The unreleased native-restore fixes after RELEASE.2026-10-26T00-00-00Z
let the lock decide who owns a restore. Since fa42a0982 a single-pool
worker holds the object lock for the whole tier download, and since
7f38bf628 a busy lock means "a restore runs". 7f38bf628 also keeps the
restore's data in a marked multipart upload. Two regressions followed:
- While an old version restored, every GET, HEAD and write of the same
key waited for the download. On a single pool this is new; with
several pools the released code already held the lock that long.
- An ordinary upload whose completion failed after staging its metadata
became InvalidUploadID for every caller, and could not be aborted for
an hour. The marker that hid restore uploads could not tell them apart
once a completion had rewritten the upload's xl.meta.
A restore now runs in two phases:
- The worker erasure-codes the tier copy into the tmp volume of the
version's set, without the object lock, as heal does. A drive that
fails any part, or misses a part when CheckParts runs before the
commit, takes no part in the commit.
- A restore lock per version tells whether a restore runs. POST takes it
before the object lock and hands it to the worker. It is a liveness
hint: only the checks under the object lock decide what is written.
- The commit takes the object lock briefly, on one pool or several. It
reads the version again and checks the attempt's ownership, the
version identity, the tier copy, the part layout and the erasure
layout. It then publishes with RenameData drive by drive, never
through erasure renameData's undo, and needs the write quorum. The
replaced data directory goes only when every published drive reports
the same one.
- A publication that reached some drives but not the write quorum
converges under the same lock hold, in bounded retries after it: back
to the unrestored version on every drive that holds it, or to a
committed restore when the copy turns out to be on a write quorum.
Each pass first checks that every drive holds what the version held
before the commit, the commit wrote, or a convergence wrote, and stops
on anything else. It deletes no directory.
Multipart goes back to main: erasure-multipart.go and
bucket-object-lock.go are identical to 69fcf50cd. The commit lock, the
marker and the restore exceptions of multipart completion and of the
null-version Object Lock check are gone, and single-pool completion
again takes the upload lock first and the object lock right before it
commits. With several pools, abandoned-data cleanup now takes the pools
lock that writes take, as heal does, so it cannot remove a directory
that a commit moved in place and has not yet referenced. The restore
capability becomes protocol 2, and a queued worker checks it again.
Recovery after a coordinator crash during publication, and storage RPCs
that complete after their caller returned, stay excluded. Main has the
same exposure, and they are tracked separately.
Tests:
- New regression tests cover both problems. They fail on f85d4a5fa,
and the multipart one also passes on main.
- Fault-injection tests cover a publication below the write quorum,
including seven of twelve drives on 8+4; lost answers; convergence
retries and stops; the old data directory; a stale inline marker; a
recreated tmp volume; a published revision read across its expiry; a
second POST; a target deleted during the download; a tier 404; and
pools-locked abandoned cleanup.
- Thirteen tests of the upload marker from 7f38bf628 are dropped, and
so is TestNativeRestoreConcurrentCompletion from fa42a0982, which
completed the restore's upload from a client. Four tests are reframed
for the new semantics, including the lock-domain restore test, which
now gates the commit.
- Three self-recursive WriteMetadata test doubles are fixed, and the
helpers of the dropped tests are removed.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Feng Ruohang <rh@vonng.com>
In a single pool, DELETE in every form, bulk delete, the CopyObject
destination, SelectObjectContent, batch key rotation and the upload-ID
guards of complete, abort, get-info and list-parts locked through the
first erasure set, while PUT, multipart commit, PutObjectPart, tag,
retention, legal-hold and metadata updates, restore, transition and heal
lock in the set the object hashes to. When the sets of a pool use
different hosts, as every automatic layout of 18 or more hosts does, the
two locks ask different lockers and do not exclude each other, and
concurrent read-modify-writes of xl.meta can lose acknowledged versions
and delete markers or leave an object on a minority of drives.
Route these locks through NewObjectNSLock, which returns the lock of
the object's own set in a single pool and the pools lock otherwise.
Upload guards choose the set by their object and keep the name
object/uploadID. True prefix deletion and the locks of names that are
not objects keep their domain.
Bulk delete groups its names by set and takes one set's lock at a time,
in ascending set index; a failed group fails alone. A single node, whose
sets share one local lock table, keeps its one bulk lock.
