Martijn Visser created FLINK-40579:
--------------------------------------
Summary: Weekly connector CI cancels all other branch lines on the
first failure and exceeds the ASF 20-concurrent-jobs policy
Key: FLINK-40579
URL: https://issues.apache.org/jira/browse/FLINK-40579
Project: Flink
Issue Type: Bug
Components: Connectors / Kafka
Affects Versions: kafka-5.1.0
Reporter: Martijn Visser
Assignee: Martijn Visser
The weekly workflow (.github/workflows/weekly.yml) fans out over five (Flink
version, connector
branch) lines x three JDKs = 30 jobs. Two things make it useless as a signal
today:
1. Both matrix levels use the GitHub default fail-fast=true. The outer
flink_branches matrix in
weekly.yml cancels every other branch line ~30 s after the first failing
line; the inner JDK
matrix in flink-connector-shared-utils ci.yml/python_ci.yml cancels the
sibling JDKs ~15 s after
the first failing JDK. Every weekly run since at least October 2025
therefore shows exactly one
Java failure, one Python failure and ~28 cancelled jobs, e.g. run
34002024546 (2026-09-06).
Because all 30 jobs start within 2 s and a Java flake surfaces at minute
18-35 of a ~31-minute
job, the cancelled jobs have already consumed 90-95 % of their runtime, so
fail-fast saves
almost no minutes while discarding up to 28 of 30 results. Run 27484280508
(2026-06-14) shows the
Java matrix fully green while only the Python line made the run red.
2. The workflow starts 30 concurrent jobs.
https://infra.apache.org/github-actions-policy.html
says a workflow MUST have a job concurrency level <= 20 across all matrices
and SHOULD be <= 15.
3. The pinned Flink patch versions 1.20.4 and 2.1.2 are superseded by 1.20.5
and 2.1.3.
Proposed change (weekly.yml only; the inner per-JDK fail-fast in shared-utils
is unchanged):
- strategy.fail-fast: false on the outer matrix of compile_and_test and
python_test.
- Release lines run one JDK (v5.0/2.2.1 x 17, v5.0/2.1.3 x 17, v4.0/2.0.2 x 17,
v3.4/1.20.5 x 11);
main/2.2-SNAPSHOT keeps 11/17/21 because push_pr.yml only covers main with
2.2.1. Python runs one
JDK per line. 12 concurrent jobs instead of 30, ~280 instead of ~600 runner
minutes per run.
None of the failures observed in the last 14 weekly runs was JDK dependent.
- Bump 1.20.4 -> 1.20.5 and 2.1.2 -> 2.1.3.
- timeout_test: 45 for the Java job (green jobs finish in at most ~36 minutes).
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
