Purushottam Sinha created FLINK-40406:
-----------------------------------------
Summary: CI: Maven build has no retry on transient Central
failures, causing spurious PR failures
Key: FLINK-40406
URL: https://issues.apache.org/jira/browse/FLINK-40406
Project: Flink
Issue Type: Improvement
Components: Build System / CI, Kubernetes Operator
Reporter: Purushottam Sinha
Problem
CI jobs fail outright on a single transient error from Maven Central (e.g. 502
Bad Gateway), with no retry, causing PRs to show a wall of failed/cancelled
checks for an infrastructure blip unrelated to the change under review.
Evidence
- Job run: Alternative namespace tests (v1_20, standalone, ...) failed with
Could not transfer artifact org.codehaus.plexus:plexus-utils:pom:3.4.1 from/to
central (https://repo.maven.apache.org/maven2): transfer failed ..., status:
502 Bad Gateway
- Dockerfile:28 runs mvn -ntp clean install ... with no retry flags; ci.yml's
three mvn -B clean install / mvn -B verify calls likewise have none
- ci.yml's concurrency: cancel-in-progress: true amplifies the appearance of
the failure: a retrigger cancels in-flight matrix jobs, making one real flake
look like ~10 failures
- No settings.xml or retry precedent exists in this repo or in apache/flink's CI
Proposed fix
- Add -Dmaven.wagon.http.retryHandler.count=3 to the mvn invocation in
Dockerfile:28 and the three mvn calls in ci.yml
- Optionally add a fallback mirror (e.g. maven-central.storage.googleapis.com,
verified reachable) as defense-in-depth
Acceptance
- A simulated/observed transient Central 502 no longer fails the build on first
occurrence
- No new third-party GitHub Actions introduced
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
