The GitHub Actions job "Flink CDC CI" on flink-cdc.git/FLINK-40647 has failed.
Run started by GitHub user lvyanquan (triggered by lvyanquan).

Head commit for run:
fe288eec0fa523d23dce5e93cc44ad40a57d77ff / lvyanquan 
<[email protected]>
[FLINK-40647][runtime] Scope schema expansion to streaming mode and harden its 
failure handling

Restrict the new existing-table schema expansion option to streaming
pipelines and fix several correctness and test-harness issues found while
validating it end to end.

* Batch mode: revert the BatchSchemaOperator wiring added by this feature,
  so batch pipelines keep the pre-existing sink behavior. The translator now
  warns when the option is set in batch mode, and the docs state that
  expansion is streaming only.
* Rename the default mode OFF to DISABLED. A bare `OFF` YAML scalar is
  parsed as the boolean `false`, so the option now rejects non-textual
  values instead of coercing them, which also drops the parser workaround.
* Fail fast when a sink lacks ExistingTableSchemaExpansionSupport instead of
  degrading to the sink's own handling under TRY_EXPAND, which previously
  hid a connector misconfiguration.
* Split TRY_EXPAND into a probing phase (delegating to the sink when the
  table cannot be expanded safely) and an applying phase (propagating
  failures). Before this change a failure while applying or verifying
  derived DDL fell back to applying the original CreateTableEvent to an
  already existing table, silently dropping the missing columns.
* Centralize the lazy Paimon catalog creation in getCatalog() and reset the
  cached instance on close so the applier stays reusable.
* E2E: the SQL client creates Paimon artifacts as the JobManager user while
  the sink runs in the TaskManager container, so widening the shared volume
  permissions is required for applyCreateTable/applyAddColumn to work. The
  Fluss source table insert additionally needs the Fluss jar shipped through
  pipeline.jars, because copyJarToFlinkLib only reaches the JobManager and
  sql-client.sh has no --jar option; the read-back loop now surfaces a
  failed insert job instead of timing out silently.
* Drop the batch expansion E2E case, whose logic is shared with the
  streaming cases and already covered by unit tests.

Co-Authored-By: Codex <[email protected]>
AI-Model: gpt-5.6-sol
AI-Contributed/Feature: 314/330
AI-Contributed/UT: 367/374

Report URL: https://github.com/apache/flink-cdc/actions/runs/35748636955

With regards,
GitHub Actions via GitBox

Reply via email to