[GSoC 2026] Kafka Streams runner: validatesRunner task; Create and Flatten suites green#39380
Conversation
…atten suites green
Adds the validatesRunner gradle task running Beam's @ValidatesRunner tests from
the SDK core test jar against TestKafkaStreamsRunner (category excludes for
unsupported features, plus a sickbay list). CreateTest passes 5/5 and
FlattenTest 13/13, with one test sickbayed pending multi-output executable
stages.
Enabling the suite forced three runner fixes:
- Impulse bootstrap topics are now per-transform: a pipeline can contain two
Impulses (an empty Create plus the dummy branch PAssert adds), and Kafka
Streams rejects registering one topic on two source nodes.
- Flatten is no longer in knownUrns(): declaring it native made
TrivialNativeTransformExpander corrupt the fused graph of a multi-branch
Flatten fed by composite expansions ("consumed but never produced" from the
fuser's own validation). Mirrors the Flink runner keeping Read out of its
known URNs for expander reasons.
- A Flatten of zero PCollections is rewritten before fusion into a primitive
Read of a new EmptyBoundedSource: a zero-input Flatten is a root of the
pipeline graph and GreedyPipelineFuser only accepts Impulse or Read roots,
and an empty read has exactly the empty-PCollection semantics (no elements,
then the terminal watermark).
|
Warning Gemini encountered an error creating the summary. You can try again by commenting |
|
/gemini summary |
Summary of ChangesThis pull request enables the 'validatesRunner' test suite for the Kafka Streams runner, significantly improving the runner's validation capabilities. The changes address critical issues discovered during testing, specifically regarding how Impulse transforms register topics and how Flatten transforms are handled within the pipeline graph to avoid fusion errors. Highlights
New Features🧠 You can now enable Memory (public preview) to help Gemini Code Assist learn from your team's feedback. This makes future code reviews more consistent and personalized to your project's style. Click here to enable Memory in your admin console. Activity
|
|
Checks are failing. Will not request review until checks are succeeding. If you'd like to override that behavior, comment |
1 similar comment
|
Checks are failing. Will not request review until checks are succeeding. If you'd like to override that behavior, comment |
The task previously ran the whole ValidatesRunner category, including suites that have not been triaged yet (windowed GroupByKey, PAssert singleton views, metric types beyond counters). Feature gaps like non-global windowing have no JUnit category to exclude, so the task now declares the enabled test classes explicitly -- CreateTest and FlattenTest -- and the list grows class by class as runner support does.
1 participant
Summary
Adds the
validatesRunnergradle task, running Beam's @ValidatesRunner tests (from the SDK core test jar) againstTestKafkaStreamsRunner, category excludes for features the runner does not support yet (side inputs, state,timers, TestStream, unbounded, SDF, ...) plus a sickbay list for known-failing tests. Part of #18479; this is the milestone the midterm pointed at.
Current results:
CreateTest: 5/5FlattenTest: 13/13, 1 sickbayed (needs multi-output executable stages -- output-tag dispatch in the stage processor; follow-up)Enabling the suite immediately caught three real bugs, fixed here:
Impulses (an empty
Createplus the dummy branchPAssertaddsinternally), and Kafka Streams rejects registering the same topic on two
source nodes.
2.
Flattenis no longer declared inknownUrns(). Declaring it native madeTrivialNativeTransformExpandercorrupt the fused graph of a multi-branchFlatten whose inputs come from composite expansions (the fuser's own
validation failed with "consumed but never produced"). This one change fixed
four failing tests. Mirrors the Flink runner, which keeps Read out of its
known URNs for expander reasons.
Read of a new
EmptyBoundedSource. A zero-input Flatten is a root of thepipeline graph, and
GreedyPipelineFuseronly accepts Impulse or Readroots; an empty read has exactly the right semantics (no elements, then the
terminal watermark) and reuses the existing Read translator.
Next: enabling the GroupByKey and ParDo suites and triaging their failures the same way, and a follow-up issue for multi-output executable stages.