AWS Big Data Blog

Amazon Managed Service for Apache Flink now supports Apache Flink version 1.18

Apache Flink is an open source distributed processing engine, offering powerful programming interfaces for both stream and batch processing, with first-class support for stateful processing and event time semantics. Apache Flink supports multiple programming languages, Java, Python, Scala, SQL, and multiple APIs with different level of abstraction, which can be used interchangeably in the same application.

Amazon Managed Service for Apache Flink, which offers a fully managed, serverless experience in running Apache Flink applications, now supports Apache Flink 1.18.1, the latest version of Apache Flink at the time of writing.

In this post, we discuss some of the interesting new features and capabilities of Apache Flink, introduced with the most recent major releases, 1.16, 1.17, and 1.18, and now supported in Managed Service for Apache Flink.

New connectors

Before we dive into the new functionalities of Apache Flink available with version 1.18.1, let’s explore the new capabilities that come from the availability of many new open source connectors.


A dedicated OpenSearch connector is now available to be included in your projects, enabling an Apache Flink application to write data directly into OpenSearch, without relying on Elasticsearch compatibility mode. This connector is compatible with Amazon OpenSearch Service provisioned and OpenSearch Service Serverless.

This new connector supports SQL and Table APIs, working with both Java and Python, and the DataStream API, for Java only. Out of the box, it provides at-least-once guarantees, synchronizing the writes with Flink checkpointing. You can achieve exactly-once semantics using deterministic IDs and upsert method.

By default, the connector uses OpenSearch version 1.x client libraries. You can switch to version 2.x by adding the correct dependencies.

Amazon DynamoDB

Apache Flink developers can now use a dedicated connector to write data into Amazon DynamoDB. This connector is based on the Apache Flink AsyncSink, developed by AWS and now an integral part of the Apache Flink project, to simplify the implementation of efficient sink connectors, using non-blocking write requests and adaptive batching.

This connector also supports both SQL and Table APIs, Java and Python, and DataStream API, for Java only. By default, the sink writes in batches to optimize throughput. A notable feature of the SQL version is support for the PARTITIONED BY clause. By specifying one or more keys, you can achieve some client-side deduplication, only sending the latest record per key with each batch write. An equivalent can be achieved with the DataStream API by specifying a list of partition keys for overwriting within each batch.

This connector only works as a sink. You cannot use it for reading from DynamoDB. To look up data in DynamoDB, you still need to implement a lookup using the Flink Async I/O API or implementing a custom user-defined function (UDF), for SQL.


Another interesting connector is for MongoDB. In this case, both source and sink are available, for both the SQL and Table APIs and DataStream API. The new connector is now officially part of the Apache Flink project and supported by the community. This new connector replaces the old one provided by MongoDB directly, which only supports older Flink Sink and Source APIs.

As for other data store connectors, the source can either be used as a bounded source, in batch mode, or for lookups. The sink works both in batch mode and streaming, supporting both upsert and append mode.

Among the many notable features of this connector, one that’s worth mentioning is the ability to enable caching when using the source for lookups. Out of the box, the sink supports at-least-once guarantees. When a primary key is defined, the sink can support exactly-once semantics via idempotent upserts. The sink connector also supports exactly-once semantics, with idempotent upserts, when the primary key is defined.

New connector versioning

Not a new feature, but an important factor to consider when updating an older Apache Flink application, is the new connector versioning. Starting from Apache Flink version 1.17, most connectors have been externalized from the main Apache Flink distribution and follow independent versioning.

To include the right dependency, you need to specify the artifact version with the form: <connector-version>-<flink-version>

For example, the latest Kafka connector, also working with Amazon Managed Streaming for Apache Kafka (Amazon MSK), at the time of writing is version 3.1.0. If you are using Apache Flink 1.18, the dependency to use will be the following:


For Amazon Kinesis, the new connector version is 4.2.0. The dependency for Apache Flink 1.18 will be the following:


In the following sections, we discuss more of the powerful new features now available in Apache Flink 1.18 and supported in Amazon Managed Service for Apache Flink.


In Apache Flink SQL, users can provide hints to join queries that can be used to suggest the optimizer to have an effect in the query plan. In particular, in streaming applications, lookup joins are used to enrich a table, representing streaming data, with data that is queried from an external system, typically a database. Since version 1.16, several improvements have been introduced for lookup joins, allowing you to adjust the behavior of the join and improve performance:

  • Lookup cache is a powerful feature, allowing you to cache in-memory the most frequently used records, reducing the pressure on the database. Previously, lookup cache was specific to some connectors. Since Apache Flink 1.16, this option has become available to all connectors internally supporting lookup (FLIP-221). As of this writing, JDBC, Hive, and HBase connectors support lookup cache. Lookup cache has three available modes: FULL, for a small dataset that can be held entirely in memory, PARTIAL, for a large dataset, only caching the most recent records, or NONE, to completely disable cache. For PARTIAL cache, you can also configure the number of rows to buffer and the time-to-live.
  • Async lookup is another feature that can greatly improve performance. Async lookup provides in Apache Flink SQL a functionality similar to Async I/O available in the DataStream API. It allows Apache Flink to emit new requests to the database without blocking the processing thread until responses to previous lookups have been received. Similarly to Async I/O, you can configure async lookup to enforce ordering or allow unordered results, or adjust the buffer capacity and the timeout.
  • You can also configure a lookup retry strategy in combination with PARTIAL or NONE lookup cache, to configure the behavior in case of a failed lookup in the external database.

