Achieve production readiness for your Managed Service for Apache Flink applications - Managed Service for Apache Flink
Load-test your applicationsDefine Max parallelism Set a UUID for all operators
Services or capabilities described in Amazon Web Services documentation might vary by Region. To see the differences applicable to the China Regions, see Getting Started with Amazon Web Services in China (PDF).

Amazon Managed Service for Apache Flink was previously known as Amazon Kinesis Data Analytics for Apache Flink.

Achieve production readiness for your Managed Service for Apache Flink applications

This is a collection of important aspects of running production applications on Managed Service for Apache Flink. It's not an exhaustive list, but rather the bare minimum of what you should pay attention to before putting an application into production.

Load-test your applications

Some problems with applications only manifest under heavy load. We have seen cases where applications seemed healthy, yet an operational event substantially amplified the load on the application. This can happen completely independent of the application itself. If the data source or the data sink is unavailable for a couple of hours, the Flink application cannot make progress. When that issue is fixed, there is a backlog of unprocessed data that has accumulated, which can completely exhaust the available resources. The load can then amplify bugs or performance issues that had not emerged before.

It is therefore essential that you run proper load tests for production applications. Questions that should be answered during those load tests include:

  • Is the application stable under sustained high load?

  • Can the application still take a savepoint under peak load?

  • How long does it take to process a backlog of 1 hour? And how long for 24 hours (depending on the max retention of the data in the stream)?

  • Does the throughput of the application increase when the application is scaled?

When consuming from a data stream, these scenarios can be simulated by producing into the stream for some time. Then start the application and have it consume data from the beginning of time. For example, use a start position of TRIM_HORIZON in the case of a Kinesis data stream.

Define Max parallelism

The max parallelism defines the maximum parallelism a stateful application can scale to. This is defined when the state is first created and there is no way of scaling the operator beyond this maximum without discarding the state.

Max parallelism is set when the state is first created.

By default, Max parallelism is set to:

  • 128, if parallelism <= 128

  • MIN(nextPowerOfTwo(parallelism + (parallelism / 2)), 2^15): if parallelism > 128

If you are planning to scale your application > 128 parallelism, you should explicitly define the Max parallelism.

You can define Max parallelism at level of application, with env.setMaxParallelism(x) or single operator. Unless differently specified, all operators inherit the Max parallelism of the application.

For more information, see Setting the Maximum Parallelism in the Apache Flink Documentation.

Set a UUID for all operators

A UUID is used in the operation in which Flink maps a savepoint back to an individual operator. Setting a specific UUID for each operator gives a stable mapping for the savepoint process to restore.

.map(...).uid("my-map-function")

For more information, see Production Readiness Checklist.