fix(ci): retry Maven Central downloads and restore E2E cache - #7781

Fixes #7767

The final approach is retry + restore + conditional save, supported by controlled fork benchmarks and a read-only analysis of the official CI environment.

Problem

Maven Central transfer failures abort CI builds without a workflow-level retry. The E2E image-build job also saves the Maven repository cache but does not restore it before building.

Impact

Dependency-download failures stop build jobs and fail their aggregate checks, blocking PRs before the affected validation completes.

A seven-day review of 198 official failed-job logs confirmed at least 17 Maven Central reset/502 terminal failures, across 14 commit SHAs and three dates. This is a recurring CI reliability problem rather than an isolated failed run.

Changes

• Add bounded retries for eligible Maven Central connection resets and HTTP 429/502 responses, including the observed Wagon Transfer failed for … 502 Bad Gateway format. Allow at most three executions, with 10-second and 20-second delays.
• Retry only when all classified terminal Maven failures are eligible Central transfer errors. Compilation, test, HTTP 404, missing-artifact, other-repository, and mixed failures are not retried.
• Apply the helper to PR builds, integration build/test steps, E2E image/example builds, the SQL list-contract check, and the MySQL, WebSocket, and Spring Cloud E2E Maven invocations.
• Retry only the Maven command, without restarting the surrounding Compose lifecycle.
• Restore the Maven repository before the E2E image build, retaining the existing cache key, path, and restore prefix.
• Use conditional saving:
  • Exact hit → skip save.
  • Fallback or miss → save the current exact key after a successful build.
• Preserve the existing Maven artifact and Docker image delivery paths.
• Add focused regression coverage, including a fixture derived from the observed Wagon 502 failure.

Benchmark results

The fork benchmark completed 12 serial runs across four configurations, plus seven concurrent runs. Every valid run used the same source revision and root Maven/Docker workload. Each serial configuration had three samples.

The build mean decreased by 39.0 seconds (24.5%). Accounting for restore time gives approximately 32.7 seconds of net build-plus-restore savings.

Exact-hit save attempts spent 2–5 seconds performing archive/reservation work, with zero new cache-content upload. Skipping these attempts removes redundant work while retaining cache publication after fallback hits and misses.

Resource findings

These measurements support restoring dependencies before building, skipping redundant exact-hit saves, and retaining miss/fallback saves for downstream reuse.

Validation

• All 21 local retry behavior cases passed.
• The existing storage Compose script test passed for all five runners.
• Bash syntax checks and Git whitespace checks passed.
• Offline replay against 198 official failed-job logs recognized 15 of the 17 target Central-transfer failures. The mixed spring-release 401 failure and the unsupported registry build-extension failure remained rejected. All 181 other scanned failures remained rejected.

The fork benchmark measured warm-cache performance; it did not naturally encounter reset/429/502 failures. Real transient-recovery success rates were therefore not measured. The subsequent conditional-save and bare-502 refinements were verified separately through local tests, source checks, and historical-log replay.

Detailed measurements and supporting evidence:

• Controlled fork benchmark
• Official environment analysis

Non-blocking note

Artifact expiration during partial reruns and unrelated test failures are outside this PR’s scope.

Unsupported registry build-extension failures also remain outside this change.

Checklist

Make sure that:

• You have read the contribution guidelines.
• You submit test cases (unit or integration tests) that back your changes.
• Your local test passed ./mvnw clean install -Dmaven.javadoc.skip=true.