702T01.export.gz", "mandatory":true},
    {"url":" s3://<<S3 Bucket>>/batch/2017/07/02/20170702T02.export.gz", "mandatory":true},
    …
    {"url":" s3://<<S3 Bucket>>/batch/2017/07/02/20170702T23.export.gz", "mandatory":true}
  ]
}

The data can be ingested using the following command:

SET wlm_query_slot_count TO <<max available concurrency in the ETL queue>>;
COPY stage_tbl FROM 's3:// <<S3 Bucket>>/batch/manifest20170702.json' iam_role 'arn:aws:iam::0123456789012:role/MyRedshiftRole' manifest;

Because the downstream ETL processes depend on this COPY command to complete, the wlm_query_slot_count is used to claim all the memory available to the queue. This helps the COPY command complete as quickly as possible.

Step 3: Transform data to create daily, weekly, and monthly datasets and load into target tables

Data is staged in the “stage_tbl” from where it can be transformed into the daily, weekly, and monthly aggregates and loaded into target tables. The following job illustrates a typical weekly process:

Begin
INSERT into ETL_LOG (..) values (..);
DELETE from weekly_tbl where dataset_week = <<current week>>;
INSERT into weekly_tbl (..)
  SELECT date_trunc('week', dataset_day) AS week_begin_dataset_date, SUM(C1) AS C1, SUM(C2) AS C2
	FROM   stage_tbl
GROUP BY date_trunc('week', dataset_day);
INSERT into AUDIT_LOG values (..);
COMMIT;
End;

As shown above, multiple steps are combined into one transaction to perform a single commit, reducing contention on the commit queue.

Step 4: Unload the daily dataset to populate the S3 data lake bucket

The transformed results are now unloaded into another S3 bucket, where they can be further processed and made available for end-user reporting using a number of different tools, including Redshift Spectrum and Amazon Athena.

unload ('SELECT * FROM weekly_tbl WHERE dataset_week = <<current week>>’) TO 's3:// <<S3 Bucket>>/datalake/weekly/20170526/' iam_role 'arn:aws:iam::0123456789012:role/MyRedshiftRole';

Summary

Amazon Redshift lets you easily operate petabyte-scale data warehouses on the cloud. This post summarized the best practices for operating scalable ETL natively within Amazon Redshift. I demonstrated efficient ways to ingest and transform data, along with close monitoring. I also demonstrated the best practices being used in a typical sample ETL workload to transform the data into Amazon Redshift.

If you have questions or suggestions, please comment below.

 


About the Author

Thiyagarajan Arumugam is a Big Data Solutions Architect at Amazon Web Services and designs customer architectures to process data at scale. Prior to AWS, he built data warehouse solutions at Amazon.com. In his free time, he enjoys all outdoor sports and practices the Indian classical drum mridangam.

 

AWS Glue Now Supports Scala Scripts

Post Syndicated from Mehul Shah original https://aws.amazon.com/blogs/big-data/aws-glue-now-supports-scala-scripts/

We are excited to announce AWS Glue support for running ETL (extract, transform, and load) scripts in Scala. Scala lovers can rejoice because they now have one more powerful tool in their arsenal. Scala is the native language for Apache Spark, the underlying engine that AWS Glue offers for performing data transformations.

Beyond its elegant language features, writing Scala scripts for AWS Glue has two main advantages over writing scripts in Python. First, Scala is faster for custom transformations that do a lot of heavy lifting because there is no need to shovel data between Python and Apache Spark’s Scala runtime (that is, the Java virtual machine, or JVM). You can build your own transformations or invoke functions in third-party libraries. Second, it’s simpler to call functions in external Java class libraries from Scala because Scala is designed to be Java-compatible. It compiles to the same bytecode, and its data structures don’t need to be converted.

To illustrate these benefits, we walk through an example that analyzes a recent sample of the GitHub public timeline available from the GitHub archive. This site is an archive of public requests to the GitHub service, recording more than 35 event types ranging from commits and forks to issues and comments.

This post shows how to build an example Scala script that identifies highly negative issues in the timeline. It pulls out issue events in the timeline sample, analyzes their titles using the sentiment prediction functions from the Stanford CoreNLP libraries, and surfaces the most negative issues.

Getting started

Before we start writing scripts, we use AWS Glue crawlers to get a sense of the data—its structure and characteristics. We also set up a development endpoint and attach an Apache Zeppelin notebook, so we can interactively explore the data and author the script.

Crawl the data

The dataset used in this example was downloaded from the GitHub archive website into our sample dataset bucket in Amazon S3, and copied to the following locations:

s3://aws-glue-datasets-<region>/examples/scala-blog/githubarchive/data/

Choose the best folder by replacing <region> with the region that you’re working in, for example, us-east-1. Crawl this folder, and put the results into a database named githubarchive in the AWS Glue Data Catalog, as described in the AWS Glue Developer Guide. This folder contains 12 hours of the timeline from January 22, 2017, and is organized hierarchically (that is, partitioned) by year, month, and day.

When finished, use the AWS Glue console to navigate to the table named data in the githubarchive database. Notice that this data has eight top-level columns, which are common to each event type, and three partition columns that correspond to year, month, and day.

Choose the payload column, and you will notice that it has a complex schema—one that reflects the union of the payloads of event types that appear in the crawled data. Also note that the schema that crawlers generate is a subset of the true schema because they sample only a subset of the data.

Set up the library, development endpoint, and notebook

Next, you need to download and set up the libraries that estimate the sentiment in a snippet of text. The Stanford CoreNLP libraries contain a number of human language processing tools, including sentiment prediction.

Download the Stanford CoreNLP libraries. Unzip the .zip file, and you’ll see a directory full of jar files. For this example, the following jars are required:

  • stanford-corenlp-3.8.0.jar
  • stanford-corenlp-3.8.0-models.jar
  • ejml-0.23.jar

Upload these files to an Amazon S3 path that is accessible to AWS Glue so that it can load these libraries when needed. For this example, they are in s3://glue-sample-other/corenlp/.

Development endpoints are static Spark-based environments that can serve as the backend for data exploration. You can attach notebooks to these endpoints to interactively send commands and explore and analyze your data. These endpoints have the same configuration as that of AWS Glue’s job execution system. So, commands and scripts that work there also work the same when registered and run as jobs in AWS Glue.

To set up an endpoint and a Zeppelin notebook to work with that endpoint, follow the instructions in the AWS Glue Developer Guide. When you are creating an endpoint, be sure to specify the locations of the previously mentioned jars in the Dependent jars path as a comma-separated list. Otherwise, the libraries will not be loaded.

After you set up the notebook server, go to the Zeppelin notebook by choosing Dev Endpoints in the left navigation pane on the AWS Glue console. Choose the endpoint that you created. Next, choose the Notebook Server URL, which takes you to the Zeppelin server. Log in using the notebook user name and password that you specified when creating the notebook. Finally, create a new note to try out this example.

Each notebook is a collection of paragraphs, and each paragraph contains a sequence of commands and the output for that command. Moreover, each notebook includes a number of interpreters. If you set up the Zeppelin server using the console, the (Python-based) pyspark and (Scala-based) spark interpreters are already connected to your new development endpoint, with pyspark as the default. Therefore, throughout this example, you need to prepend %spark at the top of your paragraphs. In this example, we omit these for brevity.

Working with the data

In this section, we use AWS Glue extensions to Spark to work with the dataset. We look at the actual schema of the data and filter out the interesting event types for our analysis.

Start with some boilerplate code to import libraries that you need:

%spark

import com.amazonaws.services.glue.DynamicRecord
import com.amazonaws.services.glue.GlueContext
import com.amazonaws.services.glue.util.GlueArgParser
import com.amazonaws.services.glue.util.Job
import com.amazonaws.services.glue.util.JsonOptions
import com.amazonaws.services.glue.types._
import org.apache.spark.SparkContext

Then, create the Spark and AWS Glue contexts needed for working with the data:

@transient val spark: SparkContext = SparkContext.getOrCreate()
val glueContext: GlueContext = new GlueContext(spark)

You need the transient decorator on the SparkContext when working in Zeppelin; otherwise, you will run into a serialization error when executing commands.

Dynamic frames

This section shows how to create a dynamic frame that contains the GitHub records in the table that you crawled earlier. A dynamic frame is the basic data structure in AWS Glue scripts. It is like an Apache Spark data frame, except that it is designed and optimized for data cleaning and transformation workloads. A dynamic frame is well-suited for representing semi-structured datasets like the GitHub timeline.

A dynamic frame is a collection of dynamic records. In Spark lingo, it is an RDD (resilient distributed dataset) of DynamicRecords. A dynamic record is a self-describing record. Each record encodes its columns and types, so every record can have a schema that is unique from all others in the dynamic frame. This is convenient and often more efficient for datasets like the GitHub timeline, where payloads can vary drastically from one event type to another.

The following creates a dynamic frame, github_events, from your table:

val github_events = glueContext
                    .getCatalogSource(database = "githubarchive", tableName = "data")
                    .getDynamicFrame()

The getCatalogSource() method returns a DataSource, which represents a particular table in the Data Catalog. The getDynamicFrame() method returns a dynamic frame from the source.

Recall that the crawler created a schema from only a sample of the data. You can scan the entire dataset, count the rows, and print the complete schema as follows:

github_events.count
github_events.printSchema()

The result looks like the following:

The data has 414,826 records. As before, notice that there are eight top-level columns, and three partition columns. If you scroll down, you’ll also notice that the payload is the most complex column.

Run functions and filter records

This section describes how you can create your own functions and invoke them seamlessly to filter records. Unlike filtering with Python lambdas, Scala scripts do not need to convert records from one language representation to another, thereby reducing overhead and running much faster.

Let’s create a function that picks only the IssuesEvents from the GitHub timeline. These events are generated whenever someone posts an issue for a particular repository. Each GitHub event record has a field, “type”, that indicates the kind of event it is. The issueFilter() function returns true for records that are IssuesEvents.

def issueFilter(rec: DynamicRecord): Boolean = { 
    rec.getField("type").exists(_ == "IssuesEvent") 
}

Note that the getField() method returns an Option[Any] type, so you first need to check that it exists before checking the type.

You pass this function to the filter transformation, which applies the function on each record and returns a dynamic frame of those records that pass.

val issue_events =  github_events.filter(issueFilter)

Now, let’s look at the size and schema of issue_events.

issue_events.count
issue_events.printSchema()

It’s much smaller (14,063 records), and the payload schema is less complex, reflecting only the schema for issues. Keep a few essential columns for your analysis, and drop the rest using the ApplyMapping() transform:

val issue_titles = issue_events.applyMapping(Seq(("id", "string", "id", "string"),
                                                 ("actor.login", "string", "actor", "string"), 
                                                 ("repo.name", "string", "repo", "string"),
                                                 ("payload.action", "string", "action", "string"),
                                                 ("payload.issue.title", "string", "title", "string")))
issue_titles.show()

The ApplyMapping() transform is quite handy for renaming columns, casting types, and restructuring records. The preceding code snippet tells the transform to select the fields (or columns) that are enumerated in the left half of the tuples and map them to the fields and types in the right half.

Estimating sentiment using Stanford CoreNLP

To focus on the most pressing issues, you might want to isolate the records with the most negative sentiments. The Stanford CoreNLP libraries are Java-based and offer sentiment-prediction functions. Accessing these functions through Python is possible, but quite cumbersome. It requires creating Python surrogate classes and objects for those found on the Java side. Instead, with Scala support, you can use those classes and objects directly and invoke their methods. Let’s see how.

First, import the libraries needed for the analysis:

import java.util.Properties
import edu.stanford.nlp.ling.CoreAnnotations
import edu.stanford.nlp.neural.rnn.RNNCoreAnnotations
import edu.stanford.nlp.pipeline.{Annotation, StanfordCoreNLP}
import edu.stanford.nlp.sentiment.SentimentCoreAnnotations
import scala.collection.convert.wrapAll._

The Stanford CoreNLP libraries have a main driver that orchestrates all of their analysis. The driver setup is heavyweight, setting up threads and data structures that are shared across analyses. Apache Spark runs on a cluster with a main driver process and a collection of backend executor processes that do most of the heavy sifting of the data.

The Stanford CoreNLP shared objects are not serializable, so they cannot be distributed easily across a cluster. Instead, you need to initialize them once for every backend executor process that might need them. Here is how to accomplish that:

val props = new Properties()
props.setProperty("annotators", "tokenize, ssplit, parse, sentiment")
props.setProperty("parse.maxlen", "70")

object myNLP {
    lazy val coreNLP = new StanfordCoreNLP(props)
}

The properties tell the libraries which annotators to execute and how many words to process. The preceding code creates an object, myNLP, with a field coreNLP that is lazily evaluated. This field is initialized only when it is needed, and only once. So, when the backend executors start processing the records, each executor initializes the driver for the Stanford CoreNLP libraries only one time.

Next is a function that estimates the sentiment of a text string. It first calls Stanford CoreNLP to annotate the text. Then, it pulls out the sentences and takes the average sentiment across all the sentences. The sentiment is a double, from 0.0 as the most negative to 4.0 as the most positive.

def estimatedSentiment(text: String): Double = {
    if ((text == null) || (!text.nonEmpty)) { return Double.NaN }
    val annotations = myNLP.coreNLP.process(text)
    val sentences = annotations.get(classOf[CoreAnnotations.SentencesAnnotation])
    sentences.foldLeft(0.0)( (csum, x) => { 
        csum + RNNCoreAnnotations.getPredictedClass(x.get(classOf[SentimentCoreAnnotations.SentimentAnnotatedTree])) 
    }) / sentences.length
}

Now, let’s estimate the sentiment of the issue titles and add that computed field as part of the records. You can accomplish this with the map() method on dynamic frames:

val issue_sentiments = issue_titles.map((rec: DynamicRecord) => { 
    val mbody = rec.getField("title")
    mbody match {
        case Some(mval: String) => { 
            rec.addField("sentiment", ScalarNode(estimatedSentiment(mval)))
            rec }
        case _ => rec
    }
})

The map() method applies the user-provided function on every record. The function takes a DynamicRecord as an argument and returns a DynamicRecord. The code above computes the sentiment, adds it in a top-level field, sentiment, to the record, and returns the record.

Count the records with sentiment and show the schema. This takes a few minutes because Spark must initialize the library and run the sentiment analysis, which can be involved.

issue_sentiments.count
issue_sentiments.printSchema()

Notice that all records were processed (14,063), and the sentiment value was added to the schema.

Finally, let’s pick out the titles that have the lowest sentiment (less than 1.5). Count them and print out a sample to see what some of the titles look like.

val pressing_issues = issue_sentiments.filter(_.getField("sentiment").exists(_.asInstanceOf[Double] < 1.5))
pressing_issues.count
pressing_issues.show(10)

Next, write them all to a file so that you can handle them later. (You’ll need to replace the output path with your own.)

glueContext.getSinkWithFormat(connectionType = "s3", 
                              options = JsonOptions("""{"path": "s3://<bucket>/out/path/"}"""), 
                              format = "json")
            .writeDynamicFrame(pressing_issues)

Take a look in the output path, and you can see the output files.

Putting it all together

Now, let’s create a job from the preceding interactive session. The following script combines all the commands from earlier. It processes the GitHub archive files and writes out the highly negative issues:

import com.amazonaws.services.glue.DynamicRecord
import com.amazonaws.services.glue.GlueContext
import com.amazonaws.services.glue.util.GlueArgParser
import com.amazonaws.services.glue.util.Job
import com.amazonaws.services.glue.util.JsonOptions
import com.amazonaws.services.glue.types._
import org.apache.spark.SparkContext
import java.util.Properties
import edu.stanford.nlp.ling.CoreAnnotations
import edu.stanford.nlp.neural.rnn.RNNCoreAnnotations
import edu.stanford.nlp.pipeline.{Annotation, StanfordCoreNLP}
import edu.stanford.nlp.sentiment.SentimentCoreAnnotations
import scala.collection.convert.wrapAll._

object GlueApp {

    object myNLP {
        val props = new Properties()
        props.setProperty("annotators", "tokenize, ssplit, parse, sentiment")
        props.setProperty("parse.maxlen", "70")

        lazy val coreNLP = new StanfordCoreNLP(props)
    }

    def estimatedSentiment(text: String): Double = {
        if ((text == null) || (!text.nonEmpty)) { return Double.NaN }
        val annotations = myNLP.coreNLP.process(text)
        val sentences = annotations.get(classOf[CoreAnnotations.SentencesAnnotation])
        sentences.foldLeft(0.0)( (csum, x) => { 
            csum + RNNCoreAnnotations.getPredictedClass(x.get(classOf[SentimentCoreAnnotations.SentimentAnnotatedTree])) 
        }) / sentences.length
    }

    def main(sysArgs: Array[String]) {
        val spark: SparkContext = SparkContext.getOrCreate()
        val glueContext: GlueContext = new GlueContext(spark)

        val dbname = "githubarchive"
        val tblname = "data"
        val outpath = "s3://<bucket>/out/path/"

        val github_events = glueContext
                            .getCatalogSource(database = dbname, tableName = tblname)
                            .getDynamicFrame()

        val issue_events =  github_events.filter((rec: DynamicRecord) => {
            rec.getField("type").exists(_ == "IssuesEvent")
        })

        val issue_titles = issue_events.applyMapping(Seq(("id", "string", "id", "string"),
                                                         ("actor.login", "string", "actor", "string"), 
                                                         ("repo.name", "string", "repo", "string"),
                                                         ("payload.action", "string", "action", "string"),
                                                         ("payload.issue.title", "string", "title", "string")))

        val issue_sentiments = issue_titles.map((rec: DynamicRecord) => { 
            val mbody = rec.getField("title")
            mbody match {
                case Some(mval: String) => { 
                    rec.addField("sentiment", ScalarNode(estimatedSentiment(mval)))
                    rec }
                case _ => rec
            }
        })

        val pressing_issues = issue_sentiments.filter(_.getField("sentiment").exists(_.asInstanceOf[Double] < 1.5))

        glueContext.getSinkWithFormat(connectionType = "s3", 
                              options = JsonOptions(s"""{"path": "$outpath"}"""), 
                              format = "json")
                    .writeDynamicFrame(pressing_issues)
    }
}

Notice that the script is enclosed in a top-level object called GlueApp, which serves as the script’s entry point for the job. (You’ll need to replace the output path with your own.) Upload the script to an Amazon S3 location so that AWS Glue can load it when needed.

To create the job, open the AWS Glue console. Choose Jobs in the left navigation pane, and then choose Add job. Create a name for the job, and specify a role with permissions to access the data. Choose An existing script that you provide, and choose Scala as the language.

For the Scala class name, type GlueApp to indicate the script’s entry point. Specify the Amazon S3 location of the script.

Choose Script libraries and job parameters. In the Dependent jars path field, enter the Amazon S3 locations of the Stanford CoreNLP libraries from earlier as a comma-separated list (without spaces). Then choose Next.

No connections are needed for this job, so choose Next again. Review the job properties, and choose Finish. Finally, choose Run job to execute the job.

You can simply edit the script’s input table and output path to run this job on whatever GitHub timeline datasets that you might have.

Conclusion

In this post, we showed how to write AWS Glue ETL scripts in Scala via notebooks and how to run them as jobs. Scala has the advantage that it is the native language for the Spark runtime. With Scala, it is easier to call Scala or Java functions and third-party libraries for analyses. Moreover, data processing is faster in Scala because there’s no need to convert records from one language runtime to another.

