This report explains why #7781 uses retry + restore + conditional save. Official CI logs establish the download failures and cache reuse opportunities; the fork benchmark measures build time and runner disk usage.
A seven-day review of 198 failed-job logs found at least 17 Maven Central transfer failures: 5 connection resets and 12 HTTP 502 responses, across 14 commit SHAs and three dates.
| Workflow | Confirmed failures |
|---|---|
| integrated-test | 7 |
| e2e-k8s | 6 |
| ci / registry_dependencies | 3 |
| native-sql-matrix | 1 |
The final helper was replayed offline against all 198 logs. It accepted 15 of the 17 Central-transfer failures and rejected all 181 other failures. The two remaining Central cases contained a mixed spring-release 401 error or an unsupported registry build-extension error.
The review also identified three E2E Maven calls with real transfer failures: MySQL storage, WebSocket, and Spring Cloud. These calls are now wrapped individually; the surrounding Compose startup, logging, and cleanup remain in place. Registry dependency checks remain outside this PR's retry coverage. Classification replay does not measure recovery after an actual retry.
Evidence: failure records and job links, scan coverage, final classifier replay.
At the audited master revision, the image builder ran build → save, without requesting restore. All ten sampled builders downloaded 2,912–2,943 Central entries and took 139–194 seconds for the build step.
Existing official jobs demonstrate reuse before a build:
| Job | Requested → restored key suffix | Result | Restored bytes |
|---|---|---|---|
| PR7793 CI | 8a2c41e7 → 8a2c41e7 |
exact | 776,804,688 |
| PR7673 CI | 2f4add9e → 8a2c41e7 |
fallback | 776,804,688 |
PR7793 was opened at 01:23:00 UTC. Its CI restored the exact key at 01:23:23, before its own E2E builder saved a PR-local entry at 01:26:23. The restored size matches the master entry created the previous day. Scope rules, timing, and size point to inheritance from master; the action log does not print the source entry ID.
The ten sampled downstream cases had 9 exact hits and 1 miss. Four additional pre-build CI jobs had 2 exact hits and 2 fallbacks. These are selected samples, not a repository-wide hit-rate estimate. The new builder restore position has no historical hit rate because the existing builder did not request restore.
Official sampled restores ranged from approximately 0.77–3.22 GB and 6–40 seconds. The fork's 6.3-second restore therefore describes its own 740.79 MiB archive. Downstream cases still downloaded additional dependencies after exact hits.
Evidence: E2E sample metrics, pre-build CI metrics, sample selection.
The inventory snapshot returned 11 Maven cache entries totaling 10,505,273,201 bytes (9.78 GiB), across 8 refs and 6 full keys.
The following entries have the same Linux-maven-8a2c41e7… key and the same cache version, but occupy separate storage:
| Entry ID | Ref | Bytes |
|---|---|---|
| 9094733854 | master | 776,804,688 |
| 9106318400 | PR7793 merge | 776,690,893 |
| 9106264157 | PR7792 merge | 776,725,406 |
| 9105358453 | PR7785 merge | 776,720,680 |
| 9105347180 | PR7789 merge | 776,596,704 |
| 9105084294 | PR7788 merge | 776,640,438 |
The five PR copies consume 3,883,374,121 bytes (about 3.88 GB) beyond the master copy. PRs can restore master/base caches; a PR-created cache is scoped to its merge ref and is not shared with other PRs. See GitHub cache scope rules.
The fork measured 2–5 seconds for redundant exact-key save attempts, with zero new cache-content upload. This supports the existing save step's new condition:
| Restore result | cache-hit output |
Save after successful build |
|---|---|---|
| Exact hit | true |
Skip |
| Prefix fallback | false |
Publish the current exact key |
| Miss | empty/unset | Publish the current exact key |
The cache action source confirms these output values. Fallback and miss still create the new key; an inherited exact hit avoids another PR-local copy.
Mainline POM changes produced 15, 32, and 42 hash transitions over 7, 14, and 30 days. Keeping fallback restore and publication of new exact keys accommodates those changes. These transition counts are not cache-upload counts.
Evidence: full inventory, POM hash history.
| Resource | Finding |
|---|---|
| Dependency downloads | Repeated cold downloads are an observed avoidable cost and the source of the reported reset/502 failures. |
| Runner disk | The fork's minimum sampled free disk was 80.8 GiB, including cache restore and both image exports. Local disk exhaustion was not a bottleneck in that workload. |
| Cache storage | Cross-ref duplicate storage is confirmed. The configured quota/budget endpoint returned HTTP 402, so the 9.78 GiB snapshot does not establish quota exhaustion or eviction. |
| Cache save work | Exact-hit archive/reservation work took 2–5 seconds without new upload. Conditional saving removes that repeated work. |
| Current build outputs | Cases receive current ShenYu Maven artifacts and Docker images through artifacts. Storage builds its independent E2E reactor and uses current images. Skipping an exact-hit dependency-cache save preserves these delivery paths. |
The usage endpoint reported the same byte total as inventory but a count of 10 instead of 11. Both responses are retained; inventory statistics use the 11 returned entries.
- Historical window: seven days ending 2026-10-18 02:45:44 UTC / 10:45:44 Beijing time. Inventory is a snapshot collected afterward.
- Official workflow source: master
f00c1571. Final helper results refer to PR commit5a2f0470f; its helper is unchanged from the locally replayed revision. - Failure scanning covered 198 jobs from failed runs and reruns. Cancelled-run child failures and older-created runs were not exhaustively scanned, so 17 is a lower bound.
- The final implementation passed 21 local retry behavior cases, the existing five-runner storage Compose script test, shell syntax checks, and Git whitespace checks. No complete Docker E2E result is claimed by this report.
- Additional source data: workflow cache map, window counts, storage-limit response, and fork benchmark.