The COPY handler takes its source and destination locks before it opens
the source, sorted, and transactionally: bounded waits while holding a
lock, release in reverse order and retry with new lock instances, one
deadline. It reads the source under them, passes NoLock to every nested
call, and runs that work in the locks' context. This also removes a
deadlock of copies in opposite directions wherever both locks were in
one domain, such as a single node or several pools.
Lock names and RPCs are unchanged, but a pool is protected only when
every node runs this change: affected pools must be upgraded and rolled
back with a drained restart.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Feng Ruohang <rh@vonng.com>
A native restore (POST ?restore, issue #281) marks the version ongoing,
and a background worker rehydrates it from the tier. The worker, its
upload and every status write could be reached or confused by other
callers. A public client could write into or complete the restore
upload. POST copied a whole metadata snapshot over newer protection
changes. An old or failed attempt could act on a newer one. Restore,
expiry and heal decided on whichever drive's metadata they read.
Each POST that starts a rehydration now writes an attempt ID into the
version's internal metadata, with the identity it resolved: version
ID, ETag, size and ModTime. The worker, its upload, its part writes and
its completion carry the attempt. A worker whose attempt no longer
owns the version stops without writing, completing or resetting. Every
write uses the versioning state and explicit version ID captured by
POST, so a null target stays null.
The restore upload carries an internal marker, decided over all drives.
Public multipart calls answer NoSuchUpload and listings hide it. An
unknown marker fails closed. A completion holds a commit lock from
staging its commit in the upload until the commit ends. Stale cleanup
decides on every upload past the stale age under that lock, and keeps
it until its rename ends, also past the drive deadline. Abort of a
committing upload takes the lock and removes a failed commit's
leftover only when every copy is staged and older than one hour.
POST, the worker, the reset, the ExpireRestored recheck and heal all
decide on the authoritative revision of the exact version. That is the
revision held by max(n/2+1, read quorum) drives, identified by the
version header and the normalized metadata digest. Status writes patch
only the restore fields of that revision. They write only the drives
that hold it, and put the old revision back on failure. Heal copies
restore fields only when a quorum holds the same fields.
A failed attempt resets itself with bounded in-memory retries. A later
POST takes over a stranded marker when no attempt runs. Otherwise it
answers RestoreAlreadyInProgress within a few seconds.
Older nodes do not know the marker. A new grid capability handler lets
the serving node ask every peer. POST answers 503 ServiceUnavailable
and writes nothing until all peers support native restore, and a
queued worker asks again before it creates its upload.
The single-pool lock-domain split between DELETE/CopyObject and restore
is closed by its own fix, which this change depends on.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Feng Ruohang <rh@vonng.com>
The merged all-pool check from #279 decided on the versions that a read
quorum of each set merged non-strictly. A drive that missed one metadata
update could therefore supply the revision used for the decision, and a
partially unreadable set could drop a committed version from the history.
Both let an expiry delete data the rule did not cover, including a version
under legal hold.
- objectVersions accepts a revision only when max(n/2+1, the version's
metadata read quorum) drives hold it, identifies revisions by the header
plus a sha256 over a canonical, length-prefixed encoding of the
normalized metadata, filters non-accepted revisions before the strict
merge, and fails closed on omitted versions, too many unreadable drives,
or cancellation.
- Lock fields are read with the evaluator's key precedence. Cross-pool
copies that disagree on hold or retention must be strictly ordered by
valid timestamps; copies with different ModTime are separate writes and
an older protected write refuses the object. The set safeguard re-checks
protection and pending purges on its fresh read for checked version
deletes.
- Duplicate copies with semantically different replication or purge state
refuse the object.
- Batch expiry announces and audits only completed or already-absent
removals.
Reviewed adversarially over seven rounds; the final round approved with no
remaining findings in scope. All-versions expiry through deletePrefix and
tiered delete-all dispatch are tracked separately.
Signed-off-by: Feng Ruohang <rh@vonng.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Reverse the first-parent diff of a3df317ae0,
including the feature branch compatibility and mover follow-up fixes.
Retain the independent multi-pool correctness fixes from #178 and migrate
their shared test fixture away from access-tier code.