You can find more example of Scala scripts in our GitHub examples repository: https://github.com/awslabs/aws-glue-samples. We encourage you to experiment with Scala scripts and let us know about any interesting ETL flows that you want to share.

Happy Glue-ing!

 


Additional Reading

If you found this post useful, be sure to check out Simplify Querying Nested JSON with the AWS Glue Relationalize Transform and Genomic Analysis with Hail on Amazon EMR and Amazon Athena.

 


About the Authors

Mehul Shah is a senior software manager for AWS Glue. His passion is leveraging the cloud to build smarter, more efficient, and easier to use data systems. He has three girls, and, therefore, he has no spare time.

 

 

 

Ben Sowell is a software development engineer at AWS Glue.

 

 

 

 
Vinay Vivili is a software development engineer for AWS Glue.

 

 

 

Daniel Miessler on My Writings about IoT Security

Post Syndicated from Bruce Schneier original https://www.schneier.com/blog/archives/2018/01/daniel_miessler.html

Daniel Miessler criticizes my writings about IoT security:

I know it’s super cool to scream about how IoT is insecure, how it’s dumb to hook up everyday objects like houses and cars and locks to the internet, how bad things can get, and I know it’s fun to be invited to talk about how everything is doom and gloom.

I absolutely respect Bruce Schneier a lot for what he’s contributed to InfoSec, which makes me that much more disappointed with this kind of position from him.

InfoSec is full of those people, and it’s beneath people like Bruce to add their voices to theirs. Everyone paying attention already knows it’s going to be a soup sandwich — a carnival of horrors — a tragedy of mistakes and abuses of trust.

It’s obvious. Not interesting. Not novel. Obvious. But obvious or not, all these things are still going to happen.

I actually agree with everything in his essay. “We should obviously try to minimize the risks, but we don’t do that by trying to shout down the entire enterprise.” Yes, definitely.

I don’t think the IoT must be stopped. I do think that the risks are considerable, and will increase as these systems become more pervasive and susceptible to class breaks. And I’m trying to write a book that will help navigate this. I don’t think I’m the prophet of doom, and don’t want to come across that way. I’ll give the manuscript another read with that in mind.

Physics cheats

Post Syndicated from Eevee original https://eev.ee/blog/2018/01/06/physics-cheats/

Anonymous asks:

something about how we tweak physics to “work” better in games?

Ho ho! Work. Get it? Like in physics…?

Hitboxes

Hitbox” is perhaps not the most accurate term, since the shape used for colliding with the environment and the shape used for detecting damage might be totally different. They’re usually the same in simple platformers, though, and that’s what most of my games have been.

The hitbox is the biggest physics fudge by far, and it exists because of a single massive approximation that (most) games make: you’re controlling a single entity in the abstract, not a physical body in great detail.

That is: when you walk with your real-world meat shell, you perform a complex dance of putting one foot in front of the other, a motion you spent years perfecting. When you walk in a video game, you press a single “walk” button. Your avatar may play an animation that moves its legs back and forth, but since you’re not actually controlling the legs independently (and since simulating them is way harder), the game just treats you like a simple shape. Fairly often, this is a box, or something very box-like.

An Eevee sprite standing on faux ground; the size of the underlying image and the hitbox are outlined

Since the player has no direct control over the exact placement of their limbs, it would be slightly frustrating to have them collide with the world. This is especially true in cases like the above, where the tail and left ear protrude significantly out from the main body. If that Eevee wanted to stand against a real-world wall, she would simply tilt her ear or tail out of the way, so there’s no reason for the ear to block her from standing against a game wall. To compensate for this, the ear and tail are left out of the collision box entirely and will simply jut into a wall if necessary — a goofy affordance that’s so common it doesn’t even register as unusual. As a bonus (assuming this same box is used for combat), she won’t take damage from projectiles that merely graze past an ear.

(One extra consideration for sprite games in particular: the hitbox ought to be horizontally symmetric around the sprite’s pivot — i.e. the point where the entity is truly considered to be standing — so that the hitbox doesn’t abruptly move when the entity turns around!)

Corners

Treating the player (and indeed most objects) as a box has one annoying side effect: boxes have corners. Corners can catch on other corners, even by a single pixel. Real-world bodies tend to be a bit rounder and squishier and this can tolerate grazing a corner; even real-world boxes will simply rotate a bit.

Ah, but in our faux physics world, we generally don’t want conscious actors (such as the player) to rotate, even with a realistic physics simulator! Real-world bodies are made of parts that will generally try to keep you upright, after all; you don’t tilt back and forth much.

One way to handle corners is to simply remove them from conscious actors. A hitbox doesn’t have to be a literal box, after all. A popular alternative — especially in Unity where it’s a standard asset — is the pill-shaped capsule, which has semicircles/hemispheres on the top and bottom and a cylindrical body in 3D. No corners, no problem.

Of course, that introduces a new problem: now the player can’t balance precariously on edges without their rounded bottom sliding them off. Alas.

If you’re stuck with corners, then, you may want to use a corner bump, a term I just made up. If the player would collide with a corner, but the collision is only by a few pixels, just nudge them to the side a bit and carry on.

An Eevee sprite trying to move sideways into a shallow ledge; the game bumps her upwards slightly, so she steps onto it instead

When the corner is horizontal, this creates stairs! This is, more or less kinda, how steps work in Doom: when the player tries to cross from one sector into another, if the height difference is 24 units or less, the game simply bumps them upwards to the height of the new floor and lets them continue on.

Implementing this in a game without Doom’s notion of sectors is a little trickier. In fact, I still haven’t done it. Collision detection based on rejection gets it for free, kinda, but it’s not very deterministic and it breaks other things. But that’s a whole other post.

Gravity

Gravity is pretty easy. Everything accelerates downwards all the time. What’s interesting are the exceptions.

Jumping

Jumping is a giant hack.

Think about how actual jumping works: you tense your legs, which generally involves bending your knees first, and then spring upwards. In a platformer, you can just leap whenever you feel like it, which is nonsense. Also you go like twenty feet into the air?

Worse, most platformers allow variable-height jumping, where your jump is lower if you let go of the jump button while you’re in the air. Normally, one would expect to have to decide how much force to put into the jump beforehand.

But of course this is about convenience of controls: when jumping is your primary action, you want to be able to do it immediately, without any windup for how high you want to jump.

(And then there’s double jumping? Come on.)

Air control is a similar phenomenon: usually you’d jump in a particular direction by controlling how you push off the ground with your feet, but in a video game, you don’t have feet! You only have the box. The compromise is to let you control your horizontal movement to a limit degree in midair, even though that doesn’t make any sense. (It’s way more fun, though, and overall gives you more movement options, which are good to have in an interactive medium.)

Air control also exposes an obvious place that game physics collide with the realistic model of serious physics engines. I’ve mentioned this before, but: if you use Real Physics™ and air control yourself into a wall, you might find that you’ll simply stick to the wall until you let go of the movement buttons. Why? Remember, player movement acts as though an external force were pushing you around (and from the perspective of a Real™ physics engine, this is exactly how you’d implement it) — so air-controlling into a wall is equivalent to pushing a book against a wall with your hand, and the friction with the wall holds you in place. Oops.

Ground sticking

Another place game physics conflict with physics engines is with running to the top of a slope. On a real hill, of course, you land on top of the slope and are probably glad of it; slopes are hard to climb!

An Eevee moves to the top of a slope, and rather than step onto the flat top, she goes flying off into the air

In a video game, you go flying. Because you’re a box. With momentum. So you hit the peak and keep going in the same direction. Which is diagonally upwards.

Projectiles

To make them more predictable, projectiles generally aren’t subject to gravity, at least as far as I’ve seen. The real world does not have such an exemption. The real world imposes gravity even on sniper rifles, which in a video game are often implemented as an instant trace unaffected by anything in the world because the bullet never actually exists in the world.

Resistance

Ah. Welcome to hell.

Water

Water is an interesting case, and offhand I don’t know the gritty details of how games implement it. In the real world, water applies a resistant drag force to movement — and that force is proportional to the square of velocity, which I’d completely forgotten until right now. I am almost positive that no game handles that correctly. But then, in real-world water, you can push against the water itself for movement, and games don’t simulate that either. What’s the rough equivalent?

The Sonic Physics Guide suggests that Sonic handles it by basically halving everything: acceleration, max speed, friction, etc. When Sonic enters water, his speed is cut; when Sonic exits water, his speed is increased.

That last bit feels validating — I could swear Metroid Prime did the same thing, and built my own solution around it, but couldn’t remember for sure. It makes no sense, of course, for a jump to become faster just because you happened to break the surface of the water, but it feels fantastic.

The thing I did was similar, except that I didn’t want to add a multiplier in a dozen places when you happen to be underwater (and remember which ones need it to be squared, etc.). So instead, I calculate everything completely as normal, so velocity is exactly the same as it would be on dry land — but the distance you would move gets halved. The effect seems to be pretty similar to most platformers with water, at least as far as I can tell. It hasn’t shown up in a published game and I only added this fairly recently, so I might be overlooking some reason this is a bad idea.

(One reason that comes to mind is that velocity is now a little white lie while underwater, so anything relying on velocity for interesting effects might be thrown off. Or maybe that’s correct, because velocity thresholds should be halved underwater too? Hm!)

Notably, air is also a fluid, so it should behave the same way (just with different constants). I definitely don’t think any games apply air drag that’s proportional to the square of velocity.

Friction

Friction is, in my experience, a little handwaved. Probably because real-world friction is so darn complicated.

Consider that in the real world, we want very high friction on the surfaces we walk on — shoes and tires are explicitly designed to increase it, even. We move by bracing a back foot against the ground and using that to push ourselves forward, so we want the ground to resist our push as much as possible.

In a game world, we are a box. We move by being pushed by some invisible outside force, so if the friction between ourselves and the ground is too high, we won’t be able to move at all! That’s complete nonsense physically, but it turns out to be handy in some cases — for example, highish friction can simulate walking through deep mud, which should be difficult due to fluid drag and low friction.

But the best-known example of the fakeness of game friction is video game ice. Walking on real-world ice is difficult because the low friction means low grip; your feet are likely to slip out from under you, and you’ll simply fall down and have trouble moving at all. In a video game, you can’t fall down, so you have the opposite experience: you spend most of your time sliding around uncontrollably. Yet ice is so common in video games (and perhaps so uncommon in places I’ve lived) that I, at least, had never really thought about this disparity until an hour or so ago.

Game friction vs real-world friction

Real-world friction is a force. It’s the normal force (which is the force exerted by the object on the surface) times some constant that depends on how the two materials interact.

Force is mass times acceleration, and platformers often ignore mass, so friction ought to be an acceleration — applied against the object’s movement, but never enough to push it backwards.

I haven’t made any games where variable friction plays a significant role, but my gut instinct is that low friction should mean the player accelerates more slowly but has a higher max speed, and high friction should mean the opposite. I see from my own source code that I didn’t even do what I just said, so let’s defer to some better-made and well-documented games: Sonic and Doom.

In Sonic, friction is a fixed value subtracted from the player’s velocity (regardless of direction) each tic. Sonic has a fixed framerate, so the units are really pixels per tic squared (i.e. acceleration), multiplied by an implicit 1 tic per tic. So far, so good.

But Sonic’s friction only applies if the player isn’t pressing or . Hang on, that isn’t friction at all; that’s just deceleration! That’s equivalent to jogging to a stop. If friction were lower, Sonic would take longer to stop, but otherwise this is only tangentially related to friction.

(In fairness, this approach would decently emulate friction for non-conscious sliding objects, which are never going to be pressing movement buttons. Also, we don’t have the Sonic source code, and the name “friction” is a fan invention; the Sonic Physics Guide already uses “deceleration” to describe the player’s acceleration when turning around.)

Okay, let’s try Doom. In Doom, the default friction is 90.625%.

Hang on, what?

Yes, in Doom, friction is a multiplier applied every tic. Doom runs at 35 tics per second, so this is a multiplier of 0.032 per second. Yikes!

This isn’t anything remotely like real friction, but it’s much easier to implement. With friction as acceleration, the game has to know both the direction of movement (so it can apply friction in the opposite direction) and the magnitude (so it doesn’t overshoot and launch the object in the other direction). That means taking a semi-costly square root and also writing extra code to cap the amount of friction. With a multiplier, neither is necessary; just multiply the whole velocity vector and you’re done.

There are some downsides. One is that objects will never actually stop, since multiplying by 3% repeatedly will never produce a result of zero — though eventually the speed will become small enough to either slip below a “minimum speed” threshold or simply no longer fit in a float representation. Another is that the units are fairly meaningless: with Doom’s default friction of 90.625%, about how long does it take for the player to stop? I have no idea, partly because “stop” is ambiguous here! If friction were an acceleration, I could divide it into the player’s max speed to get a time.

All that aside, what are the actual effects of changing Doom’s friction? What an excellent question that’s surprisingly tricky to answer. (Note that friction can’t be changed in original Doom, only in the Boom port and its derivatives.) Here’s what I’ve pieced together.

Doom’s “friction” is really two values. “Friction” itself is a multiplier applied to moving objects on every tic, but there’s also a move factor which defaults to \(\frac{1}{32} = 0.03125\) and is derived from friction for custom values.

Every tic, the player’s velocity is multiplied by friction, and then increased by their speed times the move factor.

$$
v(n) = v(n – 1) \times friction + speed \times move factor
$$

Eventually, the reduction from friction will balance out the speed boost. That happens when \(v(n) = v(n – 1)\), so we can rearrange it to find the player’s effective max speed:

$$
v = v \times friction + speed \times move factor \\
v – v \times friction = speed \times move factor \\
v = speed \times \frac{move factor}{1 – friction}
$$

For vanilla Doom’s move factor of 0.03125 and friction of 0.90625, that becomes:

$$
v = speed \times \frac{\frac{1}{32}}{1 – \frac{29}{32}} = speed \times \frac{\frac{1}{32}}{\frac{3}{32}} = \frac{1}{3} \times speed
$$

Curiously, “speed” is three times the maximum speed an actor can actually move. Doomguy’s run speed is 50, so in practice he moves a third of that, or 16⅔ units per tic. (Of course, this isn’t counting SR40, a bug that lets Doomguy run ~40% faster than intended diagonally.)

So now, what if you change friction? Even more curiously, the move factor is calculated completely differently depending on whether friction is higher or lower than the default Doom amount:

$$
move factor = \begin{cases}
\frac{133 – 128 \times friction}{544} &≈ 0.244 – 0.235 \times friction & \text{ if } friction \ge \frac{29}{32} \\
\frac{81920 \times friction – 70145}{1048576} &≈ 0.078 \times friction – 0.067 & \text{ otherwise }
\end{cases}
$$

That’s pretty weird? Complicating things further is that low friction (which means muddy terrain, remember) has an extra multiplier on its move factor, depending on how fast you’re already going — the idea is apparently that you have a hard time getting going, but it gets easier as you find your footing. The extra multiplier maxes out at 8, which makes the two halves of that function meet at the vanilla Doom value.

A graph of the relationship between friction and move factor

That very top point corresponds to the move factor from the original game. So no matter what you do to friction, the move factor becomes lower. At 0.85 and change, you can no longer move at all; below that, you move backwards.

From the formula above, it’s easy to see what changes to friction and move factor will do to Doomguy’s stable velocity. Move factor is in the numerator, so increasing it will increase stable velocity — but it can’t increase, so stable velocity can only ever decrease. Friction is in the denominator, but it’s subtracted from 1, so increasing friction will make the denominator a smaller value less than 1, i.e. increase stable velocity. Combined, we get this relationship between friction and stable velocity.

A graph showing stable velocity shooting up dramatically as friction increases

As friction approaches 1, stable velocity grows without bound. This makes sense, given the definition of \(v(n)\) — if friction is 1, the velocity from the previous tic isn’t reduced at all, so we just keep accelerating freely.

All of this is why I’m wary of using multipliers.

Anyway, this leaves me with one last question about the effects of Doom’s friction: how long does it take to reach stable velocity? Barring precision errors, we’ll never truly reach stable velocity, but let’s say within 5%. First we need a closed formula for the velocity after some number of tics. This is a simple recurrence relation, and you can write a few terms out yourself if you want to be sure this is right.

$$
v(n) = v_0 \times friction^n + speed \times move factor \times \frac{friction^n – 1}{friction – 1}
$$

Our initial velocity is zero, so the first term disappears. Set this equal to the stable formula and solve for n:

$$
speed \times move factor \times \frac{friction^n – 1}{friction – 1} = (1 – 5\%) \times speed \times \frac{move factor}{1 – friction} \\
friction^n – 1 = -(1 – 5\%) \\
n = \frac{\ln 5\%}{\ln friction}
$$

Speed” and move factor disappear entirely, which makes sense, and this is purely a function of friction (and how close we want to get). For vanilla Doom, that comes out to 30.4, which is a little less than a second. For other values of friction:

A graph of time to stability which leaps upwards dramatically towards the right

As friction increases (which in Doom terms means the surface is more slippery), it takes longer and longer to reach stable speed, which is in turn greater and greater. For lesser friction (i.e. mud), stable speed is lower, but reached fairly quickly. (Of course, the extra “getting going” multiplier while in mud adds some extra time here, but including that in the graph is a bit more complicated.)

I think this matches with my instincts above. How fascinating!

What’s that? This is way too much math and you hate it? Then don’t use multipliers in game physics.

Uh

That was a hell of a diversion!

I guess the goofiest stuff in basic game physics is really just about mapping player controls to in-game actions like jumping and deceleration; the rest consists of hacks to compensate for representing everything as a box.

How to Encrypt Amazon S3 Objects with the AWS SDK for Ruby

Post Syndicated from Doug Schwartz original https://aws.amazon.com/blogs/security/how-to-encrypt-amazon-s3-objects-with-the-aws-sdk-for-ruby/

AWS KMS image

Recently, Amazon announced some new Amazon S3 encryption and security features. The AWS Blog post showed how to use the Amazon S3 console to take advantage of these new features. However, if you have a large number of Amazon S3 buckets, using the console to implement these features could take hours, if not days. As an alternative, I created documentation topics in the AWS SDK for Ruby Developer Guide that include code examples showing you how to use the new Amazon S3 encryption features using the AWS SDK for Ruby.

What are my encryption options?

You can encrypt Amazon S3 bucket objects on a server or on a client:

  • When you encrypt objects on a server, you request that Amazon S3 encrypt the objects before saving them to disk in data centers and decrypt the objects when you download them. The main advantage of this approach is that Amazon S3 manages the entire encryption process.
  • When you encrypt objects on a client, you encrypt the objects before you upload them to Amazon S3. In this case, you manage the encryption process, the encryption keys, and related tools. Use this option when:
    • Company policy and standards require it.
    • You already have a development process in place that meets your needs.

    Encrypting on the client has always been available, but you should know the following points:

    • You must be diligent about protecting your encryption keys, which is analogous to having a burglar-proof lock on your front door. If you leave a key under the mat, your security is compromised.
    • If you lose your encryption keys, you won’t be able to decrypt your data.

    If you encrypt objects on the client, we strongly recommend that you use an AWS Key Management Service (AWS KMS) managed customer master key (CMK)

How to use encryption on a server

You can specify that Amazon S3 automatically encrypts objects as you upload them to a bucket or require that objects uploaded to an Amazon S3 bucket include encryption on a server before they are uploaded to an Amazon S3 bucket.

The advantage of these settings is that when you specify them, you ensure that objects uploaded to Amazon S3 are encrypted. Alternatively, you can have Amazon S3 encrypt individual objects on the server as you upload them to a bucket or encrypt them on the server with your own key as you upload them to a bucket.

