volume.configure.replication rolls through nodes one at a time, so a
volume legitimately reports the old and new replica placement while it
settles. Layouts differing only in rp within the same collection are a
supported transition, not a reused id.
Dial the advertised grpc port, return mark errors, and keep a
quarantine entry until the writable push succeeds so a released
copy does not stay read-only on the server.
A quarantined copy stayed writable on the volume server itself for a
client holding an existing file id, since the master-only mark never
reached the server's local write path. Quarantine now also calls
VolumeMarkReadonly (can_delete so expiry still drains), is re-pushed if
the server comes back reporting it writable, and VolumeMarkWritable
runs when the conflict resolves and the server had reported writable.
- collect all conflicting layouts, not just the first, so a vid spanning
three layouts does not leave one writable
- mark quarantine on the location list only: folding it into the stored
report made the volume digest permanently disagree with the server and
forced a full volume list every heartbeat
- isAllWritable now honors the location read-only mark so a quarantined
copy stays out of writables through resyncs and re-registration
- re-check the conflict before keeping marks, and do not clear a node's
quarantine while it still holds the vid in another layout
- log the full layout keys on both sides of the conflict
Track the latest server-reported read-only flag while quarantined so a
release restores the server s true state, and run the release check on
collection deletion and data-node removal too, not only explicit volume
unregistration.
Ec-only volumes share the global id space. Compare shard collections
against regular ones while allowing shards and regular copies of the
same encoded volume.
A cross-node volume.move briefly shows the same id on two nodes under
different disk layouts; permanent quarantine would pin a valid target
read-only. Keep the first-reported state per quarantined copy and lift
the marks once the conflicting copy unregisters, so only a persistent
collision stays read-only.
Near-simultaneous registrations let each side quarantine only the other,
so arrival order decided which copy stayed writable. A confirmed
cross-layout collision means either copy could be the impostor, so flag
every colliding copy read-only regardless of registration order.
DeferAllocation only armed after the topology id persisted, so growth
requests arriving during init ran ahead of the warmup window. Gate
NextVolumeId on an init-pending flag spanning the whole mint.
heartbeatResponse now reads MaxVolumeId and TopologyId off ms.Topo, but
the test built a MasterServer without one and panicked, failing the
package for every PR.
Generated with [Devin](https://devin.ai)
Co-authored-by: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* volume balance: unify the per-move convergence guard
Evaluate the post-move utilization comparison through serverUtilization so
the guard also applies when servers lack MaxVolumeCount info (raw-count
mode), where it previously did not run at all.
* volume balance: restrict fallback destinations to improving servers
The score-based fallback inside createBalanceTask could redirect a move to
any valid server, including ones as full as the source — between equally
full servers the move only relocates the imbalance and the next detection
cycle reverses it (#11839). Exclude from eligibleTargets any server whose
post-move utilization would exceed the source's, so a fallback destination
satisfies the same convergence guard as the loop-chosen target.
* volume balance: skip a volume when the chosen target violates placement
When the detection loop's target resolves but cannot host the replica (it
already holds one, or is out of the filtered scope), createBalanceTask now
returns nil so the loop advances to the next volume — matching shell
volume.balance, which continues to the next candidate instead of
redirecting the move to a different server.
The score-based fallback is kept for the case where the target cannot be
resolved at all; its allowed set is already limited to servers the move
would improve.
* volume balance: regression test for replica-constrained targets (#11839)
Reproduces the reported loop: the emptiest server holds the second rp=001
replica of the lowest-ID volumes on the fuller servers, so it is never a
legal target for them. Detection must skip those volumes and send the
movable ones to the under-utilized server, and re-running detection after
applying the plan must never move the same volume twice.
2026-10-19 20:52:30 -07:00
Chris LuGitHubDevin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
The master intentionally sheds an assign that triggered volume growth
once its wait window elapses (503 + Retry-After); on a slow CI box
cold-start growth can outlast it, so the single-shot assertion flaked.
Honor the documented retry contract like a real client, within a 60s
deadline.
Generated with [Devin](https://devin.ai)
Co-authored-by: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Chris LuGitHubDevin <158243242+devin-ai-integration[bot]@users.noreply.github.com>Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* s3: strip internal object-lock and version headers on auth entry
putToFiler persists Seaweed-X-Amz-Object-Lock-Mode, -Retention-Until-Date,
-Legal-Hold and -Version-Id request headers into authoritative entry
metadata without a permission check, so an s3:PutObject-only writer could
pin COMPLIANCE retention and legal hold on its own uploads, bypassing the
dedicated-permission gate on the standard x-amz-object-lock-* headers.
The server only sets these headers from validated inputs, so drop them in
authenticateRequestInternal alongside the identity headers already
scrubbed there.
