Sylwester Lachiewicz created FLINK-40511:
--------------------------------------------
Summary: Track and speed up slow integration tests
(KafkaSinkITCase and friends)
Key: FLINK-40511
URL: https://issues.apache.org/jira/browse/FLINK-40511
Project: Flink
Issue Type: Improvement
Components: Connectors / Kafka
Reporter: Sylwester Lachiewicz
Background
CI runs under a 50-minute job budget, and the integration-test module dominates
it. FLINK-40139 added a 15-minute per-method JUnit timeout as a safety net, but
the underlying suite is genuinely slow. This issue tracks measuring and
refactoring the slow ITCases so the whole module fits the budget with room to
spare.
Measurements (JDK 21, surefire per-method XML; CI matches local within ~2%)
| Class | Methods | Total |
|---|---|---|
| KafkaSinkITCase | 117 | ~712s |
| KafkaTableITCase | | ~116s |
| DynamicKafkaSourceITTest | | ~111s |
| all other classes | | <120s each |
KafkaSinkITCase is the outlier. Its slowest single methods are the exactly-once
recovery variants (~30–34s each: testRecoveryWithExactlyOnceGuarantee1 at
33.7s, siblings at 32.6/32.5/31.7s, rescaleListing6 at 19.0s) — likely bounded
by checkpoint intervals rather than test logic.
Directions to evaluate
* Split KafkaSinkITCase by feature (exactly-once recovery, rescaling, delivery
guarantees) into smaller classes.
* Reduce per-test broker/container setup cost: KafkaTestBase is all-static with
per-class prepare/shutDownServices across 12 test classes; sharing environments
across classes requires config-keyed shared environments (cluster count, secure
mode) with JVM-shutdown-hook cleanup — beware cross-class group.id/offset
leakage.
* surefire forkCount > 1 to parallelize at class level (experimented with on
branch ci-surefire-forkcount-2).
Not a blocker for FLINK-40139; the 15m default timeout already bounds the worst
case per method.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
