Apache Beam: easily implement backoff policy in your DoFn

In Apache Beam, DoFn is your swiss knife: when you don’t have an existing PTransform or CompositeTransform provided by the SDK, you can create your own function.

DoFn ?

A DoFn applies your logic in each element in the input PCollection and let you populate the elements of an output PCollection. To be included in your pipeline, it’s wrapped in a ParDo PTransform.

For instance, you can transform element using a DoFn:

pipeline.apply("ReadFromJms", JmsIO.read().withConnectionFactory(CF).withQueue("city"))         .apply("TransformJmsRecordAsPojo", ParDo.of(new DoFn<JmsRecord, MyCityPojo>() {                   @ProcessElement                   public void processElement(ProcessContext c) {                       String payload = c.element().getPayload();                       MyCityPojo city = new MyCityPojo(payload);                       c.output(city);                   }          })

We can see here the core method of DoFn: processElement() annotated with @ProcessElement.

If you have to deal with resources, other annotated methods allow you to tweak the DoFn lifecycle:

  1. @Setup is called when the DoFn is created on the worker.
  2. @StartBundle is called when the runner starts a bundle of data. The bundle size depends of the runner.
  3. @FinishBundle is called when all elements in the bundle has been processed.
  4. @Teardown is called when the DoFn is stopped on the worker.

Adding Backoff Policy

Now, in your @ProcessElement method, if an exception occurs, the current bundle is just invalidated and not processed.

So, you might want to implement a retry strategy, replaying the data processing logic.

Apache Beam provides utils allowing you to easily implement this retry policy: it’s what we name backoff policy.

FluentBackoff is a convenient util class where you can define a retry policy with:

  • an initial backoff delay between the retry (incremented at each attempt)
  • a max number of retries

To illustrate this, let’s create a DoFn using this:

public class DoFnWithBackoff extends DoFn<JmsRecord, MyCityPojo> {   private static final FluentBackoff BACKOFF =       FluentBackoff.DEFAULT           .withMaxRetries(5)           .withInitialBackoff(Duration.standardSeconds(5));   @ProcessElement   public void processElement(ProcessContext c) {     Sleeper sleeper = Sleeper.DEFAULT;     BackOff backoff = BACKOFF.backoff();     while (true) {       try {          String payload = c.element().getPayload();          MyCityPojo city = new MyCityPojo(payload); // can raise an exception          c.output(city);       } catch (Exception e) {         if (e instanceOf BusinessException.class) {           // in case of custom business exception, we don't retry           throw new BusinessException();         }         // retrying changing the content if required         if (!BackOffUtils.next(sleeper, backoff)) {           // we retried the max number of times           throw e;         }       }     }   }}

The Apache Beam backoff utils give you complete control and you can implement your own retry policy.

Comments

Post a Comment

Popular posts from this blog

Quarkus and "meta" extension

Quarkus is a great Java framework with a large ecosystem thanks to the extensions . Basically, if using Maven, your pom.xml contains the list of the extensions you want to use, meaning something like this: ... <dependency> <groupId>io.quarkus</groupId> <artifactId>quarkus-grpc</artifactId> </dependency> <dependency> <groupId>io.quarkus</groupId> <artifactId>quarkus-opentelemetry</artifactId> </dependency> ... It can quickly become large, and sometime you can have "conflict" between extensions (for example if you mix in the same dependencies set reactive and non reactive extensions). Quarkus "meta" extension To simplify the dependencies for the extensions you want to use (especially if it's basically the same extensions in all your services), you can create your own "meta" extension. This extension doesn't contain actual code, it "wraps" other exten
Read more

Getting started with Apache Karaf Minho

Image
Apache Karaf Minho Apache Karaf Minho is a new runtime, completely created from scratch. The objectives are: super easy to use and run ligtning fast, including GraalVM support Kubernetes compliant to be easily used on the cloud helping to optimize cloud costs by supporting applications colocation providing turnkey services allowing you to create Minho applications in a minute Minho is composed by: minho-boot is the core of Minho. It contains the runtime launcher and the core services (service registry, lifecycle service, configuration service, ...). minho-* services provided additional functionnalities in the runtime. You can use these services to create your own services, it's also where you will find the services managing sepcific kind of applications, allowing colocation. minho-tooling optionally provide convenient tools to create your own runtime. It includes minho-maven-plugin for instance. In this blog post, we will take a quick tour on A
Read more

Apache Karaf Minho and OpenTelemetry

Image
Telemetry is an important part of a software stack, especially on distributed system (where services call each others). OpenTelemetry is a CNCF project coming from OpenTracing and OpenCensus. It provides API to easily instrument your code and tools to collect and export telemetry data (metrics, logs and traces). Before going into how to use OpenTelemetry with Minho, we have to understand some key concepts: Span is a single building block representing an operation or some unit of work that you want to capture. They are capable of standing on their own, referencing (or following) from other spans, storing metadata, tags, etc. Trace is a visualisation of the life of a request (or series of other operations) that is made up of one or more spans. This collection of spans work together to paint a picture of the overall request, allowing you to easily reconstruct what happened within each span, etc. SpanContext is a wrapper of key values pairs that associates a trace to one or mor
Read more