Tolerate retired ILM keys and XML, read old v9 statistics while writing v8,
and document migration without moving objects or rewriting their metadata.
Include regression coverage using a historical scanner/writer v9 fixture.
Signed-off-by: Feng Ruohang <rh@vonng.com>
Issue #120 routes an existing SSE-C replica through PutObjectHandler and
NewMultipartUploadHandler. On main those handlers assigned the incoming retention
and legal hold directly, without the source-timestamp ordering #111 added to
CopyObjectHandler and without persisting the ordering timestamps, so in
active-active replication a retransmit carrying an older value could overwrite a
destination version's newer lock state.
Share #111's ordering decision as applyReplicatedObjectLock in
cmd/bucket-object-lock.go and call it from CopyObject, PUT and multipart
initiation. A request that is not an actual trusted replica keeps ordinary write
semantics (a validated value is applied and stamped now); only a real replica
update is ordered against the stored version, so a marker-only peer write no
longer drops a validated hold or default retention. CopyObject keeps its SSE-C
key-rotation encMetadata reconciliation inline. putReplicationOpts now emits a
stored retention ordering timestamp even when the value keys are absent, so a
removal recorded on the retransmit PUT path still replicates onward. replicateAll
marks Failed and carries the error when putReplicationOpts fails.
The handler decision is made against the version as it stands then, which a
concurrent lock update can outrun before the write commits, and for multipart
across the whole initiation-to-completion span. Close that window under the
object write lock the receiving erasure set holds: a trusted SSE-C replica full
write sets ObjectOptions.ReplicaLockReconcile, and erasureObjects.PutObject and
CompleteMultipartUpload re-run the ordering (reconcileStoredObjectLock, which
orders retention and legal hold independently by their reserved timestamps)
against the destination version read on that set before committing. Persisted
upload metadata records the null version as an empty VersionID, so completion
looks that up as the null version rather than the latest. The reconcile runs only
against an existing version; a not-found destination keeps the write's own
accepted lock, including a pre-upgrade upload that persisted values without
ordering timestamps, and a non-not-found read error fails the write. Scoped to
the SSE-C paths this issue enables; CopyObject is left as #111 wrote it.
Scope: this orders Object Lock against the destination version under the write
lock and is correct for a single erasure set. A multi-pool deployment -- where a
version can have duplicate copies across pools, object ModTime ties do not track
per-field lock timestamps, and the object namespace lock is per-pool -- needs a
cross-pool lock-safe reconcile and is deliberately out of scope here, tracked in
pgsty/silo#TBD-multipool-lock.
Tests: TestAPISSECReplicaRetransmitObjectLockOrdering and its multipart sibling;
TestAPIReplicaMultipartNewerHoldSurvivesCompletion and
TestReplicaPutObjectLockReconcileUnderWriteLock (a hold or retention reaching the
version after the handler decision, or after multipart initiation, survives the
commit; a pre-upgrade upload on an absent version keeps its lock);
TestReplicaLockReconcileNullVersion (a null-version completion reconciles the null
version, not a coexisting UUID version, and an absent null version keeps its
accepted lock); TestAPIReplicaMarkerOnlyAppliesObjectLock; TestReplicaStoredLock;
the timestamp-only putReplicationOpts round trip; and the retransmit, exemption
and target-head tests. The #111 CopyObject replica suite and the existing #120
suite stay green, as do the PUT/multipart handler and object-layer regression
suites. Compatibility: the shared helper preserves #111's CopyObject behavior; a
non-replica PUT or multipart initiation that sets Object Lock now also stamps the
reserved ordering timestamp, matching CopyObject since #111; only trusted SSE-C
replica writes take the in-lock reconcile.
Refs pgsty/silo#120
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>
Return AWS-compatible errors for CompleteMultipartUpload checksum failures without changing the global streaming checksum mapping. Compare explicit multipart checksum types symmetrically, distinguish missing composite part checksums, and preserve the CRC64NVME canonicalization pending a direct AWS probe.
Signed-off-by: Feng Ruohang <rh@vonng.com>
Keep NVMe/HDD pool pairs useful without remote tiering: promote objects that
are read often, demote previously moved objects once they go idle, and leave
the feature off until operators set a two-pool topology.