All these behaviors can be controlled using a LOOKUP hint, like in the following example, where we show a lookup join using async lookup:

    /*+ LOOKUP('table'='Customers', 'async'='true', 'output-mode'='allow_unordered') */ 
    O.order_id,, C.address
FROM Orders AS O 
JOIN Customers FOR SYSTEM_TIME AS OF O.proc_time AS C 
  ON O.customer_id = O.customer_id


In this section, we discuss new improvements and support in PyFlink.

Python 3.10 support

Apache Flink newest versions introduced several improvements for PyFlink users. First and foremost, Python 3.10 is now supported, and Python 3.6 support has been completely removed (FLINK-29421). Managed Service for Apache Flink currently uses Python 3.10 runtime to run PyFlink applications.

Getting closer to feature parity

From the perspective of the programming API, PyFlink is getting closer to Java on every version. The DataStream API now supports features like side outputs and broadcast state, and gaps on windowing API have been closed. PyFlink also now supports new connectors like Amazon Kinesis Data Streams directly from the DataStream API.

Thread mode improvements

PyFlink is very efficient. The overhead of running Flink API operators in PyFlink is minimal compared to Java or Scala, because the runtime actually runs the operator implementation in the JVM directly, regardless of the language of your application. But when you have a user-defined function, things are slightly different. A line of Python code as simple as lambda x: x + 1, or as complex as a Pandas function, must run in a Python runtime.

By default, Apache Flink runs a Python runtime on each Task Manager, external to the JVM. Each record is serialized, handed to the Python runtime via inter-process communication, deserialized, and processed in the Python runtime. The result is then serialized and handed back to the JVM, where it’s deserialized. This is the PyFlink PROCESS mode. It’s very stable but it introduces an overhead, and in some cases, it may become a performance bottleneck.

Since version 1.15, Apache Flink also supports THREAD mode for PyFlink. In this mode, Python user-defined functions are run within the JVM itself, removing the serialization/deserialization and inter-process communication overhead. THREAD mode has some limitations; for example, THREAD mode cannot be used for Pandas or UDAFs (user-defined aggregate functions, consisting of many input records and one output record), but can substantially improve performance of a PyFlink application.

With version 1.16, the support of THREAD mode has been substantially extended, also covering the Python DataStream API.

THREAD mode is supported by Managed Service for Apache Flink, and can be enabled directly from your PyFlink application.

Apple Silicon support

If you use Apple Silicon-based machines to develop PyFlink applications, developing for PyFlink 1.15, you have probably encountered some of the known Python dependency issues on Apple Silicon. These issues have been finally resolved (FLINK-25188). These limitations did not affect PyFlink applications running on Managed Service for Apache Flink. Before version 1.16, if you wanted to develop a PyFlink application on a machine using M1, M2, or M3 chipset, you had to use some workarounds, because it was impossible to install PyFlink 1.15 or earlier directly on the machine.

Unaligned checkpoint improvements

Apache Flink 1.15 already supported Incremental Checkpoints and Buffer Debloating. These features can be used, particularly in combination, to improve checkpoint performance, making checkpointing duration more predictable, especially in the presence of backpressure. For more information about these features, see Optimize checkpointing in your Amazon Managed Service for Apache Flink applications with buffer debloating and unaligned checkpoints.

With versions 1.16 and 1.17, several changes have been introduced to improve stability and performance.

Handling data skew

Apache Flink uses watermarks to support event-time semantics. Watermarks are special records, normally injected in the flow from the source operator, that mark the progress of event time for operators like event time windowing aggregations. A common technique is delaying watermarks from the latest observed event time, to allow events to be out of order, at least to some degree.

However, the use of watermarks comes with a challenge. When the application has multiple sources, for example it receives events from multiple partitions of a Kafka topic, watermarks are generated independently for each partition. Internally, each operator always waits for the same watermark on all input partitions, practically aligning it on the slowest partition. The drawback is that if one of the partitions is not receiving data, watermarks don’t progress, increasing the end-to-end latency. For this reason, an optional idleness timeout has been introduced in many streaming sources. After the configured timeout, watermark generation ignores any partition not receiving any record, and watermarks can progress.

You can also face a similar but opposite challenge if one source is receiving events much faster than the others. Watermarks are aligned to the slowest partition, meaning that any windowing aggregation will wait for the watermark. Records from the fast source have to wait, being buffered. This may result in buffering an excessive volume of data, and an uncontrollable growth of operator state.

To address the issue of faster sources, starting with Apache Flink 1.17, you can enable watermark alignment of source splits (FLINK-28853). This mechanism, disabled by default, makes sure that no partitions progress their watermarks too fast, compared to other partitions. You can bind together multiple sources, like multiple input topics, assigning the same alignment group ID, and configuring the duration of the maximal drift from the current watermark. If one specific partition is receiving events too fast, the source operator pauses consuming that partition until the drift is reduced below the configured threshold.