Generated with [Devin](https://devin.ai)
Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* s3: strip internal SSE headers on auth entry
X-SeaweedFS-SSE-S3-Key-Data, -SSE-S3-Base-IV and -SSE-KMS-Base-IV are read
from the request to drive multipart encryption, so a client could inject
its own key material and IVs. The server sets them itself between handler
stages of a request; drop inbound values in authenticateRequestInternal.
Generated with [Devin](https://devin.ai)
Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* s3: scrub internal headers before the auth-enabled early return
Auth and AuthPostPolicy skip authenticateRequestInternal when IAM is
disabled, so client-supplied internal headers still reached putToFiler.
Hoist the scrub into scrubInternalHeaders and run it unconditionally in
both wrappers as well as the shared auth entry point.
Generated with [Devin](https://devin.ai)
Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* s3: test internal header scrub through the auth-disabled wrapper
Generated with [Devin](https://devin.ai)
Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---------
Co-authored-by: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
2026-10-19 18:35:20 -07:00
Chris LuGitHubDevin <devin-ai-integration[bot]@users.noreply.github.com>
* filer.sync: keep the needle-404 clock from resetting on live volumes
The give-up clock for a source chunk that cannot be produced is kept per
volume so a vacuumed volume's files all inherit one wait. But a needle
404 and a volume lookup failure shared that clock, and sourceServed
cleared it whenever any needle on the same volume read fine, so on a
busy cluster the grace never elapsed for a single deleted needle on a
healthy volume and filer.sync retried it forever (#11841).
Give needle not-found answers their own per-volume clock that serving
siblings does not reset.
* filer.sync: an answered 404 needs no liveness probe before skipping
After the missing-chunk gate gives up, onReplicateChunkError only skips
the entry when a re-read of the most recently served chunk succeeds.
That probe can never have been armed on a fresh sync, and its single
stored file id can itself be vacuumed since — either way the entry stays
pinned and the offset never advances.
A needle 404 was answered by a live volume server, which is itself proof
the source is serving requests, so skip on it directly. Keep the probe
for the lookup-failure flavor, where no server answered and the silence
can still mean a source outage.
* filer.sync: skip a replayed version the source rewrote in place
Supersession was decided by entry mtime alone, so a source entry whose
chunks were swapped without an mtime bump — remote cache eviction and
re-cache, an in-place repair — never read as superseded, and its
vacuumed chunks were retried indefinitely.
When a chunk is missing, compare the replayed chunk's file id against
the live entry's resolved chunk set: absent means the replayed version
is obsolete and its event is skipped on the first failure instead of
after the give-up window. The same live-entry comparison now also
decides the entry-level skip in onCorruptChunk/onReplicateChunkError.
* filer.sync: give each missing needle its own grace clock
A 404 is one needle's answer, but the clock was kept per volume: once the
first needle's wait expired, every other needle on that volume inherited an
already-elapsed clock and could be written off on its first sighting. Track
the wait per file id and clear it only when that needle itself serves again,
so a live volume's healthy needles neither shorten nor restart it.
* filer.sync: keep the probe for filer-proxied 404s
ReadPart proxied through the source filer collapses "no volume locations"
into the same 404 a missing needle returns, so that answer cannot stand in
for the liveness probe the way a direct volume read's can. Only treat 404 as
an answered needle when the sink reads volume servers directly.
* filer.sync: only a not-older live entry can supersede a replay
Chunk-list comparison must respect the same ordering the mtime check does:
a live entry older than the replayed version means the source was rolled
back, so its differing chunks say nothing about whether the replayed data
is still needed. Gate the content check on live mtime >= replayed mtime in
both sourceMovedOn and sourceDroppedChunk.
* filer.sync: compare chunk ids without resolving manifests
resolvedChunkFileIds downloaded manifest files to expand chunk lists, so a
manifest sitting on a dead volume made the supersession check itself fail
and left the offset pinned behind a replica the source cannot read. File
ids are immutable needle identities — the unresolved sets already compare
chunk content exactly, and the check no longer performs I/O.
* filer.sync: judge nested chunks by their top-level chunk
A manifest child is not in the entry's chunk list even when the manifest
still references it, so checking the live entry for the child id reported
it dropped and returned the raw 404 — unwrapped, so the offset still
pinned. Thread the entry-level chunk id down the manifest recursion and
consult it for the dropped-chunk check instead.
* filer.sync: let the manifest-child test exit through the gate
Run the worker with a shortened grace period and assert the returned error
is marked errSourceChunkMissing, instead of racing the worker still running
against a restored RetryWaitTime.
* s3: sort ListMultipartUploads by key and honor key-marker
* s3: skip upload folders without an object key in ListMultipartUploads
* s3: order same-key multipart uploads by initiation time
The upload id suffix is now a time-ordered UUIDv7, so sorting by upload id lists a key's uploads oldest first, like S3. Pagination is unchanged: ordering and the upload-id-marker comparison share the same key.
selectMultipartUploads now builds the whole result, including NextKeyMarker and NextUploadIdMarker, and the pagination test pages with the response markers.
