The #105 T3 change (PR #156) tried to keep the resident metadata cache
monotonic by guarding peer-reload publication on lastUpdate(). But
lastUpdate() is the max of per-config timestamps and cannot order whole
records: a node caching {policy@20, CORS@10} that receives a newer CORS@15
still has lastUpdate()==20, so the guard rejects the legitimately-newer
record and the periodic refresh (same comparator) cannot repair it. A
paused reload could also resurrect deleted resident state.
Per the maintainer decision, revert the reload publication to its original
unconditional (acceptable-until-refresh) behavior:
- remove setReloaded and restore the plain Set plus notification/target
registry updates in LoadBucketMetadataHandler;
- restore refreshBucketsMetadataLoop's own lastUpdate() staleness check and
globalEventNotifier.set / globalBucketTargetSys.set publication;
- restore the unconditional GetConfig cache-miss publication;
- document the known freshness limitation at the reload site (the periodic
refresh is best-effort and cannot repair an equal-maximum-timestamp
divergence).
The T1 lifecycle merge-under-lock (UpdateExpiryLCConfig) and both T2 fixes
(DeleteBucket takes metadata.lock before deleting; saveMetadata and
loadBucketMetadataParseUnderLock recheck physical bucket existence) are
kept fully intact.
Tests:
- drop the T3 reproductions (overlapping-reload resident-cache test and the
peer-reload-preserves-current-targets publication test);
- add lockBucketMetadataAcquireHook, a nil-in-production atomic test hook in
the shared metadata.lock path, so tests can deterministically observe a
caller (notably DeleteBucket, whose lock is taken through its
erasureServerPools receiver and is invisible to an injected object layer)
reaching the lock;
- rewrite the T2 delete-race ghost test to hold metadata.lock MID-SAVE (past
saveMetadata's existence recheck) and synchronize on the delete's actual
lock attempt via the hook, so it isolates the lock-before-delete fix:
removing only DeleteBucket's metadata.lock (recheck kept) now fails it;
- rewrite the cancellation test to observe the delete's actual lock attempt,
then cancel and await its error while still holding the lock, so a
scheduling-delayed delete stopped by the canceled context can no longer
pass on a broken tree.
Refs #105. Follows #156.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Signed-off-by: Feng Ruohang <rh@vonng.com>
Audit of the three deferred #105 follow-ups. Each reproduces with a
deterministic red test in cmd/bucket-metadata-race_test.go, and each fix is
the minimal change that turns its test green while preserving the
<bucket>.lck -> metadata.lock -> .metadata.bin lock order established by #103.
1. Lifecycle expiry merge lost update (persistent). PeerBucketLCConfigHandler
and healBucketILMExpiry read the current lifecycle with an unlocked
GetConfigFromDisk, merged the replicated expiry rules with the local
transition rules, then wrote the pre-computed blob via Update. Any lifecycle
transition change committed between the merge read and the merge write was
silently lost. New BucketMetadataSys.UpdateExpiryLCConfig performs the read,
merge, and save under one metadata.lock; mergeExpiryWithLCConfig now takes
the locked snapshot and validates object-lock retention from it instead of
re-reading (avoids a re-entrant metadata load under the lock).
2. DeleteBucket ghost .metadata.bin (persistent). DeleteBucket took only
<bucket>.lck while config writers take only metadata.lock, so a writer that
was mid-save could re-create .metadata.bin after the prefix purge. The purge
now runs under metadata.lock, with a best-effort unlocked fallback so a
delete is never blocked from completing.
3. Overlapping peer reloads publishing a stale resident cache (freshness only;
the persisted record stays correct). LoadBucketMetadataHandler and the
GetConfig cache-miss path published with an unconditional Set, so a reload
that read an older revision could overwrite a newer resident record until the
next refresh. New BucketMetadataSys.setReloaded (and a matching GetConfig
guard) refuses to regress a newer resident record, mirroring
refreshBucketsMetadataLoop.
Verification: go build -tags kqueue,dev ./...; go vet ./cmd; gofmt clean;
rebrand-guard baseline unchanged; go test -tags kqueue,dev ./cmd (207s) green;
new tests plus the #103 metadata suite green under -race.
Refs #105. Parent #102. Foundation #103.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01L7qJqWwy8oFA6aCXWRzXQe
Signed-off-by: Feng Ruohang <rh@vonng.com>
Feng Ruohang