MartijnVisser opened a new pull request, #28738:
URL: https://github.com/apache/flink/pull/28738
## What is the purpose of the change
Fixes the intermittent `NullPointerException: metadataHandlerProvider`
(Calcite `RelMetadataQueryBase`) that has recurred for years, only in
flink-python (e.g. `StreamPandasConversionTests`).
Root cause: Calcite keeps the metadata handler provider in a thread-local
(`RelMetadataQueryBase.THREAD_PROVIDERS`), seeded only on the thread that
builds the cached, shared `RelOptCluster` (`FlinkRelOptClusterFactory`).
PyFlink runs optimization (`to_pandas()` ->
`ArrowUtils.collectAsPandasDataFrame` -> `table.execute().collect()`) on a py4j
gateway thread that can differ from the cluster-building thread, where the
thread-local is unset, so the first metadata-handler lookup NPEs. The
nondeterministic py4j thread assignment is why it is intermittent and
PyFlink-only.
The fix re-seeds the thread-local in `FlinkRelMetadataQuery.instance()`, the
single funnel through which all metadata queries are obtained (directly, via
`reuseOrCreate`, and via the cluster's metadata-query supplier), using the same
provider `FlinkRelOptClusterFactory` installs.
Notes for reviewers:
- Scope: this addresses the sequential cross-thread case (the observed
failures). Concurrent optimization on multiple threads against the same shared,
non-thread-safe `RelMetadataQuery` is a separate, pre-existing concern that
this change neither introduces nor fixes.
- The thread-local is set and not cleared on foreign (pooled py4j) threads,
by parity with Calcite's own `RelOptCluster.setMetadataProvider` seeding; the
value is the stateless `FlinkDefaultRelMetadataProvider` singleton chain.
- Design choice: healing the `instance()` funnel is provider-agnostic and
robust to any foreign-thread caller, versus seeding at the PyFlink entry point,
which is more invasive and PyFlink-specific.
Likely duplicates of the same NPE: FLINK-33531, FLINK-38392.
## Brief change log
- Guard `FlinkRelMetadataQuery.instance()` to re-seed
`RelMetadataQueryBase.THREAD_PROVIDERS` when unset on the current thread
- Cross-reference comment in `FlinkRelOptClusterFactory` documenting the
single provider both sites rely on
- Add `FlinkRelMetadataQueryThreadLocalTest`, a deterministic red/green
reproduction of the cross-thread NPE
## Verifying this change
This change added tests and can be verified as follows:
- `FlinkRelMetadataQueryThreadLocalTest` builds the cluster on the main
thread and issues a metadata query on a fresh worker thread; it fails with
`NullPointerException: metadataHandlerProvider` without the fix and passes with
it
- `mvn test -Dtest=FlinkRelMd*` in `flink-table-planner` passes (315 tests)
## Does this pull request potentially affect one of the following parts:
- Dependencies (does it add or upgrade a dependency): no
- The public API, i.e., is any changed class annotated with
`@Public(Evolving)`: no
- The serializers: no
- The runtime per-record code paths (performance sensitive): no
- Anything that affects deployment or recovery: JobManager (and its
components), Checkpointing, Kubernetes/Yarn, ZooKeeper: no
- The S3 file system connector: no
## Documentation
- Does this pull request introduce a new feature? no
- If yes, how is the feature documented? not applicable
---
##### Was generative AI tooling used to co-author this PR?
- [X] Yes (Claude Code with Claude Opus 4.8)
Generated-by: Claude Opus 4.8 (1M context)
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