The AWS SDK for Ruby Developer Guide now contains the following topics that explain your encryption options on a server:

How to use encryption on a client

You can encrypt objects on a client before you upload them to a bucket and decrypt them after you download them from a bucket by using the Amazon S3 encryption client.

The AWS SDK for Ruby Developer Guide now contains the following topics that explain your encryption options on the client:

Note: The Amazon S3 encryption client in the AWS SDK for Ruby is compatible with other Amazon S3 encryption clients, but it is not compatible with other AWS client-side encryption libraries, including the AWS Encryption SDK and the Amazon DynamoDB encryption client for Java. Each library returns a different ciphertext (“encrypted message”) format, so you can’t use one library to encrypt objects and a different library to decrypt them. For more information, see Protecting Data Using Client-Side Encryption.

If you have comments about this blog post, submit them in the “Comments” section below. If you have questions about encrypting objects on servers and clients, start a new thread on the Amazon S3 forum or contact AWS Support.

– Doug

Using Trusted Timestamping With Java

Post Syndicated from Bozho original https://techblog.bozho.net/using-trusted-timestamping-java/

Trusted timestamping is the process of having a trusted third party (“Time stamping authority”, TSA) certify the time of a given event in electronic form. The EU regulation eIDAS gives these timestamps legal strength – i.e. nobody can dispute the time or the content of the event if it was timestamped. It is applicable to multiple scenarios, including timestamping audit logs. (Note: timestamping is not sufficient for a good audit trail as it does not prevent a malicious actor from deleting the event altogether)

There are a number of standards for trusted timestamping, the core one being RFC 3161. As most RFCs it is hard to read. Fortunately for Java users, BouncyCastle implements the standard. Unfortunately, as with most security APIs, working with it is hard, even abysmal. I had to implement it, so I’ll share the code needed to timestamp data.

The whole gist can be found here, but I’ll try to explain the main flow. Obviously, there is a lot of code that’s there to simply follow the standard. The BouncyCastle classes are a maze that’s hard to navigate.

The main method is obviously timestamp(hash, tsaURL, username, password):

public TimestampResponseDto timestamp(byte[] hash, String tsaUrl, String tsaUsername, String tsaPassword) throws IOException {
    MessageImprint imprint = new MessageImprint(sha512oid, hash);

    TimeStampReq request = new TimeStampReq(imprint, null, new ASN1Integer(random.nextLong()),
            ASN1Boolean.TRUE, null);

    byte[] body = request.getEncoded();
    try {
        byte[] responseBytes = getTSAResponse(body, tsaUrl, tsaUsername, tsaPassword);

        ASN1StreamParser asn1Sp = new ASN1StreamParser(responseBytes);
        TimeStampResp tspResp = TimeStampResp.getInstance(asn1Sp.readObject());
        TimeStampResponse tsr = new TimeStampResponse(tspResp);

        checkForErrors(tsaUrl, tsr);

        // validate communication level attributes (RFC 3161 PKIStatus)
        tsr.validate(new TimeStampRequest(request));

        TimeStampToken token = tsr.getTimeStampToken();
            
        TimestampResponseDto response = new TimestampResponseDto();
        response.setTime(getSigningTime(token.getSignedAttributes()));
        response.setEncodedToken(Base64.getEncoder().encodeToString(token.getEncoded()));
           
        return response;
    } catch (RestClientException | TSPException | CMSException | OperatorCreationException | GeneralSecurityException e) {
        throw new IOException(e);
    }
}

It prepares the request by creating the message imprint. Note that you are passing the hash itself, but also the hashing algorithm used to make the hash. Why isn’t the API hiding that from you, I don’t know. In my case the hash is obtained in a more complicated way, so it’s useful, but still. Then we get the raw form of the request and send it to the TSA (time stamping authority). It is an HTTP request, sort of simple, but you have to take care of some request and response headers that are not necessarily consistent across TSAs. The username and password are optional, some TSAs offer the service (rate-limited) without authentication.

When you have the raw response back, you parse it to a TimeStampResponse. Again, you have to go through 2 intermediate objects (ASN1StreamParser and TimeStampResp), which may be a proper abstraction, but is not a usable API.

Then you check if the response was successful, and you also have to validate it – the TSA may have returned a bad response. Ideally all of that could’ve been hidden from you. Validation throws an exception, which in this case I just propagate by wrapping in an IOException.

Finally, you get the token and return the response. The most important thing is the content of the token, which in my case was needed as Base64, so I encode it. It could just be the raw bytes as well. If you want to get any additional data from the token (e.g. the signing time), it’s not that simple; you have to parse the low-level attributes (seen in the gist).

Okay, you have the token now, and you can store it in a database. Occasionally you may want to validate whether timestamps have not been tampered with (which is my usecase). The code is here, and I won’t even try to explain it – it’s a ton of boilerplate that is also accounting for variations in the way TSAs respond (I’ve tried a few). The fact that a DummyCertificate class is needed either means I got something very wrong, or confirms my critique for the BouncyCastle APIs. The DummyCertificate may not be needed for some TSAs, but it is for others, and you actually can’t instantiate it that easily. You need a real certificate to construct it (which is not included in the gist; using the init() method in the next gist you can create the dummy with dummyCertificate = new DummyCertificate(certificateHolder.toASN1Structure());). In my code these are all one class, but for presenting them I decided to split it, hence this little duplication.

Okay, now we can timestamp and validate timestamps. That should be enough; but for testing purposes (or limited internal use) you may want to do the timestamping locally instead of asking a TSA. The code can be found here. It uses spring, but you can instead pass the keystore details as arguments to the init method. You need a JKS store with a keypair and a certificate, and I used KeyStore Explorer to create them. If you are running your application in AWS, you may want to encrypt your keystore using KMS (Key Management Service), and then decrypt it on application load, but that’s out of the scope of this article. For the local timestamping validation works as expected, and for timestamping – instead of calling the external service, just call localTSA.timestamp(req);

How did I get to know which classes to instantiate and which parameters to pass – I don’t remember. Looking at tests, examples, answers, sources. It took a while, and so I’m sharing it, to potentially save some trouble of others.

A list of TSAs you can test with: SafeCreative, FreeTSA, time.centum.pl.

I realize this does not seem applicable to many scenarios, but I would recommend timestamping some critical pieces of your application data. And it is generally useful to have it in your “toolbox”, ready to use, rather than trying to read the standard and battling with BouncyCastle classes for days in order to achieve this allegedly simple task.

The post Using Trusted Timestamping With Java appeared first on Bozho's tech blog.

How to Enhance the Security of Sensitive Customer Data by Using Amazon CloudFront Field-Level Encryption

Post Syndicated from Alex Tomic original https://aws.amazon.com/blogs/security/how-to-enhance-the-security-of-sensitive-customer-data-by-using-amazon-cloudfront-field-level-encryption/

Amazon CloudFront is a web service that speeds up distribution of your static and dynamic web content to end users through a worldwide network of edge locations. CloudFront provides a number of benefits and capabilities that can help you secure your applications and content while meeting compliance requirements. For example, you can configure CloudFront to help enforce secure, end-to-end connections using HTTPS SSL/TLS encryption. You also can take advantage of CloudFront integration with AWS Shield for DDoS protection and with AWS WAF (a web application firewall) for protection against application-layer attacks, such as SQL injection and cross-site scripting.

Now, CloudFront field-level encryption helps secure sensitive data such as a customer phone numbers by adding another security layer to CloudFront HTTPS. Using this functionality, you can help ensure that sensitive information in a POST request is encrypted at CloudFront edge locations. This information remains encrypted as it flows to and beyond your origin servers that terminate HTTPS connections with CloudFront and throughout the application environment. In this blog post, we demonstrate how you can enhance the security of sensitive data by using CloudFront field-level encryption.

Note: This post assumes that you understand concepts and services such as content delivery networks, HTTP forms, public-key cryptography, CloudFrontAWS Lambda, and the AWS CLI. If necessary, you should familiarize yourself with these concepts and review the solution overview in the next section before proceeding with the deployment of this post’s solution.

How field-level encryption works

Many web applications collect and store data from users as those users interact with the applications. For example, a travel-booking website may ask for your passport number and less sensitive data such as your food preferences. This data is transmitted to web servers and also might travel among a number of services to perform tasks. However, this also means that your sensitive information may need to be accessed by only a small subset of these services (most other services do not need to access your data).

User data is often stored in a database for retrieval at a later time. One approach to protecting stored sensitive data is to configure and code each service to protect that sensitive data. For example, you can develop safeguards in logging functionality to ensure sensitive data is masked or removed. However, this can add complexity to your code base and limit performance.

Field-level encryption addresses this problem by ensuring sensitive data is encrypted at CloudFront edge locations. Sensitive data fields in HTTPS form POSTs are automatically encrypted with a user-provided public RSA key. After the data is encrypted, other systems in your architecture see only ciphertext. If this ciphertext unintentionally becomes externally available, the data is cryptographically protected and only designated systems with access to the private RSA key can decrypt the sensitive data.

It is critical to secure private RSA key material to prevent unauthorized access to the protected data. Management of cryptographic key material is a larger topic that is out of scope for this blog post, but should be carefully considered when implementing encryption in your applications. For example, in this blog post we store private key material as a secure string in the Amazon EC2 Systems Manager Parameter Store. The Parameter Store provides a centralized location for managing your configuration data such as plaintext data (such as database strings) or secrets (such as passwords) that are encrypted using AWS Key Management Service (AWS KMS). You may have an existing key management system in place that you can use, or you can use AWS CloudHSM. CloudHSM is a cloud-based hardware security module (HSM) that enables you to easily generate and use your own encryption keys in the AWS Cloud.

To illustrate field-level encryption, let’s look at a simple form submission where Name and Phone values are sent to a web server using an HTTP POST. A typical form POST would contain data such as the following.

POST / HTTP/1.1
Host: example.com
Content-Type: application/x-www-form-urlencoded
Content-Length:60

Name=Jane+Doe&Phone=404-555-0150

Instead of taking this typical approach, field-level encryption converts this data similar to the following.

POST / HTTP/1.1
Host: example.com
Content-Type: application/x-www-form-urlencoded
Content-Length: 1713

Name=Jane+Doe&Phone=AYABeHxZ0ZqWyysqxrB5pEBSYw4AAA...

To further demonstrate field-level encryption in action, this blog post includes a sample serverless application that you can deploy by using a CloudFormation template, which creates an application environment using CloudFront, Amazon API Gateway, and Lambda. The sample application is only intended to demonstrate field-level encryption functionality and is not intended for production use. The following diagram depicts the architecture and data flow of this sample application.

Sample application architecture and data flow

Diagram of the solution's architecture and data flow

Here is how the sample solution works:

  1. An application user submits an HTML form page with sensitive data, generating an HTTPS POST to CloudFront.
  2. Field-level encryption intercepts the form POST and encrypts sensitive data with the public RSA key and replaces fields in the form post with encrypted ciphertext. The form POST ciphertext is then sent to origin servers.
  3. The serverless application accepts the form post data containing ciphertext where sensitive data would normally be. If a malicious user were able to compromise your application and gain access to your data, such as the contents of a form, that user would see encrypted data.
  4. Lambda stores data in a DynamoDB table, leaving sensitive data to remain safely encrypted at rest.
  5. An administrator uses the AWS Management Console and a Lambda function to view the sensitive data.
  6. During the session, the administrator retrieves ciphertext from the DynamoDB table.
  7. The administrator decrypts sensitive data by using private key material stored in the EC2 Systems Manager Parameter Store.
  8. Decrypted sensitive data is transmitted over SSL/TLS via the AWS Management Console to the administrator for review.

Deployment walkthrough

The high-level steps to deploy this solution are as follows:

  1. Stage the required artifacts
    When deployment packages are used with Lambda, the zipped artifacts have to be placed in an S3 bucket in the target AWS Region for deployment. This step is not required if you are deploying in the US East (N. Virginia) Region because the package has already been staged there.
  2. Generate an RSA key pair
    Create a public/private key pair that will be used to perform the encrypt/decrypt functionality.
  3. Upload the public key to CloudFront and associate it with the field-level encryption configuration
    After you create the key pair, the public key is uploaded to CloudFront so that it can be used by field-level encryption.
  4. Launch the CloudFormation stack
    Deploy the sample application for demonstrating field-level encryption by using AWS CloudFormation.
  5. Add the field-level encryption configuration to the CloudFront distribution
    After you have provisioned the application, this step associates the field-level encryption configuration with the CloudFront distribution.
  6. Store the RSA private key in the Parameter Store
    Store the private key in the Parameter Store as a SecureString data type, which uses AWS KMS to encrypt the parameter value.

Deploy the solution

1. Stage the required artifacts

(If you are deploying in the US East [N. Virginia] Region, skip to Step 2, “Generate an RSA key pair.”)

Stage the Lambda function deployment package in an Amazon S3 bucket located in the AWS Region you are using for this solution. To do this, download the zipped deployment package and upload it to your in-region bucket. For additional information about uploading objects to S3, see Uploading Object into Amazon S3.

2. Generate an RSA key pair

In this section, you will generate an RSA key pair by using OpenSSL:

  1. Confirm access to OpenSSL.
    $ openssl version

    You should see version information similar to the following.

    OpenSSL <version> <date>

  1. Create a private key using the following command.
    $ openssl genrsa -out private_key.pem 2048

    The command results should look similar to the following.

    Generating RSA private key, 2048 bit long modulus
    ................................................................................+++
    ..........................+++
    e is 65537 (0x10001)
  1. Extract the public key from the private key by running the following command.
    $ openssl rsa -pubout -in private_key.pem -out public_key.pem

    You should see output similar to the following.

    writing RSA key
  1. Restrict access to the private key.$ chmod 600 private_key.pem Note: You will use the public and private key material in Steps 3 and 6 to configure the sample application.

3. Upload the public key to CloudFront and associate it with the field-level encryption configuration

Now that you have created the RSA key pair, you will use the AWS Management Console to upload the public key to CloudFront for use by field-level encryption. Complete the following steps to upload and configure the public key.

Note: Do not include spaces or special characters when providing the configuration values in this section.

  1. From the AWS Management Console, choose Services > CloudFront.
  2. In the navigation pane, choose Public Key and choose Add Public Key.
    Screenshot of adding a public key

Complete the Add Public Key configuration boxes:

  • Key Name: Type a name such as DemoPublicKey.
  • Encoded Key: Paste the contents of the public_key.pem file you created in Step 2c. Copy and paste the encoded key value for your public key, including the -----BEGIN PUBLIC KEY----- and -----END PUBLIC KEY----- lines.
  • Comment: Optionally add a comment.
  1. Choose Create.
  2. After adding at least one public key to CloudFront, the next step is to create a profile to tell CloudFront which fields of input you want to be encrypted. While still on the CloudFront console, choose Field-level encryption in the navigation pane.
  3. Under Profiles, choose Create profile.
    Screenshot of creating a profile

Complete the Create profile configuration boxes:

  • Name: Type a name such as FLEDemo.
  • Comment: Optionally add a comment.
  • Public key: Select the public key you configured in Step 4.b.
  • Provider name: Type a provider name such as FLEDemo.
    This information will be used when the form data is encrypted, and must be provided to applications that need to decrypt the data, along with the appropriate private key.
  • Pattern to match: Type phone. This configures field-level encryption to match based on the phone.
  1. Choose Save profile.
  2. Configurations include options for whether to block or forward a query to your origin in scenarios where CloudFront can’t encrypt the data. Under Encryption Configurations, choose Create configuration.
    Screenshot of creating a configuration

Complete the Create configuration boxes:

  • Comment: Optionally add a comment.
  • Content type: Enter application/x-www-form-urlencoded. This is a common media type for encoding form data.
  • Default profile ID: Select the profile you added in Step 3e.
  1. Choose Save configuration

4. Launch the CloudFormation stack

Launch the sample application by using a CloudFormation template that automates the provisioning process.

Input parameter Input parameter description
ProviderID Enter the Provider name you assigned in Step 3e. The ProviderID is used in field-level encryption configuration in CloudFront (letters and numbers only, no special characters)
PublicKeyName Enter the Key Name you assigned in Step 3b. This name is assigned to the public key in field-level encryption configuration in CloudFront (letters and numbers only, no special characters).
PrivateKeySSMPath Leave as the default: /cloudfront/field-encryption-sample/private-key
ArtifactsBucket The S3 bucket with artifact files (staged zip file with app code). Leave as default if deploying in us-east-1.
ArtifactsPrefix The path in the S3 bucket containing artifact files. Leave as default if deploying in us-east-1.

To finish creating the CloudFormation stack:

  1. Choose Next on the Select Template page, enter the input parameters and choose Next.
    Note: The Artifacts configuration needs to be updated only if you are deploying outside of us-east-1 (US East [N. Virginia]). See Step 1 for artifact staging instructions.
  2. On the Options page, accept the defaults and choose Next.
  3. On the Review page, confirm the details, choose the I acknowledge that AWS CloudFormation might create IAM resources check box, and then choose Create. (The stack will be created in approximately 15 minutes.)

5. Add the field-level encryption configuration to the CloudFront distribution

While still on the CloudFront console, choose Distributions in the navigation pane, and then:

    1. In the Outputs section of the FLE-Sample-App stack, look for CloudFrontDistribution and click the URL to open the CloudFront console.
    2. Choose Behaviors, choose the Default (*) behavior, and then choose Edit.
    3. For Field-level Encryption Config, choose the configuration you created in Step 3g.
      Screenshot of editing the default cache behavior
    4. Choose Yes, Edit.
    5. While still in the CloudFront distribution configuration, choose the General Choose Edit, scroll down to Distribution State, and change it to Enabled.
    6. Choose Yes, Edit.

6. Store the RSA private key in the Parameter Store

In this step, you store the private key in the EC2 Systems Manager Parameter Store as a SecureString data type, which uses AWS KMS to encrypt the parameter value. For more information about AWS KMS, see the AWS Key Management Service Developer Guide. You will need a working installation of the AWS CLI to complete this step.

  1. Store the private key in the Parameter Store with the AWS CLI by running the following command. You will find the <KMSKeyID> in the KMSKeyID in the CloudFormation stack Outputs. Substitute it for the placeholder in the following command.
    $ aws ssm put-parameter --type "SecureString" --name /cloudfront/field-encryption-sample/private-key --value file://private_key.pem --key-id "<KMSKeyID>"
    
    ------------------
    |  PutParameter  |
    +----------+-----+
    |  Version |  1  |
    +----------+-----+

  1. Verify the parameter. Your private key material should be accessible through the ssm get-parameter in the following command in the Value The key material has been truncated in the following output.
    $ aws ssm get-parameter --name /cloudfront/field-encryption-sample/private-key --with-decryption
    
    -----…
    
    ||  Value  |  -----BEGIN RSA PRIVATE KEY-----
    MIIEowIBAAKCAQEAwGRBGuhacmw+C73kM6Z…….

    Notice we use the —with decryption argument in this command. This returns the private key as cleartext.

    This completes the sample application deployment. Next, we show you how to see field-level encryption in action.

  1. Delete the private key from local storage. On Linux for example, using the shred command, securely delete the private key material from your workstation as shown below. You may also wish to store the private key material within an AWS CloudHSM or other protected location suitable for your security requirements. For production implementations, you also should implement key rotation policies.
    $ shred -zvu -n  100 private*.pem
    
    shred: private_encrypted_key.pem: pass 1/101 (random)...
    shred: private_encrypted_key.pem: pass 2/101 (dddddd)...
    shred: private_encrypted_key.pem: pass 3/101 (555555)...
    ….

