A little over a month ago, a large group of Beam community members has been
working a prototype of a portable Flink runner - that is, a runner that can
execute Beam pipelines on Flink via the Portability API
<https://s.apache.org/beam-runner-api>. The prototype was developed in
branch <https://github.com/bsidhom/beam/tree/hacking-job-server> and was
successfully demonstrated at Flink Forward, where it ran Python and Go
pipelines in a limited setting.
Since then, a smaller group of people (Ankur Goenka, Axel Magnuson, Ben
Sidhom and myself) have been working on productionizing the prototype to
address its limitations and do things "the right way", preparing to reuse
this work for developing other portable runners (e.g. Spark). This involves
a surprising amount of work, since many important design and implementation
concerns could be ignored for the purposes of a prototype. I wanted to give
an update on where we stand now.
Our immediate milestone in sight is *Run Java and Python batch WordCount
examples against a distributed remote Flink cluster*. That involves a few
moving parts, roughly in order of appearance:
- The SDK is configured to use a "portable runner", whose responsibility is
to run the pipeline against a given JobService endpoint.
- The portable runner converts the pipeline to a portable Pipeline proto
- The runner finds out which artifacts it needs to stage, and staging them
against an ArtifactStagingService
- A Flink-specific JobService receives the Pipeline proto, performs some
optimizations (e.g. fusion) and translates it to Flink datasets and
- A Flink function executes a fused chain of Beam transforms (an
"executable stage") by converting the input and the stage to bundles and
executing them against an SDK harness
- The function starts the proper SDK harness, auxiliary services (e.g.
artifact retrieval, side input handling) and wires them together
- The function feeds the data to the harness and receives data back.
*And here is our status of implementation for these parts:* basically,
almost everything is either done or in review.
- General-purpose portable runner in the Python SDK: done
<https://github.com/apache/beam/pull/5301>; Java SDK: also done
- Artifact staging from the Python SDK: in review (PR
<https://github.com/apache/beam/pull/5251>); in java, it's done also
- Flink JobService: in review <https://github.com/apache/beam/pull/5262>
- Translation from a Pipeline proto to Flink datasets and functions: done
- ArtifactStagingService implementation that stages artifacts to a location
on a distributed filesystem: in development (design is clear)
- Flink function for executing via an SDK harness: done
- APIs for managing lifecycle of an SDK harness: done
- Specific implementation of those APIs using Docker: part done
<https://github.com/apache/beam/pull/5189>, part in review
- ArtifactRetrievalService that retrieves artifacts from the location where
ArtifactStagingService staged them: in development.
We expect that the in-review parts will be done, and the in-development
parts be developed, in the next 2-3 weeks. We will, of course, update the
community when this important milestone is reached.
*After that, the next milestones include:*
- Sett up Java, Python and Go ValidatesRunner tests to run against the
portable Flink runner, and get them to pass
- Expand Python and Go to parity in terms of such test coverage
- Implement the portable Spark runner, with a similar lifecycle but reusing
almost all of the Flink work
- Add support for streaming to both (which requires SDF - that work is
progressing in parallel and by this point should be in a suitable place)
*For people who would like to get involved in this effort: *You can already
help out by improving ValidatesRunner test coverage in Python and Go. Java
has >300 such tests, Python has only a handful. There'll be a large amount
of parallelizable work once we get the VR test suites running - stay tuned.
SDF+Portability is also expected to produce a lot of parallelizable work up
for grabs within several weeks.