Flink is a versatile framework, supporting many different deployment scenarios in a mix and match fashion.
Below, we briefly explain the building blocks of a Flink cluster, their purpose and available implementations. If you just want to start Flink locally, we recommend setting up a Standalone Cluster.
Overview and Reference Architecture #
The figure below shows the building blocks of every Flink cluster. There is always somewhere a client running. It takes the code of the Flink applications, transforms it into a JobGraph and submits it to the JobManager.
The JobManager distributes the work onto the TaskManagers, where the actual operators (such as sources, transformations and sinks) are running.
When deploying Flink, there are often multiple options available for each building block. We have listed them in the table below the figure.
|Flink Client||Compiles batch or streaming applications into a dataflow graph, which it then submits to the JobManager.|
JobManager is the name of the central work coordination component of Flink. It has implementations for different resource providers, which differ on high-availability, resource allocation behavior and supported job submission modes.
JobManager modes for job submissions:
|TaskManager||TaskManagers are the services actually performing the work of a Flink job.|
|External Components (all optional)|
|High Availability Service Provider||Flink's JobManager can be run in high availability mode which allows Flink to recover from JobManager faults. In order to failover faster, multiple standby JobManagers can be started to act as backups.|
|File Storage and Persistency||For checkpointing (recovery mechanism for streaming jobs) Flink relies on external file storage systems||See FileSystems page.|
|Resource Provider||Flink can be deployed through different Resource Provider Frameworks, such as Kubernetes or YARN.||See JobManager implementations above.|
|Metrics Storage||Flink components report internal metrics and Flink jobs can report additional, job specific metrics as well.||See Metrics Reporter page.|
|Application-level data sources and sinks||While application-level data sources and sinks are not technically part of the deployment of Flink cluster components, they should be considered when planning a new Flink production deployment. Colocating frequently used data with Flink can have significant performance benefits||
Repeatable Resource Cleanup #
Once a job has reached a globally terminal state of either finished, failed or cancelled, the external component resources associated with the job are then cleaned up. In the event of a failure when cleaning up a resource, Flink will attempt to retry the cleanup. You can configure the retry strategy used. Reaching the maximum number of retries without succeeding will leave the job in a dirty state. Its artifacts would need to be cleaned up manually (see the High Availability Services / JobResultStore section for further details). Restarting the very same job (i.e. using the same job ID) will result in the cleanup being restarted without running the job again.
There is currently an issue with the cleanup of CompletedCheckpoints that failed to be deleted while subsuming them as part of the usual CompletedCheckpoint management. These artifacts are not covered by the repeatable cleanup, i.e. they have to be deleted manually, still. This is covered by FLINK-26606.
Deployment Modes #
Flink can execute applications in one of three ways:
- in Application Mode,
- in a Per-Job Mode,
- in Session Mode.
The above modes differ in:
- the cluster lifecycle and resource isolation guarantees
- whether the application’s
main()method is executed on the client or on the cluster.
Application Mode #
In all the other modes, the application’s
main() method is executed on the client side. This process
includes downloading the application’s dependencies locally, executing the
main() to extract a representation
of the application that Flink’s runtime can understand (i.e. the
JobGraph) and ship the dependencies and
JobGraph(s) to the cluster. This makes the Client a heavy resource consumer as it may need substantial
network bandwidth to download dependencies and ship binaries to the cluster, and CPU cycles to execute the
main(). This problem can be more pronounced when the Client is shared across users.
Building on this observation, the Application Mode creates a cluster per submitted application, but this time,
main() method of the application is executed on the JobManager. Creating a cluster per application can be
seen as creating a session cluster shared only among the jobs of a particular application, and torn down when
the application finishes. With this architecture, the Application Mode provides the same resource isolation
and load balancing guarantees as the Per-Job mode, but at the granularity of a whole application. Executing
main() on the JobManager allows for saving the CPU cycles required, but also save the bandwidth required
for downloading the dependencies locally. Furthermore, it allows for more even spread of the network load for
downloading the dependencies of the applications in the cluster, as there is one JobManager per application.
In the Application Mode, the
main()is executed on the cluster and not on the client, as in the other modes. This may have implications for your code as, for example, any paths you register in your environment using the
registerCachedFile()must be accessible by the JobManager of your application.
Compared to the Per-Job mode, the Application Mode allows the submission of applications consisting of
multiple jobs. The order of job execution is not affected by the deployment mode but by the call used
to launch the job. Using
execute(), which is blocking, establishes an order and it will lead to the
execution of the “next” job being postponed until “this” job finishes. Using
executeAsync(), which is
non-blocking, will lead to the “next” job starting before “this” job finishes.
The Application Mode allows for multi-
execute()applications but High-Availability is not supported in these cases. High-Availability in Application Mode is only supported for single-
Additionally, when any of multiple running jobs in Application Mode (submitted for example using
executeAsync()) gets cancelled, all jobs will be stopped and the JobManager will shut down. Regular job completions (by the sources shutting down) are supported.
Per-Job Mode #
Aiming at providing better resource isolation guarantees, the Per-Job mode uses the available resource provider framework (e.g. YARN, Kubernetes) to spin up a cluster for each submitted job. This cluster is available to that job only. When the job finishes, the cluster is torn down and any lingering resources (files, etc) are cleared up. This provides better resource isolation, as a misbehaving job can only bring down its own TaskManagers. In addition, it spreads the load of book-keeping across multiple JobManagers, as there is one per job. For these reasons, the Per-Job resource allocation model is the preferred mode by many production reasons.
Session Mode #
Session mode assumes an already running cluster and uses the resources of that cluster to execute any submitted application. Applications executed in the same (session) cluster use, and consequently compete for, the same resources. This has the advantage that you do not pay the resource overhead of spinning up a full cluster for every submitted job. But, if one of the jobs misbehaves or brings down a TaskManager, then all jobs running on that TaskManager will be affected by the failure. This, apart from a negative impact on the job that caused the failure, implies a potential massive recovery process with all the restarting jobs accessing the filesystem concurrently and making it unavailable to other services. Additionally, having a single cluster running multiple jobs implies more load for the JobManager, who is responsible for the book-keeping of all the jobs in the cluster.
In Session Mode, the cluster lifecycle is independent of that of any job running on the cluster
and the resources are shared across all jobs. The Per-Job mode pays the price of spinning up a cluster
for every submitted job, but this comes with better isolation guarantees as the resources are not shared
across jobs. In this case, the lifecycle of the cluster is bound to that of the job. Finally, the
Application Mode creates a session cluster per application and executes the application’s
method on the cluster.
Vendor Solutions #
A number of vendors offer managed or fully hosted Flink solutions. None of these vendors are officially supported or endorsed by the Apache Flink PMC. Please refer to vendor maintained documentation on how to use these products.
AliCloud Realtime Compute #
Supported Environments: AliCloud
Amazon EMR #
Supported Environments: AWS
Amazon Kinesis Data Analytics for Apache Flink #
Supported Environments: AWS
Cloudera DataFlow #
Supported Environment: AWS Azure Google On-Premise
Supported Environment: AWS
Huawei Cloud Stream Service #
Supported Environment: Huawei
Ververica Platform #
Supported Environments: AliCloud AWS Azure Google On-Premise