Fix DoFnInvoker cache collision for generic types (#37355) This fixes a bug where ByteBuddyDoFnInvokerFactory would return the same cached invoker for different generic instantiations of the same DoFn class. Changes: 1. Introduced InvokerCacheKey with TypeDescriptors to ensure unique cache entries. 2. Updated generateInvokerClass to append type-based hash suffix. 3. Added regression test (testCacheKeyCollisionProof). This PR fixes a critical issue where ByteBuddyDoFnInvokerFactory failed to distinguish between different generic instantiations of the same DoFn class (e.g., MyFn<String> vs MyFn<Integer>). 1. Cache Key Strategy: Introduced InvokerCacheKey to include input/output TypeDescriptors in the cache lookup. 2. Class Naming: Updated generateInvokerClass to append a type-based hash suffix to ensure unique class names. 3. Robustness (The Fix): Added defensive try-catch blocks when accessing TypeDescriptors. - Some internal transforms (like MapElements) throw IllegalStateException if getOutputTypeDescriptor() is called after serialization. - In these cases, the factory now gracefully falls back to using Object.class (legacy behavior), ensuring backward compatibility for transforms that do not retain type information at runtime.
Apache Beam is a unified model for defining both batch and streaming data-parallel processing pipelines, as well as a set of language-specific SDKs for constructing pipelines and Runners for executing them on distributed processing backends, including Apache Flink, Apache Spark, Google Cloud Dataflow, and Hazelcast Jet.
Beam provides a general approach to expressing embarrassingly parallel data processing pipelines and supports three categories of users, each of which have relatively disparate backgrounds and needs.
The model behind Beam evolved from several internal Google data processing projects, including MapReduce, FlumeJava, and Millwheel. This model was originally known as the “Dataflow Model”.
To learn more about the Beam Model (though still under the original name of Dataflow), see the World Beyond Batch: Streaming 101 and Streaming 102 posts on O’Reilly’s Radar site, and the VLDB 2015 paper.
The key concepts in the Beam programming model are:
PCollection: represents a collection of data, which could be bounded or unbounded in size.PTransform: represents a computation that transforms input PCollections into output PCollections.Pipeline: manages a directed acyclic graph of PTransforms and PCollections that is ready for execution.PipelineRunner: specifies where and how the pipeline should execute.Beam supports multiple language-specific SDKs for writing pipelines against the Beam Model.
Currently, this repository contains SDKs for Java, Python and Go.
Have ideas for new SDKs or DSLs? See the sdk-ideas label.
Beam supports executing programs on multiple distributed processing backends through PipelineRunners. Currently, the following PipelineRunners are available:
DirectRunner runs the pipeline on your local machine.PrismRunner runs the pipeline on your local machine using Beam Portability.DataflowRunner submits the pipeline to the Google Cloud Dataflow.FlinkRunner runs the pipeline on an Apache Flink cluster. The code has been donated from dataArtisans/flink-dataflow and is now part of Beam.SparkRunner runs the pipeline on an Apache Spark cluster.JetRunner runs the pipeline on a Hazelcast Jet cluster. The code has been donated from hazelcast/hazelcast-jet and is now part of Beam.Twister2Runner runs the pipeline on a Twister2 cluster. The code has been donated from DSC-SPIDAL/twister2 and is now part of Beam.Have ideas for new Runners? See the runner-ideas label.
Instructions for building and testing Beam itself are in the contribution guide.
Here are some resources actively maintained by the Beam community to help you get started:
To get involved with Apache Beam: