gsadmin/rustfs
1
0
Fork 0
mirror of https://github.com/rustfs/rustfs.git synced 2026-09-06 03:59:14 +00:00
Code Issues Packages Projects Releases Wiki Activity

fix(heal): reject incomplete migration lineage before staging

Browse Source 
Co-Authored-By: heihutu <heihutu@gmail.com>
Co-Authored-By: zhi22915 <qiuzgang@gmail.com>
This commit is contained in:
houseme
2026-09-05 17:03:53 +08:00
parent 408ee1be22
commit d361b3cfac
2 changed files with 330 additions and 55 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docs/testing/mrf-pending-migration.md
+3 -3
View File
@@ -2,13 +2,13 @@
`heal::mrf_queue::snapshot::migration` exposes explicit capture, staging, and readback of pending legacy responsibility evidence. Nothing invokes it from the production MRF consumer. It does not enable the committed-snapshot writer, freeze legacy ingress, acknowledge durable admission, or authorize source garbage collection.
The caller supplies every configured local disk slot, including missing slots. Missing/unformatted/duplicate disks, unavailable metadata volumes, invalid records, empty evidence, and aggregate byte/record overflow fail closed. Both legacy paths retain their original bytes, disk identity, absent-versus-empty state, and SHA-256 digest. Complete subset/superset replicas become conservative pending evidence, never a claimed newest legacy snapshot. Raw record replay preserves kind, scope and nil/absent version encodings; unknown incarnation stays unknown.
The caller supplies every configured local disk slot, including missing slots. Missing/unformatted/duplicate disks, unavailable metadata volumes, invalid records, and aggregate byte/record/source-history overflow fail closed. A valid empty source observation can inherit earlier pending responsibilities, but staging without any current or inherited responsibility is rejected. Both legacy paths retain their original bytes, disk identity, absent-versus-empty state, and SHA-256 digest. Complete subset/superset replicas become conservative pending evidence, never a claimed newest legacy snapshot. Raw record replay preserves kind, scope and nil/absent version encodings; unknown incarnation stays unknown.
Staging writes only `.heal-mrf-import-pending.{0,1}.bin`, `.heal-mrf-import-commit.{0,1}.bin`, and `.heal-mrf-import-claim.bin` under the metadata volume. It reuses the committed reader's manifest codec and the storage owner's conditional-file operation, including the configured metadata durability policy. A candidate is written before sources are revalidated, its manifest is then committed, and committed bytes plus source coverage are read back before success. Success is pending staging evidence, not a power-loss or cluster-quorum durability receipt.
A successor inherits prior source bytes even if a replay consumer has already read or admitted their records. Size limits include inherited evidence. There is no completion-based pruning. A changed source blocks recovery of that pending generation; an explicit new capture can stage a successor that retains both the previous and current responsibilities. The inactive slot is replaced while the preceding committed slot remains intact.
A successor inherits prior source bytes even if a replay consumer has already read or admitted their records. Independent byte, record, and source-history limits include inherited evidence, and source identities use a hash index with full-byte conflict checks. There is no completion-based pruning. A changed source blocks recovery of that pending generation; an explicit new capture can stage a successor that retains both the previous and current responsibilities. The inactive slot is replaced while the preceding committed slot remains intact. Any corrupt, unsupported or over-budget slot blocks selection rather than falling back to an older generation. A payload without a manifest is repairable only when its length and digest match the exact retry candidate or a payload independently validated through another committed replica; retries also repair missing replica manifests.
All participating disks are claimed in disk-identity order through CAS. Normal completion conditionally releases only the current invocation's claim. Cancellation, process death, or an ambiguous claim/release I/O failure may leave a claim behind. Read-only recovery remains available, but further staging is blocked until a separate storage-fenced recovery procedure is implemented. Process liveness and the legacy ingress lease do not authorize taking over or deleting a claim.
All participating disks are claimed in disk-identity order through CAS. Normal completion conditionally releases only the current invocation's claim, attempts every acquired claim even after a release error, and reports the first release error. Cancellation, process death, or an ambiguous claim/release I/O failure may leave a claim behind. Read-only recovery remains available when the snapshot/source proof is valid, but further staging is blocked until a separate storage-fenced recovery procedure is implemented. Process liveness and the legacy ingress lease do not authorize taking over or deleting a claim.
Run the focused fixtures with a nonzero test count: