The GitHub Actions job "Flink CDC CI" on 
flink-cdc.git/FLINK-36605_debezium_bump has failed.
Run started by GitHub user eskabetxe (triggered by eskabetxe).

Head commit for run:
0fd133606ab808390e49b22cdf73233e459c934f / Joao Boto <[email protected]>
[hotfix][ci] TEMPORARY: split shards per connector, report every module, bound 
hung forks

DO NOT MERGE - diagnostic aid for the Debezium 2.x bump, revert before merging.

Three changes, all aimed at making it visible which connector actually passes.

1. Split the bundled shards. "Source Unit Tests" and its 2.x twin go from 4 
shards
   to one job per database (mysql-source, postgres-source, oceanbase-source, 
tidb,
   vitess, oracle, sqlserver, db2, mongodb). Today a red
   "postgres-source, oceanbase-source, tidb, vitess" job means Vitess hung while
   Postgres and TiDB passed, which is impossible to tell from the job name.
   For the pipeline shards only the three connectors this bump touches
   (mysql-pipeline, postgres-pipeline, oracle-pipeline) are pulled out into 
their
   own jobs; the ten unrelated connectors stay grouped. That isolates the 
affected
   ones and takes their runtime out of a shard that exceeded the 90 minute 
budget.

2. Pass -fae to the test invocation. Maven fails fast, so the first failing 
module
   aborts the rest of its shard: the "oracle, sqlserver" job stopped after
   flink-connector-oracle-cdc (module 1 of 4) and "db2, mongodb" stopped after
   flink-connector-db2-cdc, leaving SqlServer and MongoDB completely 
unexercised.
   Each job still builds two modules, so this matters even after the split.

3. Bound a hung forked test JVM, via a new surefire.fork.timeout property so the
   limit can differ per module. A single deadlocked test otherwise consumes the
   whole job budget - the Vitess connector hangs in a gRPC call inside 
Debezium's
   connection validation and took Postgres, OceanBase and TiDB down with it in
   both the Flink 1.x and 2.x shards.

   Connector modules default to 30 minutes. Their integration-tests execution 
sets
   reuseForks=false, so there the limit applies per ITCase, roughly 2x headroom
   over the slowest class observed (postgres NewlyAddedTableITCase, ~14 
minutes).

   The two e2e modules override it to 90 minutes. Their end-to-end-tests 
execution
   sets forkCount=1 but inherits reuseForks=true, so one fork runs the entire
   suite and the limit applies to that whole run, not to a single test: on 
master
   the pipeline e2e module legitimately takes ~82 minutes. The step timeout goes
   from 90 to 110 minutes so surefire can hit its own limit first and name the
   class that stalled, instead of the job being killed with no test report.

Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>

Report URL: https://github.com/apache/flink-cdc/actions/runs/33372963286

With regards,
GitHub Actions via GitBox

Reply via email to