Test the sample application

Use the following steps to test the sample application with field-level encryption:

  1. Open sample application in your web browser by clicking the ApplicationURL link in the CloudFormation stack Outputs. (for example, https:d199xe5izz82ea.cloudfront.net/prod/). Note that it may take several minutes for the CloudFront distribution to reach the Deployed Status from the previous step, during which time you may not be able to access the sample application.
  2. Fill out and submit the HTML form on the page:
    1. Complete the three form fields: Full Name, Email Address, and Phone Number.
    2. Choose Submit.
      Screenshot of completing the sample application form
      Notice that the application response includes the form values. The phone number returns the following ciphertext encryption using your public key. This ciphertext has been stored in DynamoDB.
      Screenshot of the phone number as ciphertext
  3. Execute the Lambda decryption function to download ciphertext from DynamoDB and decrypt the phone number using the private key:
    1. In the CloudFormation stack Outputs, locate DecryptFunction and click the URL to open the Lambda console.
    2. Configure a test event using the “Hello World” template.
    3. Choose the Test button.
  4. View the encrypted and decrypted phone number data.
    Screenshot of the encrypted and decrypted phone number data

Summary

In this blog post, we showed you how to use CloudFront field-level encryption to encrypt sensitive data at edge locations and help prevent access from unauthorized systems. The source code for this solution is available on GitHub. For additional information about field-level encryption, see the documentation.

If you have comments about this post, submit them in the “Comments” section below. If you have questions about or issues implementing this solution, please start a new thread on the CloudFront forum.

– Alex and Cameron

How to Easily Apply Amazon Cloud Directory Schema Changes with In-Place Schema Upgrades

Post Syndicated from Mahendra Chheda original https://aws.amazon.com/blogs/security/how-to-easily-apply-amazon-cloud-directory-schema-changes-with-in-place-schema-upgrades/

Now, Amazon Cloud Directory makes it easier for you to apply schema changes across your directories with in-place schema upgrades. Your directory now remains available while Cloud Directory applies backward-compatible schema changes such as the addition of new fields. Without migrating data between directories or applying code changes to your applications, you can upgrade your schemas. You also can view the history of your schema changes in Cloud Directory by using version identifiers, which help you track and audit schema versions across directories. If you have multiple instances of a directory with the same schema, you can view the version history of schema changes to manage your directory fleet and ensure that all directories are running with the same schema version.

In this blog post, I demonstrate how to perform an in-place schema upgrade and use schema versions in Cloud Directory. I add additional attributes to an existing facet and add a new facet to a schema. I then publish the new schema and apply it to running directories, upgrading the schema in place. I also show how to view the version history of a directory schema, which helps me to ensure my directory fleet is running the same version of the schema and has the correct history of schema changes applied to it.

Note: I share Java code examples in this post. I assume that you are familiar with the AWS SDK and can use Java-based code to build a Cloud Directory code example. You can apply the concepts I cover in this post to other programming languages such as Python and Ruby.

Cloud Directory fundamentals

I will start by covering a few Cloud Directory fundamentals. If you are already familiar with the concepts behind Cloud Directory facets, schemas, and schema lifecycles, you can skip to the next section.

Facets: Groups of attributes. You use facets to define object types. For example, you can define a device schema by adding facets such as computers, phones, and tablets. A computer facet can track attributes such as serial number, make, and model. You can then use the facets to create computer objects, phone objects, and tablet objects in the directory to which the schema applies.

Schemas: Collections of facets. Schemas define which types of objects can be created in a directory (such as users, devices, and organizations) and enforce validation of data for each object class. All data within a directory must conform to the applied schema. As a result, the schema definition is essentially a blueprint to construct a directory with an applied schema.

Schema lifecycle: The four distinct states of a schema: Development, Published, Applied, and Deleted. Schemas in the Published and Applied states have version identifiers and cannot be changed. Schemas in the Applied state are used by directories for validation as applications insert or update data. You can change schemas in the Development state as many times as you need them to. In-place schema upgrades allow you to apply schema changes to an existing Applied schema in a production directory without the need to export and import the data populated in the directory.

How to add attributes to a computer inventory application schema and perform an in-place schema upgrade

To demonstrate how to set up schema versioning and perform an in-place schema upgrade, I will use an example of a computer inventory application that uses Cloud Directory to store relationship data. Let’s say that at my company, AnyCompany, we use this computer inventory application to track all computers we give to our employees for work use. I previously created a ComputerSchema and assigned its version identifier as 1. This schema contains one facet called ComputerInfo that includes attributes for SerialNumber, Make, and Model, as shown in the following schema details.

Schema: ComputerSchema
Version: 1

Facet: ComputerInfo
Attribute: SerialNumber, type: Integer
Attribute: Make, type: String
Attribute: Model, type: String

AnyCompany has offices in Seattle, Portland, and San Francisco. I have deployed the computer inventory application for each of these three locations. As shown in the lower left part of the following diagram, ComputerSchema is in the Published state with a version of 1. The Published schema is applied to SeattleDirectory, PortlandDirectory, and SanFranciscoDirectory for AnyCompany’s three locations. Implementing separate directories for different geographic locations when you don’t have any queries that cross location boundaries is a good data partitioning strategy and gives your application better response times with lower latency.

Diagram of ComputerSchema in Published state and applied to three directories

Legend for the diagrams in this post

The following code example creates the schema in the Development state by using a JSON file, publishes the schema, and then creates directories for the Seattle, Portland, and San Francisco locations. For this example, I assume the schema has been defined in the JSON file. The createSchema API creates a schema Amazon Resource Name (ARN) with the name defined in the variable, SCHEMA_NAME. I can use the putSchemaFromJson API to add specific schema definitions from the JSON file.

// The utility method to get valid Cloud Directory schema JSON
String validJson = getJsonFile("ComputerSchema_version_1.json")

String SCHEMA_NAME = "ComputerSchema";

String developmentSchemaArn = client.createSchema(new CreateSchemaRequest()
        .withName(SCHEMA_NAME))
        .getSchemaArn();

// Put the schema document in the Development schema
PutSchemaFromJsonResult result = client.putSchemaFromJson(new PutSchemaFromJsonRequest()
        .withSchemaArn(developmentSchemaArn)
        .withDocument(validJson));

The following code example takes the schema that is currently in the Development state and publishes the schema, changing its state to Published.

String SCHEMA_VERSION = "1";
String publishedSchemaArn = client.publishSchema(
        new PublishSchemaRequest()
        .withDevelopmentSchemaArn(developmentSchemaArn)
        .withVersion(SCHEMA_VERSION))
        .getPublishedSchemaArn();

// Our Published schema ARN is as follows
// arn:aws:clouddirectory:us-west-2:XXXXXXXXXXXX:schema/published/ComputerSchema/1

The following code example creates a directory named SeattleDirectory and applies the published schema. The createDirectory API call creates a directory by using the published schema provided in the API parameters. Note that Cloud Directory stores a version of the schema in the directory in the Applied state. I will use similar code to create directories for PortlandDirectory and SanFranciscoDirectory.

String DIRECTORY_NAME = "SeattleDirectory"; 

CreateDirectoryResult directory = client.createDirectory(
        new CreateDirectoryRequest()
        .withName(DIRECTORY_NAME)
        .withSchemaArn(publishedSchemaArn));

String directoryArn = directory.getDirectoryArn();
String appliedSchemaArn = directory.getAppliedSchemaArn();

// This code section can be reused to create directories for Portland and San Francisco locations with the appropriate directory names

// Our directory ARN is as follows 
// arn:aws:clouddirectory:us-west-2:XXXXXXXXXXXX:directory/XX_DIRECTORY_GUID_XX

// Our applied schema ARN is as follows 
// arn:aws:clouddirectory:us-west-2:XXXXXXXXXXXX:directory/XX_DIRECTORY_GUID_XX/schema/ComputerSchema/1

Revising a schema

Now let’s say my company, AnyCompany, wants to add more information for computers and to track which employees have been assigned a computer for work use. I modify the schema to add two attributes to the ComputerInfo facet: Description and OSVersion (operating system version). I make Description optional because it is not important for me to track this attribute for the computer objects I create. I make OSVersion mandatory because it is critical for me to track it for all computer objects so that I can make changes such as applying security patches or making upgrades. Because I make OSVersion mandatory, I must provide a default value that Cloud Directory will apply to objects that were created before the schema revision, in order to handle backward compatibility. Note that you can replace the value in any object with a different value.

I also add a new facet to track computer assignment information, shown in the following updated schema as the ComputerAssignment facet. This facet tracks these additional attributes: Name (the name of the person to whom the computer is assigned), EMail (the email address of the assignee), Department, and department CostCenter. Note that Cloud Directory refers to the previously available version identifier as the Major Version. Because I can now add a minor version to a schema, I also denote the changed schema as Minor Version A.

Schema: ComputerSchema
Major Version: 1
Minor Version: A 

Facet: ComputerInfo
Attribute: SerialNumber, type: Integer 
Attribute: Make, type: String
Attribute: Model, type: Integer
Attribute: Description, type: String, required: NOT_REQUIRED
Attribute: OSVersion, type: String, required: REQUIRED_ALWAYS, default: "Windows 7"

Facet: ComputerAssignment
Attribute: Name, type: String
Attribute: EMail, type: String
Attribute: Department, type: String
Attribute: CostCenter, type: Integer

The following diagram shows the changes that were made when I added another facet to the schema and attributes to the existing facet. The highlighted area of the diagram (bottom left) shows that the schema changes were published.

Diagram showing that schema changes were published

The following code example revises the existing Development schema by adding the new attributes to the ComputerInfo facet and by adding the ComputerAssignment facet. I use a new JSON file for the schema revision, and for the purposes of this example, I am assuming the JSON file has the full schema including planned revisions.

// The utility method to get a valid CloudDirectory schema JSON
String schemaJson = getJsonFile("ComputerSchema_version_1_A.json")

// Put the schema document in the Development schema
PutSchemaFromJsonResult result = client.putSchemaFromJson(
        new PutSchemaFromJsonRequest()
        .withSchemaArn(developmentSchemaArn)
        .withDocument(schemaJson));

Upgrading the Published schema

The following code example performs an in-place schema upgrade of the Published schema with schema revisions (it adds new attributes to the existing facet and another facet to the schema). The upgradePublishedSchema API upgrades the Published schema with backward-compatible changes from the Development schema.

// From an earlier code example, I know the publishedSchemaArn has this value: "arn:aws:clouddirectory:us-west-2:XXXXXXXXXXXX:schema/published/ComputerSchema/1"

// Upgrade publishedSchemaArn to minorVersion A. The Development schema must be backward compatible with 
// the existing publishedSchemaArn. 

String minorVersion = "A"

UpgradePublishedSchemaResult upgradePublishedSchemaResult = client.upgradePublishedSchema(new UpgradePublishedSchemaRequest()
        .withDevelopmentSchemaArn(developmentSchemaArn)
        .withPublishedSchemaArn(publishedSchemaArn)
        .withMinorVersion(minorVersion));

String upgradedPublishedSchemaArn = upgradePublishedSchemaResult.getUpgradedSchemaArn();

// The Published schema ARN after the upgrade shows a minor version as follows 
// arn:aws:clouddirectory:us-west-2:XXXXXXXXXXXX:schema/published/ComputerSchema/1/A

Upgrading the Applied schema

The following diagram shows the in-place schema upgrade for the SeattleDirectory directory. I am performing the schema upgrade so that I can reflect the new schemas in all three directories. As a reminder, I added new attributes to the ComputerInfo facet and also added the ComputerAssignment facet. After the schema and directory upgrade, I can create objects for the ComputerInfo and ComputerAssignment facets in the SeattleDirectory. Any objects that were created with the old facet definition for ComputerInfo will now use the default values for any additional attributes defined in the new schema.

Diagram of the in-place schema upgrade for the SeattleDirectory directory

I use the following code example to perform an in-place upgrade of the SeattleDirectory to a Major Version of 1 and a Minor Version of A. Note that you should change a Major Version identifier in a schema to make backward-incompatible changes such as changing the data type of an existing attribute or dropping a mandatory attribute from your schema. Backward-incompatible changes require directory data migration from a previous version to the new version. You should change a Minor Version identifier in a schema to make backward-compatible upgrades such as adding additional attributes or adding facets, which in turn may contain one or more attributes. The upgradeAppliedSchema API lets me upgrade an existing directory with a different version of a schema.

// This upgrades ComputerSchema version 1 of the Applied schema in SeattleDirectory to Major Version 1 and Minor Version A
// The schema must be backward compatible or the API will fail with IncompatibleSchemaException

UpgradeAppliedSchemaResult upgradeAppliedSchemaResult = client.upgradeAppliedSchema(new UpgradeAppliedSchemaRequest()
        .withDirectoryArn(directoryArn)
        .withPublishedSchemaArn(upgradedPublishedSchemaArn));

String upgradedAppliedSchemaArn = upgradeAppliedSchemaResult.getUpgradedSchemaArn();

// The Applied schema ARN after the in-place schema upgrade will appear as follows
// arn:aws:clouddirectory:us-west-2:XXXXXXXXXXXX:directory/XX_DIRECTORY_GUID_XX/schema/ComputerSchema/1

// This code section can be reused to upgrade directories for the Portland and San Francisco locations with the appropriate directory ARN

Note: Cloud Directory has excluded returning the Minor Version identifier in the Applied schema ARN for backward compatibility and to enable the application to work across older and newer versions of the directory.

The following diagram shows the changes that are made when I perform an in-place schema upgrade in the two remaining directories, PortlandDirectory and SanFranciscoDirectory. I make these calls sequentially, upgrading PortlandDirectory first and then upgrading SanFranciscoDirectory. I use the same code example that I used earlier to upgrade SeattleDirectory. Now, all my directories are running the most current version of the schema. Also, I made these schema changes without having to migrate data and while maintaining my application’s high availability.

Diagram showing the changes that are made with an in-place schema upgrade in the two remaining directories

Schema revision history

I can now view the schema revision history for any of AnyCompany’s directories by using the listAppliedSchemaArns API. Cloud Directory maintains the five most recent versions of applied schema changes. Similarly, to inspect the current Minor Version that was applied to my schema, I use the getAppliedSchemaVersion API. The listAppliedSchemaArns API returns the schema ARNs based on my schema filter as defined in withSchemaArn.

I use the following code example to query an Applied schema for its version history.

// This returns the five most recent Minor Versions associated with a Major Version
ListAppliedSchemaArnsResult listAppliedSchemaArnsResult = client.listAppliedSchemaArns(new ListAppliedSchemaArnsRequest()
        .withDirectoryArn(directoryArn)
        .withSchemaArn(upgradedAppliedSchemaArn));

// Note: The listAppliedSchemaArns API without the SchemaArn filter returns all the Major Versions in a directory

The listAppliedSchemaArns API returns the two ARNs as shown in the following output.

arn:aws:clouddirectory:us-west-2:XXXXXXXXXXXX:directory/XX_DIRECTORY_GUID_XX/schema/ComputerSchema/1
arn:aws:clouddirectory:us-west-2:XXXXXXXXXXXX:directory/XX_DIRECTORY_GUID_XX/schema/ComputerSchema/1/A

The following code example queries an Applied schema for current Minor Version by using the getAppliedSchemaVersion API.

// This returns the current Applied schema's Minor Version ARN 

GetAppliedSchemaVersion getAppliedSchemaVersionResult = client.getAppliedSchemaVersion(new GetAppliedSchemaVersionRequest()
	.withSchemaArn(upgradedAppliedSchemaArn));

The getAppliedSchemaVersion API returns the current Applied schema ARN with a Minor Version, as shown in the following output.

arn:aws:clouddirectory:us-west-2:XXXXXXXXXXXX:directory/XX_DIRECTORY_GUID_XX/schema/ComputerSchema/1/A

If you have a lot of directories, schema revision API calls can help you audit your directory fleet and ensure that all directories are running the same version of a schema. Such auditing can help you ensure high integrity of directories across your fleet.

Summary

You can use in-place schema upgrades to make changes to your directory schema as you evolve your data set to match the needs of your application. An in-place schema upgrade allows you to maintain high availability for your directory and applications while the upgrade takes place. For more information about in-place schema upgrades, see the in-place schema upgrade documentation.

If you have comments about this blog post, submit them in the “Comments” section below. If you have questions about implementing the solution in this post, start a new thread in the Directory Service forum or contact AWS Support.

– Mahendra

 

Glenn’s Take on re:Invent 2017 – Part 3

Post Syndicated from Glenn Gore original https://aws.amazon.com/blogs/architecture/glenns-take-on-reinvent-2017-part-3/

Glenn Gore here, Chief Architect for AWS. I was in Las Vegas last week — with 43K others — for re:Invent 2017. I checked in to the Architecture blog here and here with my take on what was interesting about some of the bigger announcements from a cloud-architecture perspective.

In the excitement of so many new services being launched, we sometimes overlook feature updates that, while perhaps not as exciting as Amazon DeepLens, have significant impact on how you architect and develop solutions on AWS.

Amazon DynamoDB is used by more than 100,000 customers around the world, handling over a trillion requests every day. From the start, DynamoDB has offered high availability by natively spanning multiple Availability Zones within an AWS Region. As more customers started building and deploying truly-global applications, there was a need to replicate a DynamoDB table to multiple AWS Regions, allowing for read/write operations to occur in any region where the table was replicated. This update is important for providing a globally-consistent view of information — as users may transition from one region to another — or for providing additional levels of availability, allowing for failover between AWS Regions without loss of information.

There are some interesting concurrency-design aspects you need to be aware of and ensure you can handle correctly. For example, we support the “last writer wins” reconciliation where eventual consistency is being used and an application updates the same item in different AWS Regions at the same time. If you require strongly-consistent read/writes then you must perform all of your read/writes in the same AWS Region. The details behind this can be found in the DynamoDB documentation. Providing a globally-distributed, replicated DynamoDB table simplifies many different use cases and allows for the logic of replication, which may have been pushed up into the application layers to be simplified back down into the data layer.

The other big update for DynamoDB is that you can now back up your DynamoDB table on demand with no impact to performance. One of the features I really like is that when you trigger a backup, it is available instantly, regardless of the size of the table. Behind the scenes, we use snapshots and change logs to ensure a consistent backup. While backup is instant, restoring the table could take some time depending on its size and ranges — from minutes to hours for very large tables.

This feature is super important for those of you who work in regulated industries that often have strict requirements around data retention and backups of data, which sometimes limited the use of DynamoDB or required complex workarounds to implement some sort of backup feature in the past. This often incurred significant, additional costs due to increased read transactions on their DynamoDB tables.

Amazon Simple Storage Service (Amazon S3) was our first-released AWS service over 11 years ago, and it proved the simplicity and scalability of true API-driven architectures in the cloud. Today, Amazon S3 stores trillions of objects, with transactional requests per second reaching into the millions! Dealing with data as objects opened up an incredibly diverse array of use cases ranging from libraries of static images, game binary downloads, and application log data, to massive data lakes used for big data analytics and business intelligence. With Amazon S3, when you accessed your data in an object, you effectively had to write/read the object as a whole or use the range feature to retrieve a part of the object — if possible — in your individual use case.

Now, with Amazon S3 Select, an SQL-like query language is used that can work with delimited text and JSON files, as well as work with GZIP compressed files. We don’t support encryption during the preview of Amazon S3 Select.

Amazon S3 Select provides two major benefits:

  • Faster access
  • Lower running costs

Serverless Lambda functions, where every millisecond matters when you are being charged, will benefit greatly from Amazon S3 Select as data retrieval and processing of your Lambda function will experience significant speedups and cost reductions. For example, we have seen 2x speed improvement and 80% cost reduction with the Serverless MapReduce code.