Signed-off-by: mr javad seydi <seydi.birjand@gmail.com>
Use Walk(), which is a recursive listing with versioning, to check if
the bucket has some objects before being removed. This is beneficial
because the bucket can contain multiple dangling objects in multiple
drives.
Also, this will prevent a bug where a bucket is deleted in a deployment
that has many erasure sets but the bucket contains one or few objects
not spread to enough erasure sets.
Rebalance-stop can race with ongoing rebalance operations. This change
prevents these operations from overwriting objects by checking the source
and destination pool indices are different.
Currently, bucket metadata is being loaded serially inside ListBuckets
Objet API. Fix that by loading the bucket metadata as the number of
erasure sets * 10, which is a good approximation.
skip healing properly in scanner when drive is hotplugged
due to how the state is passed around the SkipHealing
might not be the true state() of the system always, causing
a situation where we might healing from the scanner on the
same drive which is being. Due to this competing heals get
triggered that slow each other down.
If used, 'opts.Marker` will cause many missed entries since results are returned
unsorted, and pools are serialized.
Switch to fully concurrent listing and merging across pools to return sorted entries.
This reverts commit 928c0181bf.
This change was not correct, reverting.
We track 3 states with the ProxyRequest header - if replication process wants
to know if object is already replicated with a HEAD, it shouldn't proxy back
- Poorna
there can be a sudden spike in tiny allocations,
due to too much auditing being done, also don't hang
on the
```
h.logCh <- entry
```
after initializing workers if you do not have a way to
dequeue for some reason.
- Use a shared worker pool for all ILM expiry tasks
- Free version cleanup executes in a separate goroutine
- Add a free version only if removing the remote object fails
- Add ILM expiry metrics to the node namespace
- Move tier journal tasks to expiryState
- Remove unused on-disk journal for tiered objects pending deletion
- Distribute expiry tasks across workers such that the expiry of versions of
the same object serialized
- Ability to resize worker pool without server restart
- Make scaling down of expiryState workers' concurrency safe; Thanks
@klauspost
- Add error logs when expiryState and transition state are not
initialized (yet)
* metrics: Add missed tier journal entry tasks
* Initialize the ILM worker pool after the object layer
This allows batch replication to basically do not
attempt to copy objects that do not have read quorum.
This PR also allows walk() to provide custom
values for quorum under batch replication, and
key rotation.
Optionally allows customers to enable
- Enable an external cache to catch GET/HEAD responses
- Enable skipping disks that are slow to respond in GET/HEAD
when we have already achieved a quorum
AccountInfo is quite frequently called by the Console UI
login attempts, when many users are logging in it is important
that we provide them with better responsiveness.
- ListBuckets information is cached every second
- Bucket usage info is cached for up to 10 seconds
- Prefix usage (optional) info is cached for up to 10 secs
Failure to update after cache expiration, would still
allow login which would end up providing information
previously cached.
This allows for seamless responsiveness for the Console UI
logins, and overall responsiveness on a heavily loaded
system.
* optimize deletePrefix, use direct set location via object name
instead of fanning out the calls for an object force delete
we can assume the set location and not do fan-out calls
* Apply suggestions from code review
Co-authored-by: Krishnan Parthasarathi <krisis@users.noreply.github.com>
---------
Co-authored-by: Krishnan Parthasarathi <krisis@users.noreply.github.com>
this randomness is needed to avoid scanning
the same buckets across different erasure sets,
in the same order.
allow random buckets to be scanned instead
allowing a wider spread of ILM, replication
checks.
Additionally do not loop over twice to fill
the channel, fill the channel regardless of
having bucket new or old.
Optimize DeleteObject API to avoid extra
GetObjectInfo call on the replicating side.
For receiving side, it is just a regular
DeleteObject call.
Bonus: Fix a corner case where version purged is
absent on target (either due to replication not yet
complete or target version already deleted in a
one-way replication or when replication was disabled).
In such cases, mark version purge complete.
This PR also returns the replication status in
proxy calls and defers replication attempt if
HEAD on object version returned a error different
from NoSuchKey
Removes the bloom filter since it has so limited usability, often gets saturated anyway and adds a bunch of complexity to the scanner.
Also removes a tiny bit of CPU by each write operation.
Feng Ruohang