Deploying as an application with durable state - Managed Service for Apache Flink
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.

Deploying as an application with durable state

You can build your code and export it to Amazon S3. You can promote the code that you wrote in your note to a continuously running stream processing application. There are two modes of running an Apache Flink application on Managed Service for Apache Flink: With a Studio notebook, you have the ability to develop your code interactively, view results of your code in real time, and visualize it within your note. After you deploy a note to run in streaming mode, Managed Service for Apache Flink creates an application for you that runs continuously, reads data from your sources, writes to your destinations, maintains long-running application state, and autoscales automatically based on the throughput of your source streams.


The S3 bucket to which you export your application code must be in the same Region as your Studio notebook.

You can only deploy a note from your Studio notebook if it meets the following criteria:

  • Paragraphs must be ordered sequentially. When you deploy your application, all paragraphs within a note will be executed sequentially (left-to-right, top-to-bottom) as they appear in your note. You can check this order by choosing Run All Paragraphs in your note.

  • Your code is a combination of Python and SQL or Scala and SQL. We do not support Python and Scala together at this time for deploy-as-application.

  • Your note should have only the following interpreters: %flink, %flink.ssql, %flink.pyflink, %flink.ipyflink, %md.

  • The use of the Zeppelin context object z is not supported. Methods that return nothing will do nothing except log a warning. Other methods will raise Python exceptions or fail to compile in Scala.

  • A note must result in a single Apache Flink job.

  • Notes with dynamic forms are unsupported for deploying as an application.

  • %md (Markdown) paragraphs will be skipped in deploying as an application, as these are expected to contain human-readable documentation that is unsuitable for running as part of the resulting application.

  • Paragraphs disabled for running within Zeppelin will be skipped in deploying as an application. Even if a disabled paragraph uses an incompatible interpreter, for example, %flink.ipyflink in a note with %flink and %flink.ssql interpreters, it will be skipped while deploying the note as an application, and will not result in an error.

  • There must be at least one paragraph present with source code (Flink SQL, PyFlink or Flink Scala) that is enabled for running for the application deployment to succeed.

  • Setting parallelism in the interpreter directive within a paragraph (e.g. %flink.ssql(parallelism=32)) will be ignored in applications deployed from a note. Instead, you can update the deployed application through the Amazon Web Services Management Console, Amazon Command Line Interface or Amazon API to change the Parallelism and/or ParallelismPerKPU settings according to the level of parallelism your application requires, or you can enable autoscaling for your deployed application.

  • If you are deploying as an application with durable state your VPC must have internet access. If your VPC does not have internet access, see Deploying as an application with durable state in a VPC with no internet access.

Scala/Python criteria

  • In your Scala or Python code, use the Blink planner (senv, stenv for Scala; s_env, st_env for Python) and not the older "Flink" planner (stenv_2 for Scala, st_env_2 for Python). The Apache Flink project recommends the use of the Blink planner for production use cases, and this is the default planner in Zeppelin and in Flink.

  • Your Python paragraphs must not use shell invocations/assignments using ! or IPython magic commands like %timeit or %conda in notes meant to be deployed as applications.

  • You can't use Scala case classes as parameters of functions passed to higher-order dataflow operators like map and filter. For information about Scala case classes, see CASE CLASSES in the Scala documentation.

SQL criteria

  • Simple SELECT statements are not permitted, as there’s nowhere equivalent to a paragraph’s output section where the data can be delivered.

  • In any given paragraph, DDL statements (USE, CREATE, ALTER, DROP, SET, RESET) must precede DML (INSERT) statements. This is because DML statements in a paragraph must be submitted together as a single Flink job.

  • There should be at most one paragraph that has DML statements in it. This is because, for the deploy-as-application feature, we only support submitting a single job to Flink.

For more information and an example, see Translate, redact and analyze streaming data using SQL functions with Amazon Managed Service for Apache Flink, Amazon Translate, and Amazon Comprehend.