Other AWS services such as Amazon Athena, Amazon Redshift, and Amazon EMR will support Amazon S3 Select as well as partner offerings including Cloudera and Hortonworks. If you are using Amazon Glacier for longer-term data archival, you will be able to use Amazon Glacier Select to retrieve a subset of your content from within Amazon Glacier.

As the volume of data that can be stored within Amazon S3 and Amazon Glacier continues to scale on a daily basis, we will continue to innovate and develop improved and optimized services that will allow you to work with these magnificently-large data sets while reducing your costs (retrieval and processing). I believe this will also allow you to simplify the transformation and storage of incoming data into Amazon S3 in basic, semi-structured formats as a single copy vs. some of the duplication and reformatting of data sometimes required to do upfront optimizations for downstream processing. Amazon S3 Select largely removes the need for this upfront optimization and instead allows you to store data once and process it based on your individual Amazon S3 Select query per application or transaction need.

Thanks for reading!

Glenn contemplating why CSV format is still relevant in 2017 (Italy).

Glenn’s Take on re:Invent 2017 Part 1

Post Syndicated from Glenn Gore original https://aws.amazon.com/blogs/architecture/glenns-take-on-reinvent-2017-part-1/

GREETINGS FROM LAS VEGAS

Glenn Gore here, Chief Architect for AWS. I’m in Las Vegas this week — with 43K others — for re:Invent 2017. We have a lot of exciting announcements this week. I’m going to post to the AWS Architecture blog each day with my take on what’s interesting about some of the announcements from a cloud architectural perspective.

Why not start at the beginning? At the Midnight Madness launch on Sunday night, we announced Amazon Sumerian, our platform for VR, AR, and mixed reality. The hype around VR/AR has existed for many years, though for me, it is a perfect example of how a working end-to-end solution often requires innovation from multiple sources. For AR/VR to be successful, we need many components to come together in a coherent manner to provide a great experience.

First, we need lightweight, high-definition goggles with motion tracking that are comfortable to wear. Second, we need to track movement of our body and hands in a 3-D space so that we can interact with virtual objects in the virtual world. Third, we need to build the virtual world itself and populate it with assets and define how the interactions will work and connect with various other systems.

There has been rapid development of the physical devices for AR/VR, ranging from iOS devices to Oculus Rift and HTC Vive, which provide excellent capabilities for the first and second components defined above. With the launch of Amazon Sumerian we are solving for the third area, which will help developers easily build their own virtual worlds and start experimenting and innovating with how to apply AR/VR in new ways.

Already, within 48 hours of Amazon Sumerian being announced, I have had multiple discussions with customers and partners around some cool use cases where VR can help in training simulations, remote-operator controls, or with new ideas around interacting with complex visual data sets, which starts bringing concepts straight out of sci-fi movies into the real (virtual) world. I am really excited to see how Sumerian will unlock the creative potential of developers and where this will lead.

Amazon MQ
I am a huge fan of distributed architectures where asynchronous messaging is the backbone of connecting the discrete components together. Amazon Simple Queue Service (Amazon SQS) is one of my favorite services due to its simplicity, scalability, performance, and the incredible flexibility of how you can use Amazon SQS in so many different ways to solve complex queuing scenarios.

While Amazon SQS is easy to use when building cloud-native applications on AWS, many of our customers running existing applications on-premises required support for different messaging protocols such as: Java Message Service (JMS), .Net Messaging Service (NMS), Advanced Message Queuing Protocol (AMQP), MQ Telemetry Transport (MQTT), Simple (or Streaming) Text Orientated Messaging Protocol (STOMP), and WebSockets. One of the most popular applications for on-premise message brokers is Apache ActiveMQ. With the release of Amazon MQ, you can now run Apache ActiveMQ on AWS as a managed service similar to what we did with Amazon ElastiCache back in 2012. For me, there are two compelling, major benefits that Amazon MQ provides:

  • Integrate existing applications with cloud-native applications without having to change a line of application code if using one of the supported messaging protocols. This removes one of the biggest blockers for integration between the old and the new.
  • Remove the complexity of configuring Multi-AZ resilient message broker services as Amazon MQ provides out-of-the-box redundancy by always storing messages redundantly across Availability Zones. Protection is provided against failure of a broker through to complete failure of an Availability Zone.

I believe that Amazon MQ is a major component in the tools required to help you migrate your existing applications to AWS. Having set up cross-data center Apache ActiveMQ clusters in the past myself and then testing to ensure they work as expected during critical failure scenarios, technical staff working on migrations to AWS benefit from the ease of deploying a fully redundant, managed Apache ActiveMQ cluster within minutes.

Who would have thought I would have been so excited to revisit Apache ActiveMQ in 2017 after using SQS for many, many years? Choice is a wonderful thing.

Amazon GuardDuty
Maintaining application and information security in the modern world is increasingly complex and is constantly evolving and changing as new threats emerge. This is due to the scale, variety, and distribution of services required in a competitive online world.

At Amazon, security is our number one priority. Thus, we are always looking at how we can increase security detection and protection while simplifying the implementation of advanced security practices for our customers. As a result, we released Amazon GuardDuty, which provides intelligent threat detection by using a combination of multiple information sources, transactional telemetry, and the application of machine learning models developed by AWS. One of the biggest benefits of Amazon GuardDuty that I appreciate is that enabling this service requires zero software, agents, sensors, or network choke points. which can all impact performance or reliability of the service you are trying to protect. Amazon GuardDuty works by monitoring your VPC flow logs, AWS CloudTrail events, DNS logs, as well as combing other sources of security threats that AWS is aggregating from our own internal and external sources.

The use of machine learning in Amazon GuardDuty allows it to identify changes in behavior, which could be suspicious and require additional investigation. Amazon GuardDuty works across all of your AWS accounts allowing for an aggregated analysis and ensuring centralized management of detected threats across accounts. This is important for our larger customers who can be running many hundreds of AWS accounts across their organization, as providing a single common threat detection of their organizational use of AWS is critical to ensuring they are protecting themselves.

Detection, though, is only the beginning of what Amazon GuardDuty enables. When a threat is identified in Amazon GuardDuty, you can configure remediation scripts or trigger Lambda functions where you have custom responses that enable you to start building automated responses to a variety of different common threats. Speed of response is required when a security incident may be taking place. For example, Amazon GuardDuty detects that an Amazon Elastic Compute Cloud (Amazon EC2) instance might be compromised due to traffic from a known set of malicious IP addresses. Upon detection of a compromised EC2 instance, we could apply an access control entry restricting outbound traffic for that instance, which stops loss of data until a security engineer can assess what has occurred.

Whether you are a customer running a single service in a single account, or a global customer with hundreds of accounts with thousands of applications, or a startup with hundreds of micro-services with hourly release cycle in a devops world, I recommend enabling Amazon GuardDuty. We have a 30-day free trial available for all new customers of this service. As it is a monitor of events, there is no change required to your architecture within AWS.

Stay tuned for tomorrow’s post on AWS Media Services and Amazon Neptune.

 

Glenn during the Tour du Mont Blanc

Object models

Post Syndicated from Eevee original https://eev.ee/blog/2017/11/28/object-models/

Anonymous asks, with dollars:

More about programming languages!

Well then!

I’ve written before about what I think objects are: state and behavior, which in practice mostly means method calls.

I suspect that the popular impression of what objects are, and also how they should work, comes from whatever C++ and Java happen to do. From that point of view, the whole post above is probably nonsense. If the baseline notion of “object” is a rigid definition woven tightly into the design of two massively popular languages, then it doesn’t even make sense to talk about what “object” should mean — it does mean the features of those languages, and cannot possibly mean anything else.

I think that’s a shame! It piles a lot of baggage onto a fairly simple idea. Polymorphism, for example, has nothing to do with objects — it’s an escape hatch for static type systems. Inheritance isn’t the only way to reuse code between objects, but it’s the easiest and fastest one, so it’s what we get. Frankly, it’s much closer to a speed tradeoff than a fundamental part of the concept.

We could do with more experimentation around how objects work, but that’s impossible in the languages most commonly thought of as object-oriented.

Here, then, is a (very) brief run through the inner workings of objects in four very dynamic languages. I don’t think I really appreciated objects until I’d spent some time with Python, and I hope this can help someone else whet their own appetite.

Python 3

Of the four languages I’m going to touch on, Python will look the most familiar to the Java and C++ crowd. For starters, it actually has a class construct.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
class Vector:
    def __init__(self, x, y):
        self.x = x
        self.y = y

    def __neg__(self):
        return Vector(-self.x, -self.y)

    def __div__(self, denom):
        return Vector(self.x / denom, self.y / denom)

    @property
    def magnitude(self):
        return (self.x ** 2 + self.y ** 2) ** 0.5

    def normalized(self):
        return self / self.magnitude

The __init__ method is an initializer, which is like a constructor but named differently (because the object already exists in a usable form by the time the initializer is called). Operator overloading is done by implementing methods with other special __dunder__ names. Properties can be created with @property, where the @ is syntax for applying a wrapper function to a function as it’s defined. You can do inheritance, even multiply:

1
2
3
4
class Foo(A, B, C):
    def bar(self, x, y, z):
        # do some stuff
        super().bar(x, y, z)

Cool, a very traditional object model.

Except… for some details.

Some details

For one, Python objects don’t have a fixed layout. Code both inside and outside the class can add or remove whatever attributes they want from whatever object they want. The underlying storage is just a dict, Python’s mapping type. (Or, rather, something like one. Also, it’s possible to change, which will probably be the case for everything I say here.)

If you create some attributes at the class level, you’ll start to get a peek behind the curtains:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
class Foo:
    values = []

    def add_value(self, value):
        self.values.append(value)

a = Foo()
b = Foo()
a.add_value('a')
print(a.values)  # ['a']
b.add_value('b')
print(b.values)  # ['a', 'b']

The [] assigned to values isn’t a default assigned to each object. In fact, the individual objects don’t know about it at all! You can use vars(a) to get at the underlying storage dict, and you won’t see a values entry in there anywhere.

Instead, values lives on the class, which is a value (and thus an object) in its own right. When Python is asked for self.values, it checks to see if self has a values attribute; in this case, it doesn’t, so Python keeps going and asks the class for one.

Python’s object model is secretly prototypical — a class acts as a prototype, as a shared set of fallback values, for its objects.

In fact, this is also how method calls work! They aren’t syntactically special at all, which you can see by separating the attribute lookup from the call.

1
2
3
print("abc".startswith("a"))  # True
meth = "abc".startswith
print(meth("a"))  # True

Reading obj.method looks for a method attribute; if there isn’t one on obj, Python checks the class. Here, it finds one: it’s a function from the class body.

Ah, but wait! In the code I just showed, meth seems to “know” the object it came from, so it can’t just be a plain function. If you inspect the resulting value, it claims to be a “bound method” or “built-in method” rather than a function, too. Something funny is going on here, and that funny something is the descriptor protocol.

Descriptors

Python allows attributes to implement their own custom behavior when read from or written to. Such an attribute is called a descriptor. I’ve written about them before, but here’s a quick overview.

If Python looks up an attribute, finds it in a class, and the value it gets has a __get__ method… then instead of using that value, Python will use the return value of its __get__ method.

The @property decorator works this way. The magnitude property in my original example was shorthand for doing this:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
class MagnitudeDescriptor:
    def __get__(self, instance, owner):
        if instance is None:
            return self
        return (instance.x ** 2 + instance.y ** 2) ** 0.5

class Vector:
    def __init__(self, x, y):
        self.x = x
        self.y = y

    magnitude = MagnitudeDescriptor()

When you ask for somevec.magnitude, Python checks somevec but doesn’t find magnitude, so it consults the class instead. The class does have a magnitude, and it’s a value with a __get__ method, so Python calls that method and somevec.magnitude evaluates to its return value. (The instance is None check is because __get__ is called even if you get the descriptor directly from the class via Vector.magnitude. A descriptor intended to work on instances can’t do anything useful in that case, so the convention is to return the descriptor itself.)

You can also intercept attempts to write to or delete an attribute, and do absolutely whatever you want instead. But note that, similar to operating overloading in Python, the descriptor must be on a class; you can’t just slap one on an arbitrary object and have it work.

This brings me right around to how “bound methods” actually work. Functions are descriptors! The function type implements __get__, and when a function is retrieved from a class via an instance, that __get__ bundles the function and the instance together into a tiny bound method object. It’s essentially:

1
2
3
4
5
class FunctionType:
    def __get__(self, instance, owner):
        if instance is None:
            return self
        return functools.partial(self, instance)

The self passed as the first argument to methods is not special or magical in any way. It’s built out of a few simple pieces that are also readily accessible to Python code.

Note also that because obj.method() is just an attribute lookup and a call, Python doesn’t actually care whether method is a method on the class or just some callable thing on the object. You won’t get the auto-self behavior if it’s on the object, but otherwise there’s no difference.

More attribute access, and the interesting part

Descriptors are one of several ways to customize attribute access. Classes can implement __getattr__ to intervene when an attribute isn’t found on an object; __setattr__ and __delattr__ to intervene when any attribute is set or deleted; and __getattribute__ to implement unconditional attribute access. (That last one is a fantastic way to create accidental recursion, since any attribute access you do within __getattribute__ will of course call __getattribute__ again.)

Here’s what I really love about Python. It might seem like a magical special case that descriptors only work on classes, but it really isn’t. You could implement exactly the same behavior yourself, in pure Python, using only the things I’ve just told you about. Classes are themselves objects, remember, and they are instances of type, so the reason descriptors only work on classes is that type effectively does this:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
class type:
    def __getattribute__(self, name):
        value = super().__getattribute__(name)
        # like all op overloads, __get__ must be on the type, not the instance
        ty = type(value)
        if hasattr(ty, '__get__'):
            # it's a descriptor!  this is a class access so there is no instance
            return ty.__get__(value, None, self)
        else:
            return value

You can even trivially prove to yourself that this is what’s going on by skipping over types behavior:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
class Descriptor:
    def __get__(self, instance, owner):
        print('called!')

class Foo:
    bar = Descriptor()

Foo.bar  # called!
type.__getattribute__(Foo, 'bar')  # called!
object.__getattribute__(Foo, 'bar')  # ...

And that’s not all! The mysterious super function, used to exhaustively traverse superclass method calls even in the face of diamond inheritance, can also be expressed in pure Python using these primitives. You could write your own superclass calling convention and use it exactly the same way as super.

This is one of the things I really like about Python. Very little of it is truly magical; virtually everything about the object model exists in the types rather than the language, which means virtually everything can be customized in pure Python.

Class creation and metaclasses

A very brief word on all of this stuff, since I could talk forever about Python and I have three other languages to get to.

The class block itself is fairly interesting. It looks like this:

1
2
class Name(*bases, **kwargs):
    # code

I’ve said several times that classes are objects, and in fact the class block is one big pile of syntactic sugar for calling type(...) with some arguments to create a new type object.

The Python documentation has a remarkably detailed description of this process, but the gist is:

  • Python determines the type of the new class — the metaclass — by looking for a metaclass keyword argument. If there isn’t one, Python uses the “lowest” type among the provided base classes. (If you’re not doing anything special, that’ll just be type, since every class inherits from object and object is an instance of type.)

  • Python executes the class body. It gets its own local scope, and any assignments or method definitions go into that scope.

  • Python now calls type(name, bases, attrs, **kwargs). The name is whatever was right after class; the bases are position arguments; and attrs is the class body’s local scope. (This is how methods and other class attributes end up on the class.) The brand new type is then assigned to Name.

Of course, you can mess with most of this. You can implement __prepare__ on a metaclass, for example, to use a custom mapping as storage for the local scope — including any reads, which allows for some interesting shenanigans. The only part you can’t really implement in pure Python is the scoping bit, which has a couple extra rules that make sense for classes. (In particular, functions defined within a class block don’t close over the class body; that would be nonsense.)

Object creation

Finally, there’s what actually happens when you create an object — including a class, which remember is just an invocation of type(...).

Calling Foo(...) is implemented as, well, a call. Any type can implement calls with the __call__ special method, and you’ll find that type itself does so. It looks something like this:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
# oh, a fun wrinkle that's hard to express in pure python: type is a class, so
# it's an instance of itself
class type:
    def __call__(self, *args, **kwargs):
        # remember, here 'self' is a CLASS, an instance of type.
        # __new__ is a true constructor: object.__new__ allocates storage
        # for a new blank object
        instance = self.__new__(self, *args, **kwargs)
        # you can return whatever you want from __new__ (!), and __init__
        # is only called on it if it's of the right type
        if isinstance(instance, self):
            instance.__init__(*args, **kwargs)
        return instance

Again, you can trivially confirm this by asking any type for its __call__ method. Assuming that type doesn’t implement __call__ itself, you’ll get back a bound version of types implementation.

1
2
>>> list.__call__
<method-wrapper '__call__' of type object at 0x7fafb831a400>

You can thus implement __call__ in your own metaclass to completely change how subclasses are created — including skipping the creation altogether, if you like.

And… there’s a bunch of stuff I haven’t even touched on.

The Python philosophy

Python offers something that, on the surface, looks like a “traditional” class/object model. Under the hood, it acts more like a prototypical system, where failed attribute lookups simply defer to a superclass or metaclass.

The language also goes to almost superhuman lengths to expose all of its moving parts. Even the prototypical behavior is an implementation of __getattribute__ somewhere, which you are free to completely replace in your own types. Proxying and delegation are easy.

Also very nice is that these features “bundle” well, by which I mean a library author can do all manner of convoluted hijinks, and a consumer of that library doesn’t have to see any of it or understand how it works. You only need to inherit from a particular class (which has a metaclass), or use some descriptor as a decorator, or even learn any new syntax.

This meshes well with Python culture, which is pretty big on the principle of least surprise. These super-advanced features tend to be tightly confined to single simple features (like “makes a weak attribute“) or cordoned with DSLs (e.g., defining a form/struct/database table with a class body). In particular, I’ve never seen a metaclass in the wild implement its own __call__.

I have mixed feelings about that. It’s probably a good thing overall that the Python world shows such restraint, but I wonder if there are some very interesting possibilities we’re missing out on. I implemented a metaclass __call__ myself, just once, in an entity/component system that strove to minimize fuss when communicating between components. It never saw the light of day, but I enjoyed seeing some new things Python could do with the same relatively simple syntax. I wouldn’t mind seeing, say, an object model based on composition (with no inheritance) built atop Python’s primitives.

Lua

Lua doesn’t have an object model. Instead, it gives you a handful of very small primitives for building your own object model. This is pretty typical of Lua — it’s a very powerful language, but has been carefully constructed to be very small at the same time. I’ve never encountered anything else quite like it, and “but it starts indexing at 1!” really doesn’t do it justice.

The best way to demonstrate how objects work in Lua is to build some from scratch. We need two key features. The first is metatables, which bear a passing resemblance to Python’s metaclasses.

Tables and metatables

The table is Lua’s mapping type and its primary data structure. Keys can be any value other than nil. Lists are implemented as tables whose keys are consecutive integers starting from 1. Nothing terribly surprising. The dot operator is sugar for indexing with a string key.

