Hyukjin Kwon created SPARK-57959:
------------------------------------
Summary: Deflake MetricsFailureInjectionSuite non-deterministic
stage block failure injection test
Key: SPARK-57959
URL: https://issues.apache.org/jira/browse/SPARK-57959
Project: Spark
Issue Type: Test
Components: SQL
Affects Versions: 5.0.0
Reporter: Hyukjin Kwon
MetricsFailureInjectionSuite 'Non-deterministic stage block failure injection -
injectFailure=true' is flaky (~1/6 in the full suite; passes 10/10 in
isolation). Evidence from macOS diagnostics: the failing run shows stage1=300
stage2=5 (a clean run, no recompute) and completes in ~370ms vs ~860ms on a
passing run, and the test's own shuffle receives no injected FetchFailed. The
INJECT_SHUFFLE_FETCH_FAILURES machinery corrupts mapper-0 of the shuffle map
stage, but whether the downstream reducer observes the FetchFailed (forcing the
stage-1 recompute that inflates the raw metric) depends on task scheduling in
the shared SparkContext, and it occasionally does not fire. Fix (test-only):
re-run the query, resetting metrics each attempt, until the injection forces a
recompute (stage1 > 300), up to 10 attempts, then run the original assertions.
Does not touch the shared DAGScheduler injection machinery used by other
suites. Validated 8x full-suite green on macos-15.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