You can enable it for each source separately. All you need is to specify an alignment group ID, which will bind together all sources that have the same ID, and the duration of the maximal drift from the current minimal watermark. This will pause consuming from the source subtask that are advancing too fast, until the drift is lower than the threshold specified.

The following code snippet shows how you can set up watermark alignment of source splits on a Kafka source emitting bounded-out-of-orderness watermarks:

KafkaSource<Event> kafkaSource = ...
DataStream<Event> stream = env.fromSource(
    WatermarkStrategy.<Event>forBoundedOutOfOrderness( Duration.ofSeconds(20))
        .withWatermarkAlignment("alignment-group-1", Duration.ofSeconds(20), Duration.ofSeconds(1)),
    "Kafka source"));

This feature is only available with FLIP-217 compatible sources, supporting watermark alignment of source splits. As of writing, among major streaming source connectors, only Kafka source supports this feature.

Direct support for Protobuf format

The SQL and Table APIs now directly support Protobuf format. To use this format, you need to generate the Protobuf Java classes from the .proto schema definition files and include them as dependencies in your application.

The Protobuf format only works with the SQL and Table APIs and only to read or write Protobuf-serialized data from a source or to a sink. Currently, Flink doesn’t directly support Protobuf to serialize state directly and it doesn’t support schema evolution, as it does for Avro, for example. You still need to register a custom serializer with some overhead for your application.

Keeping Apache Flink open source

Apache Flink internally relies on Akka for sending data between subtasks. In 2022, Lightbend, the company behind Akka, announced a license change for future Akka versions, from Apache 2.0 to a more restrictive license, and that Akka 2.6, the version used by Apache Flink, would not receive any further security update or fix.

Although Akka has been historically very stable and doesn’t require frequent updates, this license change represented a risk for the Apache Flink project. The decision of the Apache Flink community was to replace Akka with a fork of the version 2.6, called Apache Pekko (FLINK-32468). This fork will retain the Apache 2.0 license and receive any required updates by the community. In the meantime, the Apache Flink community will consider whether to remove the dependency on Akka or Pekko completely.

State compression

Apache Flink offers optional compression (default: off) for all checkpoints and savepoints. Apache Flink identified a bug in Flink 1.18.1 where the operator state couldn’t be properly restored when snapshot compression is enabled. This could result in either data loss or inability to restore from checkpoint. To resolve this, Managed Service for Apache Flink has backported the fix that will be included in future versions of Apache Flink.

In-place version upgrades with Managed Service for Apache Flink

If you are currently running an application on Managed Service for Apache Flink using Apache Flink 1.15 or older, you can now upgrade it in-place to 1.18 without losing the state, using the AWS Command Line Interface (AWS CLI), AWS CloudFormation or AWS Cloud Development Kit (AWS CDK), or any tool that uses the AWS API.

The UpdateApplication API action now supports updating the Apache Flink runtime version of an existing Managed Service for Apache Flink application. You can use UpdateApplication directly on a running application.

Before proceeding with the in-place update, you need to verify and update the dependencies included in your application, making sure they are compatible with the new Apache Flink version. In particular, you need to update any Apache Flink library, connectors, and possibly Scala version.

Also, we recommend testing the updated application before proceeding with the update. We recommend testing locally and in a non-production environment, using the target Apache Flink runtime version, to ensure no regressions were introduced.

And finally, if your application is stateful, we recommend taking a snapshot of the running application state. This will enable you to roll back to the previous application version.

When you’re ready, you can now use the UpdateApplication API action or update-application AWS CLI command to update the runtime version of the application and point it to the new application artifact, JAR, or zip file, with the updated dependencies.

For more detailed information about the process and the API, refer to In-place version upgrade for Apache Flink. The documentation includes a step by step instructions and a video to guide you through the upgrade process.


In this post, we examined some of the new features of Apache Flink, supported in Amazon Managed Service for Apache Flink. This list is not comprehensive. Apache Flink also introduced some very promising features, like operator-level TTL for the SQL and Table API [FLIP-292] and Time Travel [FLIP-308], but these are not yet supported by the API, and not really accessible to users yet. For this reason, we decided not to cover them in this post.

With the support of Apache Flink 1.18, Managed Service for Apache Flink now supports the latest released Apache Flink version. We have seen some of the interesting new features and new connectors available with Apache Flink 1.18 and how Managed Service for Apache Flink helps you upgrade an existing application in place.

You can find more details about recent releases from the Apache Flink blog and release notes:

If you are new to Apache Flink, we recommend our guide to choosing the right API and language and following the getting started guide to start using Managed Service for Apache Flink.

About the Authors

Lorenzo NicoraLorenzo Nicora works as Senior Streaming Solution Architect at AWS, helping customers across EMEA. He has been building cloud-native, data-intensive systems for over 25 years, working in the finance industry both through consultancies and for FinTech product companies. He has leveraged open-source technologies extensively and contributed to several projects, including Apache Flink.

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