1
2
3
4
5
local t = { a = 1, b = 2 }
print(t['a'])  -- 1
print(t.b)  -- 2
t.c = 3
print(t['c'])  -- 3

A metatable is a table that can be associated with another value (usually another table) to change its behavior. For example, operator overloading is implemented by assigning a function to a special key in a metatable.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
local t = { a = 1, b = 2 }
--print(t + 0)  -- error: attempt to perform arithmetic on a table value

local mt = {
    __add = function(left, right)
        return 12
    end,
}
setmetatable(t, mt)
print(t + 0)  -- 12

Now, the interesting part: one of the special keys is __index, which is consulted when the base table is indexed by a key it doesn’t contain. Here’s a table that claims every key maps to itself.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
local t = {}
local mt = {
    __index = function(table, key)
        return key
    end,
}
setmetatable(t, mt)
print(t.foo)  -- foo
print(t.bar)  -- bar
print(t[3])  -- 3

__index doesn’t have to be a function, either. It can be yet another table, in which case that table is simply indexed with the key. If the key still doesn’t exist and that table has a metatable with an __index, the process repeats.

With this, it’s easy to have several unrelated tables that act as a single table. Call the base table an object, fill the __index table with functions and call it a class, and you have half of an object system. You can even get prototypical inheritance by chaining __indexes together.

At this point things are a little confusing, since we have at least three tables going on, so here’s a diagram. Keep in mind that Lua doesn’t actually have anything called an “object”, “class”, or “method” — those are just convenient nicknames for a particular structure we might build with Lua’s primitives.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
                    ╔═══════════╗        ...
                    ║ metatable ║         ║
                    ╟───────────╢   ┌─────╨───────────────────────┐
                    ║ __index   ╫───┤ lookup table ("superclass") │
                    ╚═══╦═══════╝   ├─────────────────────────────┤
  ╔═══════════╗         ║           │ some other method           ┼─── function() ... end
  ║ metatable ║         ║           └─────────────────────────────┘
  ╟───────────╢   ┌─────╨──────────────────┐
  ║ __index   ╫───┤ lookup table ("class") │
  ╚═══╦═══════╝   ├────────────────────────┤
      ║           │ some method            ┼─── function() ... end
      ║           └────────────────────────┘
┌─────╨─────────────────┐
│ base table ("object") │
└───────────────────────┘

Note that a metatable is not the same as a class; it defines behavior, not methods. Conversely, if you try to use a class directly as a metatable, it will probably not do much. (This is pretty different from e.g. Python, where operator overloads are just methods with funny names. One nice thing about the Lua approach is that you can keep interface-like functionality separate from methods, and avoid clogging up arbitrary objects’ namespaces. You could even use a dummy table as a key and completely avoid name collisions.)

Anyway, code!

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
local class = {
    foo = function(a)
        print("foo got", a)
    end,
}
local mt = { __index = class }
-- setmetatable returns its first argument, so this is nice shorthand
local obj1 = setmetatable({}, mt)
local obj2 = setmetatable({}, mt)
obj1.foo(7)  -- foo got 7
obj2.foo(9)  -- foo got 9

Wait, wait, hang on. Didn’t I call these methods? How do they get at the object? Maybe Lua has a magical this variable?

Methods, sort of

Not quite, but this is where the other key feature comes in: method-call syntax. It’s the lightest touch of sugar, just enough to have method invocation.

1
2
3
4
5
6
7
8
9
-- note the colon!
a:b(c, d, ...)

-- exactly equivalent to this
-- (except that `a` is only evaluated once)
a.b(a, c, d, ...)

-- which of course is really this
a["b"](a, c, d, ...)

Now we can write methods that actually do something.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
local class = {
    bar = function(self)
        print("our score is", self.score)
    end,
}
local mt = { __index = class }
local obj1 = setmetatable({ score = 13 }, mt)
local obj2 = setmetatable({ score = 25 }, mt)
obj1:bar()  -- our score is 13
obj2:bar()  -- our score is 25

And that’s all you need. Much like Python, methods and data live in the same namespace, and Lua doesn’t care whether obj:method() finds a function on obj or gets one from the metatable’s __index. Unlike Python, the function will be passed self either way, because self comes from the use of : rather than from the lookup behavior.

(Aside: strictly speaking, any Lua value can have a metatable — and if you try to index a non-table, Lua will always consult the metatable’s __index. Strings all have the string library as a metatable, so you can call methods on them: try ("%s %s"):format(1, 2). I don’t think Lua lets user code set the metatable for non-tables, so this isn’t that interesting, but if you’re writing Lua bindings from C then you can wrap your pointers in metatables to give them methods implemented in C.)

Bringing it all together

Of course, writing all this stuff every time is a little tedious and error-prone, so instead you might want to wrap it all up inside a little function. No problem.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
local function make_object(body)
    -- create a metatable
    local mt = { __index = body }
    -- create a base table to serve as the object itself
    local obj = setmetatable({}, mt)
    -- and, done
    return obj
end

-- you can leave off parens if you're only passing in 
local Dog = {
    -- this acts as a "default" value; if obj.barks is missing, __index will
    -- kick in and find this value on the class.  but if obj.barks is assigned
    -- to, it'll go in the object and shadow the value here.
    barks = 0,

    bark = function(self)
        self.barks = self.barks + 1
        print("woof!")
    end,
}

local mydog = make_object(Dog)
mydog:bark()  -- woof!
mydog:bark()  -- woof!
mydog:bark()  -- woof!
print(mydog.barks)  -- 3
print(Dog.barks)  -- 0

It works, but it’s fairly barebones. The nice thing is that you can extend it pretty much however you want. I won’t reproduce an entire serious object system here — lord knows there are enough of them floating around — but the implementation I have for my LÖVE games lets me do this:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
local Animal = Object:extend{
    cries = 0,
}

-- called automatically by Object
function Animal:init()
    print("whoops i couldn't think of anything interesting to put here")
end

-- this is just nice syntax for adding a first argument called 'self', then
-- assigning this function to Animal.cry
function Animal:cry()
    self.cries = self.cries + 1
end

local Cat = Animal:extend{}

function Cat:cry()
    print("meow!")
    Cat.__super.cry(self)
end

local cat = Cat()
cat:cry()  -- meow!
cat:cry()  -- meow!
print(cat.cries)  -- 2

When I say you can extend it however you want, I mean that. I could’ve implemented Python (2)-style super(Cat, self):cry() syntax; I just never got around to it. I could even make it work with multiple inheritance if I really wanted to — or I could go the complete opposite direction and only implement composition. I could implement descriptors, customizing the behavior of individual table keys. I could add pretty decent syntax for composition/proxying. I am trying very hard to end this section now.

The Lua philosophy

Lua’s philosophy is to… not have a philosophy? It gives you the bare minimum to make objects work, and you can do absolutely whatever you want from there. Lua does have something resembling prototypical inheritance, but it’s not so much a first-class feature as an emergent property of some very simple tools. And since you can make __index be a function, you could avoid the prototypical behavior and do something different entirely.

The very severe downside, of course, is that you have to find or build your own object system — which can get pretty confusing very quickly, what with the multiple small moving parts. Third-party code may also have its own object system with subtly different behavior. (Though, in my experience, third-party code tries very hard to avoid needing an object system at all.)

It’s hard to say what the Lua “culture” is like, since Lua is an embedded language that’s often a little different in each environment. I imagine it has a thousand millicultures, instead. I can say that the tedium of building my own object model has led me into something very “traditional”, with prototypical inheritance and whatnot. It’s partly what I’m used to, but it’s also just really dang easy to get working.

Likewise, while I love properties in Python and use them all the dang time, I’ve yet to use a single one in Lua. They wouldn’t be particularly hard to add to my object model, but having to add them myself (or shop around for an object model with them and also port all my code to use it) adds a huge amount of friction. I’ve thought about designing an interesting ECS with custom object behavior, too, but… is it really worth the effort? For all the power and flexibility Lua offers, the cost is that by the time I have something working at all, I’m too exhausted to actually use any of it.

JavaScript

JavaScript is notable for being preposterously heavily used, yet not having a class block.

Well. Okay. Yes. It has one now. It didn’t for a very long time, and even the one it has now is sugar.

Here’s a vector class again:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
class Vector {
    constructor(x, y) {
        this.x = x;
        this.y = y;
    }

    get magnitude() {
        return Math.sqrt(this.x * this.x + this.y * this.y);
    }

    dot(other) {
        return this.x * other.x + this.y * other.y;
    }
}

In “classic” JavaScript, this would be written as:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
function Vector(x, y) {
    this.x = x;
    this.y = y;
}

Object.defineProperty(Vector.prototype, 'magnitude', {
    configurable: true,
    enumerable: true,
    get: function() {
        return Math.sqrt(this.x * this.x + this.y * this.y);
    },
});


Vector.prototype.dot = function(other) {
    return this.x * other.x + this.y * other.y;
};

Hm, yes. I can see why they added class.

The JavaScript model

In JavaScript, a new type is defined in terms of a function, which is its constructor.

Right away we get into trouble here. There is a very big difference between these two invocations, which I actually completely forgot about just now after spending four hours writing about Python and Lua:

1
2
let vec = Vector(3, 4);
let vec = new Vector(3, 4);

The first calls the function Vector. It assigns some properties to this, which here is going to be window, so now you have a global x and y. It then returns nothing, so vec is undefined.

The second calls Vector with this set to a new empty object, then evaluates to that object. The result is what you’d actually expect.

(You can detect this situation with the strange new.target expression, but I have never once remembered to do so.)

From here, we have true, honest-to-god, first-class prototypical inheritance. The word “prototype” is even right there. When you write this:

1
vec.dot(vec2)

JavaScript will look for dot on vec and (presumably) not find it. It then consults vecs prototype, an object you can see for yourself by using Object.getPrototypeOf(). Since vec is a Vector, its prototype is Vector.prototype.

I stress that Vector.prototype is not the prototype for Vector. It’s the prototype for instances of Vector.

(I say “instance”, but the true type of vec here is still just object. If you want to find Vector, it’s automatically assigned to the constructor property of its own prototype, so it’s available as vec.constructor.)

Of course, Vector.prototype can itself have a prototype, in which case the process would continue if dot were not found. A common (and, arguably, very bad) way to simulate single inheritance is to set Class.prototype to an instance of a superclass to get the prototype right, then tack on the methods for Class. Nowadays we can do Object.create(Superclass.prototype).

Now that I’ve been through Python and Lua, though, this isn’t particularly surprising. I kinda spoiled it.

I suppose one difference in JavaScript is that you can tack arbitrary attributes directly onto Vector all you like, and they will remain invisible to instances since they aren’t in the prototype chain. This is kind of backwards from Lua, where you can squirrel stuff away in the metatable.

Another difference is that every single object in JavaScript has a bunch of properties already tacked on — the ones in Object.prototype. Every object (and by “object” I mean any mapping) has a prototype, and that prototype defaults to Object.prototype, and it has a bunch of ancient junk like isPrototypeOf.

(Nit: it’s possible to explicitly create an object with no prototype via Object.create(null).)

Like Lua, and unlike Python, JavaScript doesn’t distinguish between keys found on an object and keys found via a prototype. Properties can be defined on prototypes with Object.defineProperty(), but that works just as well directly on an object, too. JavaScript doesn’t have a lot of operator overloading, but some things like Symbol.iterator also work on both objects and prototypes.

About this

You may, at this point, be wondering what this is. Unlike Lua and Python (and the last language below), this is a special built-in value — a context value, invisibly passed for every function call.

It’s determined by where the function came from. If the function was the result of an attribute lookup, then this is set to the object containing that attribute. Otherwise, this is set to the global object, window. (You can also set this to whatever you want via the call method on functions.)

This decision is made lexically, i.e. from the literal source code as written. There are no Python-style bound methods. In other words:

1
2
3
4
5
// this = obj
obj.method()
// this = window
let meth = obj.method
meth()

Also, because this is reassigned on every function call, it cannot be meaningfully closed over, which makes using closures within methods incredibly annoying. The old approach was to assign this to some other regular name like self (which got syntax highlighting since it’s also a built-in name in browsers); then we got Function.bind, which produced a callable thing with a fixed context value, which was kind of nice; and now finally we have arrow functions, which explicitly close over the current this when they’re defined and don’t change it when called. Phew.

Class syntax

I already showed class syntax, and it’s really just one big macro for doing all the prototype stuff The Right Way. It even prevents you from calling the type without new. The underlying model is exactly the same, and you can inspect all the parts.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
class Vector { ... }

console.log(Vector.prototype);  // { dot: ..., magnitude: ..., ... }
let vec = new Vector(3, 4);
console.log(Object.getPrototypeOf(vec));  // same as Vector.prototype

// i don't know why you would subclass vector but let's roll with it
class Vectest extends Vector { ... }

console.log(Vectest.prototype);  // { ... }
console.log(Object.getPrototypeOf(Vectest.prototype))  // same as Vector.prototype

Alas, class syntax has a couple shortcomings. You can’t use the class block to assign arbitrary data to either the type object or the prototype — apparently it was deemed too confusing that mutations would be shared among instances. Which… is… how prototypes work. How Python works. How JavaScript itself, one of the most popular languages of all time, has worked for twenty-two years. Argh.

You can still do whatever assignment you want outside of the class block, of course. It’s just a little ugly, and not something I’d think to look for with a sugary class.

A more subtle result of this behavior is that a class block isn’t quite the same syntax as an object literal. The check for data isn’t a runtime thing; class Foo { x: 3 } fails to parse. So JavaScript now has two largely but not entirely identical styles of key/value block.

Attribute access

Here’s where things start to come apart at the seams, just a little bit.

JavaScript doesn’t really have an attribute protocol. Instead, it has two… extension points, I suppose.

One is Object.defineProperty, seen above. For common cases, there’s also the get syntax inside a property literal, which does the same thing. But unlike Python’s @property, these aren’t wrappers around some simple primitives; they are the primitives. JavaScript is the only language of these four to have “property that runs code on access” as a completely separate first-class concept.

If you want to intercept arbitrary attribute access (and some kinds of operators), there’s a completely different primitive: the Proxy type. It doesn’t let you intercept attribute access or operators; instead, it produces a wrapper object that supports interception and defers to the wrapped object by default.

It’s cool to see composition used in this way, but also, extremely weird. If you want to make your own type that overloads in or calling, you have to return a Proxy that wraps your own type, rather than actually returning your own type. And (unlike the other three languages in this post) you can’t return a different type from a constructor, so you have to throw that away and produce objects only from a factory. And instanceof would be broken, but you can at least fix that with Symbol.hasInstance — which is really operator overloading, implement yet another completely different way.

I know the design here is a result of legacy and speed — if any object could intercept all attribute access, then all attribute access would be slowed down everywhere. Fair enough. It still leaves the surface area of the language a bit… bumpy?

The JavaScript philosophy

It’s a little hard to tell. The original idea of prototypes was interesting, but it was hidden behind some very awkward syntax. Since then, we’ve gotten a bunch of extra features awkwardly bolted on to reflect the wildly varied things the built-in types and DOM API were already doing. We have class syntax, but it’s been explicitly designed to avoid exposing the prototype parts of the model.

I admit I don’t do a lot of heavy JavaScript, so I might just be overlooking it, but I’ve seen virtually no code that makes use of any of the recent advances in object capabilities. Forget about custom iterators or overloading call; I can’t remember seeing any JavaScript in the wild that even uses properties yet. I don’t know if everyone’s waiting for sufficient browser support, nobody knows about them, or nobody cares.

The model has advanced recently, but I suspect JavaScript is still shackled to its legacy of “something about prototypes, I don’t really get it, just copy the other code that’s there” as an object model. Alas! Prototypes are so good. Hopefully class syntax will make it a bit more accessible, as it has in Python.

Perl 5

Perl 5 also doesn’t have an object system and expects you to build your own. But where Lua gives you two simple, powerful tools for building one, Perl 5 feels more like a puzzle with half the pieces missing. Clearly they were going for something, but they only gave you half of it.

In brief, a Perl object is a reference that has been blessed with a package.

I need to explain a few things. Honestly, one of the biggest problems with the original Perl object setup was how many strange corners and unique jargon you had to understand just to get off the ground.

(If you want to try running any of this code, you should stick a use v5.26; as the first line. Perl is very big on backwards compatibility, so you need to opt into breaking changes, and even the mundane say builtin is behind a feature gate.)

References

A reference in Perl is sort of like a pointer, but its main use is very different. See, Perl has the strange property that its data structures try very hard to spill their contents all over the place. Despite having dedicated syntax for arrays — @foo is an array variable, distinct from the single scalar variable $foo — it’s actually impossible to nest arrays.

1
2
3
my @foo = (1, 2, 3, 4);
my @bar = (@foo, @foo);
# @bar is now a flat list of eight items: 1, 2, 3, 4, 1, 2, 3, 4

The idea, I guess, is that an array is not one thing. It’s not a container, which happens to hold multiple things; it is multiple things. Anywhere that expects a single value, such as an array element, cannot contain an array, because an array fundamentally is not a single value.

And so we have “references”, which are a form of indirection, but also have the nice property that they’re single values. They add containment around arrays, and in general they make working with most of Perl’s primitive types much more sensible. A reference to a variable can be taken with the \ operator, or you can use [ ... ] and { ... } to directly create references to anonymous arrays or hashes.

1
2
3
my @foo = (1, 2, 3, 4);
my @bar = (\@foo, \@foo);
# @bar is now a nested list of two items: [1, 2, 3, 4], [1, 2, 3, 4]

(Incidentally, this is the sole reason I initially abandoned Perl for Python. Non-trivial software kinda requires nesting a lot of data structures, so you end up with references everywhere, and the syntax for going back and forth between a reference and its contents is tedious and ugly.)

A Perl object must be a reference. Perl doesn’t care what kind of reference — it’s usually a hash reference, since hashes are a convenient place to store arbitrary properties, but it could just as well be a reference to an array, a scalar, or even a sub (i.e. function) or filehandle.

I’m getting a little ahead of myself. First, the other half: blessing and packages.

Packages and blessing

Perl packages are just namespaces. A package looks like this:

1
2
3
4
5
6
7
package Foo::Bar;

sub quux {
    say "hi from quux!";
}

# now Foo::Bar::quux() can be called from anywhere

Nothing shocking, right? It’s just a named container. A lot of the details are kind of weird, like how a package exists in some liminal quasi-value space, but the basic idea is a Bag Of Stuff.

The final piece is “blessing,” which is Perl’s funny name for binding a package to a reference. A very basic class might look like this:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
package Vector;

# the name 'new' is convention, not special
sub new {
    # perl argument passing is weird, don't ask
    my ($class, $x, $y) = @_;

    # create the object itself -- here, unusually, an array reference makes sense
    my $self = [ $x, $y ];

    # associate the package with that reference
    # note that $class here is just the regular string, 'Vector'
    bless $self, $class;

    return $self;
}

sub x {
    my ($self) = @_;
    return $self->[0];
}

sub y {
    my ($self) = @_;
    return $self->[1];
}

sub magnitude {
    my ($self) = @_;
    return sqrt($self->x ** 2 + $self->y ** 2);
}

# switch back to the "default" package
package main;

# -> is method call syntax, which passes the invocant as the first argument;
# for a package, that's just the package name
my $vec = Vector->new(3, 4);
say $vec->magnitude;  # 5

A few things of note here. First, $self->[0] has nothing to do with objects; it’s normal syntax for getting the value of a index 0 out of an array reference called $self. (Most classes are based on hashrefs and would use $self->{value} instead.) A blessed reference is still a reference and can be treated like one.

In general, -> is Perl’s dereferencey operator, but its exact behavior depends on what follows. If it’s followed by brackets, then it’ll apply the brackets to the thing in the reference: ->{} to index a hash reference, ->[] to index an array reference, and ->() to call a function reference.

