Distributed Architecture

A Stateful Functions deployment consists of a few components interacting together. Here we describe these pieces and their relationship to each other and the Apache Flink runtime.

High-level View

A Stateful Functions deployment consists of a set of Apache Flink Stateful Functions processes and, optionally, various deployments that execute remote functions.

The Flink worker processes (TaskManagers) receive the events from the ingress systems (Kafka, Kinesis, etc.) and route them to the target functions. They invoke the functions and route the resulting messages to the next respective target functions. Messages designated for egress are written to an egress system (again, Kafka, Kinesis, …).


The heavy lifting is done by the Apache Flink processes, which manage the state, handle the messaging, and invoke the stateful functions. The Flink cluster consists typically of one master and multiple workers (TaskManagers).

In addition to the Apache Flink processes, a full deployment requires ZooKeeper (for master failover) and bulk storage (S3, HDFS, NAS, GCS, Azure Blob Store, etc.) to store Flink’s checkpoints. In turn, the deployment requires no database, and Flink processes do not require persistent volumes.

Logical Co-location, Physical Separation

A core principle of many Stream Processors is that application logic and the application state must be co-located. That approach is the basis for their out-of-the box consistency. Stateful Function takes a unique approach to that by logically co-locating state and compute, but allowing to physically separate them.

  • Logical co-location: Messaging, state access/updates and function invocations are managed tightly together, in the same way as in Flink’s DataStream API. State is sharded by key, and messages are routed to the state by key. There is a single writer per key at a time, also scheduling the function invocations.

  • Physical separation: Functions can be executed remotely, with message and state access provided as part of the invocation request. This way, functions can be managed independently, like stateless processes.

Deployment Styles for Functions

The stateful functions themselves can be deployed in various ways that trade off certain properties with each other: loose coupling and independent scaling on the one hand with performance overhead on the other hand. Each module of functions can be of a different kind, so some functions can run remote, while others could run embedded.

Remote Functions

Remote Functions use the above-mentioned principle of physical separation while maintaining logical co-location. The state/messaging tier (i.e., the Flink processes) and the function tier are deployed, managed, and scaled independently.

Function invocations happen through an HTTP / gRPC protocol and go through a service that routes invocation requests to any available endpoint, for example a Kubernetes (load-balancing) service, the AWS request gateway for Lambda, etc. Because invocations are self-contained (contain message, state, access to timers, etc.) the target functions can treated like any stateless application.

Refer to the documentation on the Python SDK and remote modules for details.

Co-located Functions

An alternative way of deploying functions is co-location with the Flink JVM processes. In such a setup, each Flink TaskManager would talk to one Function process sitting “next to it”. A common way to do this is to use a system like Kubernetes and deploy pods consisting of a Flink container and the function side-car container; the two communicate via the pod-local network.

This mode supports different languages while avoiding to route invocations through a Service/LoadBalancer, but it cannot scale the state and compute parts independently.

This style of deployment is similar to how Flink’s Table API and API Beam’s portability layer deploy and execute non-JVM functions.

Embedded Functions

Embedded Functions are similar to the execution mode of Stateful Functions 1.0 and to Flink’s Java/Scala stream processing APIs. Functions are run in the JVM and are directly invoked with the messages and state access. This is the most performant way, though at the cost of only supporting JVM languages. Updates to functions mean updating the Flink cluster.

Following the database analogy, Embedded Functions are a bit like Stored Procedures, but in a more principled way: The Functions here are normal Java/Scala/Kotlin functions implementing standard interfaces, and can be developed/tested in any IDE.