But if -> is followed by an identifier, then it’s a method call. For packages, that means calling a function in the package and passing the package name as the first argument. For objects — blessed references — that means calling a function in the associated package and passing the object as the first argument.

This is a little weird! A blessed reference is a superposition of two things: its normal reference behavior, and some completely orthogonal object behavior. Also, object behavior has no notion of methods vs data; it only knows about methods. Perl lets you omit parentheses in a lot of places, including when calling a method with no arguments, so $vec->magnitude is really $vec->magnitude().

Perl’s blessing bears some similarities to Lua’s metatables, but ultimately Perl is much closer to Ruby’s “message passing” approach than the above three languages’ approaches of “get me something and maybe it’ll be callable”. (But this is no surprise — Ruby is a spiritual successor to Perl 5.)

All of this leads to one little wrinkle: how do you actually expose data? Above, I had to write x and y methods. Am I supposed to do that for every single attribute on my type?

Yes! But don’t worry, there are third-party modules to help with this incredibly fundamental task. Take Class::Accessor::Fast, so named because it’s faster than Class::Accessor:

1
2
3
package Foo;
use base qw(Class::Accessor::Fast);
__PACKAGE__->mk_accessors(qw(fred wilma barney));

(__PACKAGE__ is the lexical name of the current package; qw(...) is a list literal that splits its contents on whitespace.)

This assumes you’re using a hashref with keys of the same names as the attributes. $obj->fred will return the fred key from your hashref, and $obj->fred(4) will change it to 4.

You also, somewhat bizarrely, have to inherit from Class::Accessor::Fast. Speaking of which,

Inheritance

Inheritance is done by populating the package-global @ISA array with some number of (string) names of parent packages. Most code instead opts to write use base ...;, which does the same thing. Or, more commonly, use parent ...;, which… also… does the same thing.

Every package implicitly inherits from UNIVERSAL, which can be freely modified by Perl code.

A method can call its superclass method with the SUPER:: pseudo-package:

1
2
3
4
sub foo {
    my ($self) = @_;
    $self->SUPER::foo;
}

However, this does a depth-first search, which means it almost certainly does the wrong thing when faced with multiple inheritance. For a while the accepted solution involved a third-party module, but Perl eventually grew an alternative you have to opt into: C3, which may be more familiar to you as the order Python uses.

1
2
3
4
5
6
use mro 'c3';

sub foo {
    my ($self) = @_;
    $self->next::method;
}

Offhand, I’m not actually sure how next::method works, seeing as it was originally implemented in pure Perl code. I suspect it involves peeking at the caller’s stack frame. If so, then this is a very different style of customizability from e.g. Python — the MRO was never intended to be pluggable, and the use of a special pseudo-package means it isn’t really, but someone was determined enough to make it happen anyway.

Operator overloading and whatnot

Operator overloading looks a little weird, though really it’s pretty standard Perl.

1
2
3
4
5
6
7
8
package MyClass;

use overload '+' => \&_add;

sub _add {
    my ($self, $other, $swap) = @_;
    ...
}

use overload here is a pragma, where “pragma” means “regular-ass module that does some wizardry when imported”.

\&_add is how you get a reference to the _add sub so you can pass it to the overload module. If you just said &_add or _add, that would call it.

And that’s it; you just pass a map of operators to functions to this built-in module. No worry about name clashes or pollution, which is pretty nice. You don’t even have to give references to functions that live in the package, if you don’t want them to clog your namespace; you could put them in another package, or even inline them anonymously.

One especially interesting thing is that Perl lets you overload every operator. Perl has a lot of operators. It considers some math builtins like sqrt and trig functions to be operators, or at least operator-y enough that you can overload them. You can also overload the “file text” operators, such as -e $path to test whether a file exists. You can overload conversions, including implicit conversion to a regex. And most fascinating to me, you can overload dereferencing — that is, the thing Perl does when you say $hashref->{key} to get at the underlying hash. So a single object could pretend to be references of multiple different types, including a subref to implement callability. Neat.

Somewhat related: you can overload basic operators (indexing, etc.) on basic types (not references!) with the tie function, which is designed completely differently and looks for methods with fixed names. Go figure.

You can intercept calls to nonexistent methods by implementing a function called AUTOLOAD, within which the $AUTOLOAD global will contain the name of the method being called. Originally this feature was, I think, intended for loading binary components or large libraries on-the-fly only when needed, hence the name. Offhand I’m not sure I ever saw it used the way __getattr__ is used in Python.

Is there a way to intercept all method calls? I don’t think so, but it is Perl, so I must be forgetting something.

Actually no one does this any more

Like a decade ago, a council of elder sages sat down and put together a whole whizbang system that covers all of it: Moose.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
package Vector;
use Moose;

has x => (is => 'rw', isa => 'Int');
has y => (is => 'rw', isa => 'Int');

sub magnitude {
    my ($self) = @_;
    return sqrt($self->x ** 2 + $self->y ** 2);
}

Moose has its own way to do pretty much everything, and it’s all built on the same primitives. Moose also adds metaclasses, somehow, despite that the underlying model doesn’t actually support them? I’m not entirely sure how they managed that, but I do remember doing some class introspection with Moose and it was much nicer than the built-in way.

(If you’re wondering, the built-in way begins with looking at the hash called %Vector::. No, that’s not a typo.)

I really cannot stress enough just how much stuff Moose does, but I don’t want to delve into it here since Moose itself is not actually the language model.

The Perl philosophy

I hope you can see what I meant with what I first said about Perl, now. It has multiple inheritance with an MRO, but uses the wrong one by default. It has extensive operator overloading, which looks nothing like how inheritance works, and also some of it uses a totally different mechanism with special method names instead. It only understands methods, not data, leaving you to figure out accessors by hand.

There’s 70% of an object system here with a clear general design it was gunning for, but none of the pieces really look anything like each other. It’s weird, in a distinctly Perl way.

The result is certainly flexible, at least! It’s especially cool that you can use whatever kind of reference you want for storage, though even as I say that, I acknowledge it’s no different from simply subclassing list or something in Python. It feels different in Perl, but maybe only because it looks so different.

I haven’t written much Perl in a long time, so I don’t know what the community is like any more. Moose was already ubiquitous when I left, which you’d think would let me say “the community mostly focuses on the stuff Moose can do” — but even a decade ago, Moose could already do far more than I had ever seen done by hand in Perl. It’s always made a big deal out of roles (read: interfaces), for instance, despite that I’d never seen anyone care about them in Perl before Moose came along. Maybe their presence in Moose has made them more popular? Who knows.

Also, I wrote Perl seriously, but in the intervening years I’ve only encountered people who only ever used Perl for one-offs. Maybe it’ll come as a surprise to a lot of readers that Perl has an object model at all.

End

Well, that was fun! I hope any of that made sense.

Special mention goes to Rust, which doesn’t have an object model you can fiddle with at runtime, but does do things a little differently.

It’s been really interesting thinking about how tiny differences make a huge impact on what people do in practice. Take the choice of storage in Perl versus Python. Perl’s massively common URI class uses a string as the storage, nothing else; I haven’t seen anything like that in Python aside from markupsafe, which is specifically designed as a string type. I would guess this is partly because Perl makes you choose — using a hashref is an obvious default, but you have to make that choice one way or the other. In Python (especially 3), inheriting from object and getting dict-based storage is the obvious thing to do; the ability to use another type isn’t quite so obvious, and doing it “right” involves a tiny bit of extra work.

Or, consider that Lua could have descriptors, but the extra bit of work (especially design work) has been enough of an impediment that I’ve never implemented them. I don’t think the object implementations I’ve looked at have included them, either. Super weird!

In that light, it’s only natural that objects would be so strongly associated with the features Java and C++ attach to them. I think that makes it all the more important to play around! Look at what Moose has done. No, really, you should bear in mind my description of how Perl does stuff and flip through the Moose documentation. It’s amazing what they’ve built.

Using Amazon Redshift Spectrum, Amazon Athena, and AWS Glue with Node.js in Production

Post Syndicated from Rafi Ton original https://aws.amazon.com/blogs/big-data/using-amazon-redshift-spectrum-amazon-athena-and-aws-glue-with-node-js-in-production/

This is a guest post by Rafi Ton, founder and CEO of NUVIAD. NUVIAD is, in their own words, “a mobile marketing platform providing professional marketers, agencies and local businesses state of the art tools to promote their products and services through hyper targeting, big data analytics and advanced machine learning tools.”

At NUVIAD, we’ve been using Amazon Redshift as our main data warehouse solution for more than 3 years.

We store massive amounts of ad transaction data that our users and partners analyze to determine ad campaign strategies. When running real-time bidding (RTB) campaigns in large scale, data freshness is critical so that our users can respond rapidly to changes in campaign performance. We chose Amazon Redshift because of its simplicity, scalability, performance, and ability to load new data in near real time.

Over the past three years, our customer base grew significantly and so did our data. We saw our Amazon Redshift cluster grow from three nodes to 65 nodes. To balance cost and analytics performance, we looked for a way to store large amounts of less-frequently analyzed data at a lower cost. Yet, we still wanted to have the data immediately available for user queries and to meet their expectations for fast performance. We turned to Amazon Redshift Spectrum.

In this post, I explain the reasons why we extended Amazon Redshift with Redshift Spectrum as our modern data warehouse. I cover how our data growth and the need to balance cost and performance led us to adopt Redshift Spectrum. I also share key performance metrics in our environment, and discuss the additional AWS services that provide a scalable and fast environment, with data available for immediate querying by our growing user base.

Amazon Redshift as our foundation

The ability to provide fresh, up-to-the-minute data to our customers and partners was always a main goal with our platform. We saw other solutions provide data that was a few hours old, but this was not good enough for us. We insisted on providing the freshest data possible. For us, that meant loading Amazon Redshift in frequent micro batches and allowing our customers to query Amazon Redshift directly to get results in near real time.

The benefits were immediately evident. Our customers could see how their campaigns performed faster than with other solutions, and react sooner to the ever-changing media supply pricing and availability. They were very happy.

However, this approach required Amazon Redshift to store a lot of data for long periods, and our data grew substantially. In our peak, we maintained a cluster running 65 DC1.large nodes. The impact on our Amazon Redshift cluster was evident, and we saw our CPU utilization grow to 90%.

Why we extended Amazon Redshift to Redshift Spectrum

Redshift Spectrum gives us the ability to run SQL queries using the powerful Amazon Redshift query engine against data stored in Amazon S3, without needing to load the data. With Redshift Spectrum, we store data where we want, at the cost that we want. We have the data available for analytics when our users need it with the performance they expect.

Seamless scalability, high performance, and unlimited concurrency

Scaling Redshift Spectrum is a simple process. First, it allows us to leverage Amazon S3 as the storage engine and get practically unlimited data capacity.

Second, if we need more compute power, we can leverage Redshift Spectrum’s distributed compute engine over thousands of nodes to provide superior performance – perfect for complex queries running against massive amounts of data.

Third, all Redshift Spectrum clusters access the same data catalog so that we don’t have to worry about data migration at all, making scaling effortless and seamless.

Lastly, since Redshift Spectrum distributes queries across potentially thousands of nodes, they are not affected by other queries, providing much more stable performance and unlimited concurrency.

Keeping it SQL

Redshift Spectrum uses the same query engine as Amazon Redshift. This means that we did not need to change our BI tools or query syntax, whether we used complex queries across a single table or joins across multiple tables.

An interesting capability introduced recently is the ability to create a view that spans both Amazon Redshift and Redshift Spectrum external tables. With this feature, you can query frequently accessed data in your Amazon Redshift cluster and less-frequently accessed data in Amazon S3, using a single view.

Leveraging Parquet for higher performance

Parquet is a columnar data format that provides superior performance and allows Redshift Spectrum (or Amazon Athena) to scan significantly less data. With less I/O, queries run faster and we pay less per query. You can read all about Parquet at https://parquet.apache.org/ or https://en.wikipedia.org/wiki/Apache_Parquet.

Lower cost

From a cost perspective, we pay standard rates for our data in Amazon S3, and only small amounts per query to analyze data with Redshift Spectrum. Using the Parquet format, we can significantly reduce the amount of data scanned. Our costs are now lower, and our users get fast results even for large complex queries.

What we learned about Amazon Redshift vs. Redshift Spectrum performance

When we first started looking at Redshift Spectrum, we wanted to put it to the test. We wanted to know how it would compare to Amazon Redshift, so we looked at two key questions:

  1. What is the performance difference between Amazon Redshift and Redshift Spectrum on simple and complex queries?
  2. Does the data format impact performance?

During the migration phase, we had our dataset stored in Amazon Redshift and S3 as CSV/GZIP and as Parquet file formats. We tested three configurations:

  • Amazon Redshift cluster with 28 DC1.large nodes
  • Redshift Spectrum using CSV/GZIP
  • Redshift Spectrum using Parquet

We performed benchmarks for simple and complex queries on one month’s worth of data. We tested how much time it took to perform the query, and how consistent the results were when running the same query multiple times. The data we used for the tests was already partitioned by date and hour. Properly partitioning the data improves performance significantly and reduces query times.

Simple query

First, we tested a simple query aggregating billing data across a month:

SELECT 
  user_id, 
  count(*) AS impressions, 
  SUM(billing)::decimal /1000000 AS billing 
FROM <table_name> 
WHERE 
  date >= '2017-08-01' AND 
  date <= '2017-08-31'  
GROUP BY 
  user_id;

We ran the same query seven times and measured the response times (red marking the longest time and green the shortest time):

Execution Time (seconds)
  Amazon Redshift Redshift Spectrum
CSV
Redshift Spectrum Parquet
Run #1 39.65 45.11 11.92
Run #2 15.26 43.13 12.05
Run #3 15.27 46.47 13.38
Run #4 21.22 51.02 12.74
Run #5 17.27 43.35 11.76
Run #6 16.67 44.23 13.67
Run #7 25.37 40.39 12.75
Average 21.53  44.82 12.61

For simple queries, Amazon Redshift performed better than Redshift Spectrum, as we thought, because the data is local to Amazon Redshift.

What was surprising was that using Parquet data format in Redshift Spectrum significantly beat ‘traditional’ Amazon Redshift performance. For our queries, using Parquet data format with Redshift Spectrum delivered an average 40% performance gain over traditional Amazon Redshift. Furthermore, Redshift Spectrum showed high consistency in execution time with a smaller difference between the slowest run and the fastest run.

Comparing the amount of data scanned when using CSV/GZIP and Parquet, the difference was also significant:

Data Scanned (GB)
CSV (Gzip) 135.49
Parquet 2.83

Because we pay only for the data scanned by Redshift Spectrum, the cost saving of using Parquet is evident and substantial.

Complex query

Next, we compared the same three configurations with a complex query.

Execution Time (seconds)
  Amazon Redshift Redshift Spectrum CSV Redshift Spectrum Parquet
Run #1 329.80 84.20 42.40
Run #2 167.60 65.30 35.10
Run #3 165.20 62.20 23.90
Run #4 273.90 74.90 55.90
Run #5 167.70 69.00 58.40
Average 220.84 71.12 43.14

This time, Redshift Spectrum using Parquet cut the average query time by 80% compared to traditional Amazon Redshift!

Bottom line: For complex queries, Redshift Spectrum provided a 67% performance gain over Amazon Redshift. Using the Parquet data format, Redshift Spectrum delivered an 80% performance improvement over Amazon Redshift. For us, this was substantial.

Optimizing the data structure for different workloads

Because the cost of S3 is relatively inexpensive and we pay only for the data scanned by each query, we believe that it makes sense to keep our data in different formats for different workloads and different analytics engines. It is important to note that we can have any number of tables pointing to the same data on S3. It all depends on how we partition the data and update the table partitions.

Data permutations

For example, we have a process that runs every minute and generates statistics for the last minute of data collected. With Amazon Redshift, this would be done by running the query on the table with something as follows:

SELECT 
  user, 
  COUNT(*) 
FROM 
  events_table 
WHERE 
  ts BETWEEN ‘2017-08-01 14:00:00’ AND ‘2017-08-01 14:00:59’ 
GROUP BY 
  user;

(Assuming ‘ts’ is your column storing the time stamp for each event.)

With Redshift Spectrum, we pay for the data scanned in each query. If the data is partitioned by the minute instead of the hour, a query looking at one minute would be 1/60th the cost. If we use a temporary table that points only to the data of the last minute, we save that unnecessary cost.

Creating Parquet data efficiently

On the average, we have 800 instances that process our traffic. Each instance sends events that are eventually loaded into Amazon Redshift. When we started three years ago, we would offload data from each server to S3 and then perform a periodic copy command from S3 to Amazon Redshift.

Recently, Amazon Kinesis Firehose added the capability to offload data directly to Amazon Redshift. While this is now a viable option, we kept the same collection process that worked flawlessly and efficiently for three years.

This changed, however, when we incorporated Redshift Spectrum. With Redshift Spectrum, we needed to find a way to:

  • Collect the event data from the instances.
  • Save the data in Parquet format.
  • Partition the data effectively.

To accomplish this, we save the data as CSV and then transform it to Parquet. The most effective method to generate the Parquet files is to:

  1. Send the data in one-minute intervals from the instances to Kinesis Firehose with an S3 temporary bucket as the destination.
  2. Aggregate hourly data and convert it to Parquet using AWS Lambda and AWS Glue.
  3. Add the Parquet data to S3 by updating the table partitions.

With this new process, we had to give more attention to validating the data before we sent it to Kinesis Firehose, because a single corrupted record in a partition fails queries on that partition.

Data validation

To store our click data in a table, we considered the following SQL create table command:

create external TABLE spectrum.blog_clicks (
    user_id varchar(50),
    campaign_id varchar(50),
    os varchar(50),
    ua varchar(255),
    ts bigint,
    billing float
)
partitioned by (date date, hour smallint)  
stored as parquet
location 's3://nuviad-temp/blog/clicks/';

The above statement defines a new external table (all Redshift Spectrum tables are external tables) with a few attributes. We stored ‘ts’ as a Unix time stamp and not as Timestamp, and billing data is stored as float and not decimal (more on that later). We also said that the data is partitioned by date and hour, and then stored as Parquet on S3.

First, we need to get the table definitions. This can be achieved by running the following query:

SELECT 
  * 
FROM 
  svv_external_columns 
WHERE 
  tablename = 'blog_clicks';

This query lists all the columns in the table with their respective definitions:

schemaname tablename columnname external_type columnnum part_key
spectrum blog_clicks user_id varchar(50) 1 0
spectrum blog_clicks campaign_id varchar(50) 2 0
spectrum blog_clicks os varchar(50) 3 0
spectrum blog_clicks ua varchar(255) 4 0
spectrum blog_clicks ts bigint 5 0
spectrum blog_clicks billing double 6 0
spectrum blog_clicks date date 7 1
spectrum blog_clicks hour smallint 8 2

Now we can use this data to create a validation schema for our data:

const rtb_request_schema = {
    "name": "clicks",
    "items": {
        "user_id": {
            "type": "string",
            "max_length": 100
        },
        "campaign_id": {
            "type": "string",
            "max_length": 50
        },
        "os": {
            "type": "string",
            "max_length": 50            
        },
        "ua": {
            "type": "string",
            "max_length": 255            
        },
        "ts": {
            "type": "integer",
            "min_value": 0,
            "max_value": 9999999999999
        },
        "billing": {
            "type": "float",
            "min_value": 0,
            "max_value": 9999999999999
        }
    }
};

Next, we create a function that uses this schema to validate data:

function valueIsValid(value, item_schema) {
    if (schema.type == 'string') {
        return (typeof value == 'string' && value.length <= schema.max_length);
    }
    else if (schema.type == 'integer') {
        return (typeof value == 'number' && value >= schema.min_value && value <= schema.max_value);
    }
    else if (schema.type == 'float' || schema.type == 'double') {
        return (typeof value == 'number' && value >= schema.min_value && value <= schema.max_value);
    }
    else if (schema.type == 'boolean') {
        return typeof value == 'boolean';
    }
    else if (schema.type == 'timestamp') {
        return (new Date(value)).getTime() > 0;
    }
    else {
        return true;
    }
}

Near real-time data loading with Kinesis Firehose

On Kinesis Firehose, we created a new delivery stream to handle the events as follows:

Delivery stream name: events
Source: Direct PUT
S3 bucket: nuviad-events
S3 prefix: rtb/
IAM role: firehose_delivery_role_1
Data transformation: Disabled
Source record backup: Disabled
S3 buffer size (MB): 100
S3 buffer interval (sec): 60
S3 Compression: GZIP
S3 Encryption: No Encryption
Status: ACTIVE
Error logging: Enabled

This delivery stream aggregates event data every minute, or up to 100 MB, and writes the data to an S3 bucket as a CSV/GZIP compressed file. Next, after we have the data validated, we can safely send it to our Kinesis Firehose API:

if (validated) {
    let itemString = item.join('|')+'\n'; //Sending csv delimited by pipe and adding new line

    let params = {
        DeliveryStreamName: 'events',
        Record: {
            Data: itemString
        }
    };

    firehose.putRecord(params, function(err, data) {
        if (err) {
            console.error(err, err.stack);        
        }
        else {
            // Continue to your next step 
        }
    });
}

Now, we have a single CSV file representing one minute of event data stored in S3. The files are named automatically by Kinesis Firehose by adding a UTC time prefix in the format YYYY/MM/DD/HH before writing objects to S3. Because we use the date and hour as partitions, we need to change the file naming and location to fit our Redshift Spectrum schema.

Automating data distribution using AWS Lambda

We created a simple Lambda function triggered by an S3 put event that copies the file to a different location (or locations), while renaming it to fit our data structure and processing flow. As mentioned before, the files generated by Kinesis Firehose are structured in a pre-defined hierarchy, such as:

S3://your-bucket/your-prefix/2017/08/01/20/events-4-2017-08-01-20-06-06-536f5c40-6893-4ee4-907d-81e4d3b09455.gz

All we need to do is parse the object name and restructure it as we see fit. In our case, we did the following (the event is an object received in the Lambda function with all the data about the object written to S3):

/*
	object key structure in the event object:
your-prefix/2017/08/01/20/event-4-2017-08-01-20-06-06-536f5c40-6893-4ee4-907d-81e4d3b09455.gz
	*/

let key_parts = event.Records[0].s3.object.key.split('/'); 

let event_type = key_parts[0];
let date = key_parts[1] + '-' + key_parts[2] + '-' + key_parts[3];
let hour = key_parts[4];
if (hour.indexOf('0') == 0) {
 		hour = parseInt(hour, 10) + '';
}
    
let parts1 = key_parts[5].split('-');
let minute = parts1[7];
if (minute.indexOf('0') == 0) {
        minute = parseInt(minute, 10) + '';
}

Now, we can redistribute the file to the two destinations we need—one for the minute processing task and the other for hourly aggregation:

    copyObjectToHourlyFolder(event, date, hour, minute)
        .then(copyObjectToMinuteFolder.bind(null, event, date, hour, minute))
        .then(addPartitionToSpectrum.bind(null, event, date, hour, minute))
        .then(deleteOldMinuteObjects.bind(null, event))
        .then(deleteStreamObject.bind(null, event))        
        .then(result => {
            callback(null, { message: 'done' });            
        })
        .catch(err => {
            console.error(err);
            callback(null, { message: err });            
        }); 

Kinesis Firehose stores the data in a temporary folder. We copy the object to another folder that holds the data for the last processed minute. This folder is connected to a small Redshift Spectrum table where the data is being processed without needing to scan a much larger dataset. We also copy the data to a folder that holds the data for the entire hour, to be later aggregated and converted to Parquet.

Because we partition the data by date and hour, we created a new partition on the Redshift Spectrum table if the processed minute is the first minute in the hour (that is, minute 0). We ran the following:

ALTER TABLE 
  spectrum.events 
ADD partition
  (date='2017-08-01', hour=0) 
  LOCATION 's3://nuviad-temp/events/2017-08-01/0/';

After the data is processed and added to the table, we delete the processed data from the temporary Kinesis Firehose storage and from the minute storage folder.

Migrating CSV to Parquet using AWS Glue and Amazon EMR

The simplest way we found to run an hourly job converting our CSV data to Parquet is using Lambda and AWS Glue (and thanks to the awesome AWS Big Data team for their help with this).

Creating AWS Glue jobs

What this simple AWS Glue script does:

  • Gets parameters for the job, date, and hour to be processed
  • Creates a Spark EMR context allowing us to run Spark code
  • Reads CSV data into a DataFrame
  • Writes the data as Parquet to the destination S3 bucket
  • Adds or modifies the Redshift Spectrum / Amazon Athena table partition for the table
import sys
import sys
from awsglue.transforms import *
from awsglue.utils import getResolvedOptions
from pyspark.context import SparkContext
from awsglue.context import GlueContext
from awsglue.job import Job
import boto3

## @params: [JOB_NAME]
args = getResolvedOptions(sys.argv, ['JOB_NAME','day_partition_key', 'hour_partition_key', 'day_partition_value', 'hour_partition_value' ])

#day_partition_key = "partition_0"
#hour_partition_key = "partition_1"
#day_partition_value = "2017-08-01"
#hour_partition_value = "0"

day_partition_key = args['day_partition_key']
hour_partition_key = args['hour_partition_key']
day_partition_value = args['day_partition_value']
hour_partition_value = args['hour_partition_value']

print("Running for " + day_partition_value + "/" + hour_partition_value)

sc = SparkContext()
glueContext = GlueContext(sc)
spark = glueContext.spark_session
job = Job(glueContext)
job.init(args['JOB_NAME'], args)

df = spark.read.option("delimiter","|").csv("s3://nuviad-temp/events/"+day_partition_value+"/"+hour_partition_value)
df.registerTempTable("data")

df1 = spark.sql("select _c0 as user_id, _c1 as campaign_id, _c2 as os, _c3 as ua, cast(_c4 as bigint) as ts, cast(_c5 as double) as billing from data")

df1.repartition(1).write.mode("overwrite").parquet("s3://nuviad-temp/parquet/"+day_partition_value+"/hour="+hour_partition_value)

client = boto3.client('athena', region_name='us-east-1')

response = client.start_query_execution(
    QueryString='alter table parquet_events add if not exists partition(' + day_partition_key + '=\'' + day_partition_value + '\',' + hour_partition_key + '=' + hour_partition_value + ')  location \'s3://nuviad-temp/parquet/' + day_partition_value + '/hour=' + hour_partition_value + '\'' ,
    QueryExecutionContext={
        'Database': 'spectrumdb'
    },
    ResultConfiguration={
        'OutputLocation': 's3://nuviad-temp/convertresults'
    }
)

response = client.start_query_execution(
    QueryString='alter table parquet_events partition(' + day_partition_key + '=\'' + day_partition_value + '\',' + hour_partition_key + '=' + hour_partition_value + ') set location \'s3://nuviad-temp/parquet/' + day_partition_value + '/hour=' + hour_partition_value + '\'' ,
    QueryExecutionContext={
        'Database': 'spectrumdb'
    },
    ResultConfiguration={
        'OutputLocation': 's3://nuviad-temp/convertresults'
    }
)

job.commit()

Note: Because Redshift Spectrum and Athena both use the AWS Glue Data Catalog, we could use the Athena client to add the partition to the table.

Here are a few words about float, decimal, and double. Using decimal proved to be more challenging than we expected, as it seems that Redshift Spectrum and Spark use them differently. Whenever we used decimal in Redshift Spectrum and in Spark, we kept getting errors, such as:

S3 Query Exception (Fetch). Task failed due to an internal error. File 'https://s3-external-1.amazonaws.com/nuviad-temp/events/2017-08-01/hour=2/part-00017-48ae5b6b-906e-4875-8cde-bc36c0c6d0ca.c000.snappy.parquet has an incompatible Parquet schema for column 's3://nuviad-events/events.lat'. Column type: DECIMAL(18, 8), Parquet schema:\noptional float lat [i:4 d:1 r:0]\n (https://s3-external-1.amazonaws.com/nuviad-temp/events/2017-08-01/hour=2/part-00017-48ae5b6b-906e-4875-8cde-bc36c0c6d0ca.c000.snappy.parq

We had to experiment with a few floating-point formats until we found that the only combination that worked was to define the column as double in the Spark code and float in Spectrum. This is the reason you see billing defined as float in Spectrum and double in the Spark code.

Creating a Lambda function to trigger conversion

Next, we created a simple Lambda function to trigger the AWS Glue script hourly using a simple Python code:

import boto3
import json
from datetime import datetime, timedelta
 
client = boto3.client('glue')
 
def lambda_handler(event, context):
    last_hour_date_time = datetime.now() - timedelta(hours = 1)
    day_partition_value = last_hour_date_time.strftime("%Y-%m-%d") 
    hour_partition_value = last_hour_date_time.strftime("%-H") 
    response = client.start_job_run(
    JobName='convertEventsParquetHourly',
    Arguments={
         '--day_partition_key': 'date',
         '--hour_partition_key': 'hour',
         '--day_partition_value': day_partition_value,
         '--hour_partition_value': hour_partition_value
         }
    )

Using Amazon CloudWatch Events, we trigger this function hourly. This function triggers an AWS Glue job named ‘convertEventsParquetHourly’ and runs it for the previous hour, passing job names and values of the partitions to process to AWS Glue.

Redshift Spectrum and Node.js

Our development stack is based on Node.js, which is well-suited for high-speed, light servers that need to process a huge number of transactions. However, a few limitations of the Node.js environment required us to create workarounds and use other tools to complete the process.

Node.js and Parquet

The lack of Parquet modules for Node.js required us to implement an AWS Glue/Amazon EMR process to effectively migrate data from CSV to Parquet. We would rather save directly to Parquet, but we couldn’t find an effective way to do it.

One interesting project in the works is the development of a Parquet NPM by Marc Vertes called node-parquet (https://www.npmjs.com/package/node-parquet). It is not in a production state yet, but we think it would be well worth following the progress of this package.

Timestamp data type

According to the Parquet documentation, Timestamp data are stored in Parquet as 64-bit integers. However, JavaScript does not support 64-bit integers, because the native number type is a 64-bit double, giving only 53 bits of integer range.

The result is that you cannot store Timestamp correctly in Parquet using Node.js. The solution is to store Timestamp as string and cast the type to Timestamp in the query. Using this method, we did not witness any performance degradation whatsoever.

Lessons learned

You can benefit from our trial-and-error experience.

Lesson #1: Data validation is critical

As mentioned earlier, a single corrupt entry in a partition can fail queries running against this partition, especially when using Parquet, which is harder to edit than a simple CSV file. Make sure that you validate your data before scanning it with Redshift Spectrum.

Lesson #2: Structure and partition data effectively

One of the biggest benefits of using Redshift Spectrum (or Athena for that matter) is that you don’t need to keep nodes up and running all the time. You pay only for the queries you perform and only for the data scanned per query.

Keeping different permutations of your data for different queries makes a lot of sense in this case. For example, you can partition your data by date and hour to run time-based queries, and also have another set partitioned by user_id and date to run user-based queries. This results in faster and more efficient performance of your data warehouse.

Storing data in the right format

Use Parquet whenever you can. The benefits of Parquet are substantial. Faster performance, less data to scan, and much more efficient columnar format. However, it is not supported out-of-the-box by Kinesis Firehose, so you need to implement your own ETL. AWS Glue is a great option.

Creating small tables for frequent tasks

When we started using Redshift Spectrum, we saw our Amazon Redshift costs jump by hundreds of dollars per day. Then we realized that we were unnecessarily scanning a full day’s worth of data every minute. Take advantage of the ability to define multiple tables on the same S3 bucket or folder, and create temporary and small tables for frequent queries.

Lesson #3: Combine Athena and Redshift Spectrum for optimal performance

Moving to Redshift Spectrum also allowed us to take advantage of Athena as both use the AWS Glue Data Catalog. Run fast and simple queries using Athena while taking advantage of the advanced Amazon Redshift query engine for complex queries using Redshift Spectrum.

Redshift Spectrum excels when running complex queries. It can push many compute-intensive tasks, such as predicate filtering and aggregation, down to the Redshift Spectrum layer, so that queries use much less of your cluster’s processing capacity.

Lesson #4: Sort your Parquet data within the partition

We achieved another performance improvement by sorting data within the partition using sortWithinPartitions(sort_field). For example:

df.repartition(1).sortWithinPartitions("campaign_id")…

Conclusion

We were extremely pleased with using Amazon Redshift as our core data warehouse for over three years. But as our client base and volume of data grew substantially, we extended Amazon Redshift to take advantage of scalability, performance, and cost with Redshift Spectrum.

Redshift Spectrum lets us scale to virtually unlimited storage, scale compute transparently, and deliver super-fast results for our users. With Redshift Spectrum, we store data where we want at the cost we want, and have the data available for analytics when our users need it with the performance they expect.


About the Author

With 7 years of experience in the AdTech industry and 15 years in leading technology companies, Rafi Ton is the founder and CEO of NUVIAD. He enjoys exploring new technologies and putting them to use in cutting edge products and services, in the real world generating real money. Being an experienced entrepreneur, Rafi believes in practical-programming and fast adaptation of new technologies to achieve a significant market advantage.

 

 

AWS Media Services – Process, Store, and Monetize Cloud-Based Video

Post Syndicated from Jeff Barr original https://aws.amazon.com/blogs/aws/aws-media-services-process-store-and-monetize-cloud-based-video/

Do you remember what web video was like in the early days? Standalone players, video no larger than a postage stamp, slow & cantankerous connections, overloaded servers, and the ever-present buffering messages were the norm less than two decades ago.

Today, thanks to technological progress and a broad array of standards, things are a lot better. Video consumers are now in control. They use devices of all shapes, sizes, and vintages to enjoy live and recorded content that is broadcast, streamed, or sent over-the-top (OTT, as they say), and expect immediate access to content that captures and then holds their attention. Meeting these expectations presents a challenge for content creators and distributors. Instead of generating video in a one-size-fits-all format, they (or their media servers) must be prepared to produce video that spans a broad range of sizes, formats, and bit rates, taking care to be ready to deal with planned or unplanned surges in demand. In the face of all of this complexity, they must backstop their content with a monetization model that supports the content and the infrastructure to deliver it.

New AWS Media Services
Today we are launching an array of broadcast-quality media services, each designed to address one or more aspects of the challenge that I outlined above. You can use them together to build a complete end-to-end video solution or you can use one or more in building-block style. In true AWS fashion, you can spend more time innovating and less time setting up and running infrastructure, leaving you ready to focus on creating, delivering, and monetizing your content. The services are all elastic, allowing you to ramp up processing power, connections, and storage and giving you the ability to handle million-user (and beyond) spikes with ease.

Here are the services (all accessible from a set of interactive consoles as well as through a comprehensive set of APIs):

AWS Elemental MediaConvert – File-based transcoding for OTT, broadcast, or archiving, with support for a long list of formats and codecs. Features include multi-channel audio, graphic overlays, closed captioning, and several DRM options.

AWS Elemental MediaLive – Live encoding to deliver video streams in real time to both televisions and multiscreen devices. Allows you to deploy highly reliable live channels in minutes, with full control over encoding parameters. It supports ad insertion, multi-channel audio, graphic overlays, and closed captioning.

AWS Elemental MediaPackage – Video origination and just-in-time packaging. Starting from a single input, produces output for multiple devices representing a long list of current and legacy formats. Supports multiple monetization models, time-shifted live streaming, ad insertion, DRM, and blackout management.

AWS Elemental MediaStore – Media-optimized storage that enables high performance and low latency applications such as live streaming, while taking advantage of the scale and durability of Amazon Simple Storage Service (S3).

AWS Elemental MediaTailor – Monetization service that supports ad serving and server-side ad insertion, a broad range of devices, transcoding, and accurate reporting of server-side and client-side ad insertion.

Instead of listing out all of the features in the sections below, I’ve simply included as many screen shots as possible with the expectation that this will give you a better sense of the rich set of features, parameters, and settings that you get with this set of services.

AWS Elemental MediaConvert
MediaConvert allows you to transcode content that is stored in files. You can process individual files or entire media libraries, or anything in-between. You simply create a conversion job that specifies the content and the desired outputs, and submit it to MediaConvert. There’s no software to install or patch and the service scales to meet your needs without affecting turnaround time or performance.

The MediaConvert Console lets you manage Output presets, Job templates, Queues, and Jobs:

You can use a built-in system preset or you can make one of your own. You have full control over the settings when you make your own:

Jobs templates are named, and produce one or more output groups. You can add a new group to a template with a click:

When everything is ready to go, you create a job and make some final selections, then click on Create:

Each account starts with a default queue for jobs, where incoming work is processed in parallel using all processing resources available to the account. Adding queues does not add processing resources, but does cause them to be apportioned across queues. You can temporarily pause one queue in order to devote more resources to the others. You can submit jobs to paused queues and you can also cancel any that have yet to start.

Pricing for this service is based on the amount of video that you process and the features that you use.

AWS Elemental MediaLive
This service is for live encoding, and can be run 24×7. MediaLive channels are deployed on redundant resources distributed in two physically separated Availability Zones in order to provide the reliability expected by our customers in the broadcast industry. You can specify your inputs and define your channels in the MediaLive Console:

After you create an Input, you create a Channel and attach it to the Input:

You have full control over the settings for each channel:

 

AWS Elemental MediaPackage
This service lets you deliver video to many devices from a single source. It focuses on protection and just-in-time packaging, giving you the ability to provide your users with the desired content on the device of their choice. You simply create a channel to get started:

Then you add one or more endpoints. Once again, plenty of options and full control, including a startover window and a time delay:

You find the input URL, user name, and password for your channel and route your live video stream to it for packaging:

AWS Elemental MediaStore
MediaStore offers the performance, consistency, and latency required for live and on-demand media delivery. Objects are written and read into a new “temporal” tier of object storage for a limited amount of time, then move silently into S3 for long-lived durability. You simply create a storage container to group your media content:

The container is available within a minute or so:

Like S3 buckets, MediaStore containers have access policies and no limits on the number of objects or storage capacity.

MediaStore helps you to take full advantage of S3 by managing the object key names so as to maximize storage and retrieval throughput, in accord with the Request Rate and Performance Considerations.

AWS Elemental MediaTailor
This service takes care of server-side ad insertion while providing a broadcast-quality viewer experience by transcoding ad assets on the fly. Your customer’s video player asks MediaTailor for a playlist. MediaTailor, in turn, calls your Ad Decision Server and returns a playlist that references the origin server for your original video and the ads recommended by the Ad Decision Server. The video player makes all of its requests to a single endpoint in order to ensure that client-side ad-blocking is ineffective. You simply create a MediaTailor Configuration:

Context information is passed to the Ad Decision Server in the URL:

Despite the length of this post I have barely scratched the surface of the AWS Media Services. Once AWS re:Invent is in the rear view mirror I hope to do a deep dive and show you how to use each of these services.

Available Now
The entire set of AWS Media Services is available now and you can start using them today! Pricing varies by service, but is built around a pay-as-you-go model.

Jeff;