---------
Co-authored-by: Chris Lu <chris.lu@gmail.com>
2026-10-19 18:10:06 -07:00
Chris LuGitHubDevin <devin-ai-integration[bot]@users.noreply.github.com>
The HEAD on the completed multipart object occasionally gets a 403 from a
signature verification mismatch before the object read settles; retry the
download instead of failing the run.
Co-authored-by: Devin <devin-ai-integration[bot]@users.noreply.github.com>
* pb: heartbeat fields for the federated volume-id bound (#11763)
Heartbeat carries max_volume_id_seen and last_seen_topology_id up, and
HeartbeatResponse echoes max_volume_id and topology_id down, so the
allocation bound is replicated to every volume server.
* master: echo the volume-id bound and fold reported bounds in (#11763)
Every heartbeat response carries the leader current max volume id and
topology id; every heartbeat reported max_volume_id_seen is folded into
the topology bound via UpAdjustMaxVolumeId.
* volume: persist a per-dir vid stamp of the highest id ever seen (#11763)
The store keeps max_volume_id_seen + last_seen_topology_id in a
.vid_stamp file per volume dir, fed by heartbeat echoes, AllocateVolume,
and local volume/EC creation, seeded at boot from stamps and local
holdings, and reported in every heartbeat. A rebuilt master recovers
the bound from the first surviving veteran report.
* volume-rust: mirror the vid stamp and heartbeat echo handling (#11763)
Same mechanism as the Go volume server: .vid_stamp per dir, bound folded
in from add_location scans and allocate/mount calls, echoed fields
applied on heartbeat responses, reported on every heartbeat.
* volume: serialize vid stamp writes and fsync the directory (#11763)
* volume-rust: serialize vid stamp writes and fsync the directory (#11763)
* master: echo the allocation bound to connected servers on change
The bound was only sent at connect time or on digest mismatch, so a
healthy stream never learned later increases. If the high-id server then
went offline and the master lost state, survivors restored a stale bound
and the id could be reissued. Resend the heartbeat response whenever the
bound or topology id moved.
* volume: retry failed vid stamp saves on later updates
A transient write failure left the in-memory bound raised but the stamp
stale, and repeat updates at the same bound never retried. Track a dirty
flag so any later update re-attempts the save.
* volume: cover stamp bootstrap for pre-stamp dirs in test
* volume-rust: retry failed vid stamp saves, mirroring Go
* volume-rust: stamp every dir with the combined bound on add_location
A dir whose own stamp sat below the combined bound kept its stale value
after fold; losing the higher-stamped disk then restarted the server
below ids it had already learned. Rewrite all stamps after each location
folds in.
* volume: stamp mounted volume and ec shard ids (#11763)
* volume-rust: stamp mounted volume ids, mirroring Go (#11763)
* volume: retry dirty vid stamp save on unchanged topology id (#11763)
* volume-rust: retry dirty vid stamp save on unchanged topology id (#11763)
* master: record echoed bound from the sent heartbeat response (#11763)
* s3tables: accept .puffin data files for Iceberg v3 deletion vectors
Iceberg v3 writes deletion vectors as Puffin files under a table's data/
directory, so merge-on-read DELETE and UPDATE on v3 tables were refused
with AccessDenied by the table-bucket upload validator.
Fixes#11842
* s3tables: name all accepted data formats in rejection error
The message stopped at .avro even though .lance and now .puffin are
accepted.
* s3tables: cover deletion-vector filenames in data layout test
Uses the names real engines emit: Spark/Iceberg Java
{seq}-{task}-{uuid}-{counter}-deletes.puffin, UUID-suffixed names, and
the same under a partition directory.
* s3api: check the direct peer against the trusted-proxy allowlist
* s3: verify X-Forwarded-Prefix only from trusted proxies
* s3: note that trusted proxies must sanitize forwarded headers
* s3tables: honor a table policy Deny for maintenance operations
authorizeMaintenanceTable evaluated only the bucket policy, so a Deny in the table policy never vetoed PutTableMaintenanceConfiguration, GetTableMaintenanceConfiguration or GetTableMaintenanceJobStatus even though they act on the table ARN. The stored table policy is now checked for a matching Deny first via deniedByPolicy, a veto helper that keeps the admin exemption, and resource and bucket tags are loaded into the policy context so tag-conditioned Denies evaluate. readBucketPolicy is renamed to readPolicy and reused for the table path.
Generated with [Devin](https://devin.ai)
Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* s3tables: honor the stored policy's Deny on policy management operations
PutTablePolicy, GetTablePolicy and DeleteTablePolicy only evaluated the bucket policy, so a Deny in the table's own policy was ignored and a bucket-level Allow could strip or replace the policy guarding a table. The existing table policy is now loaded and a matching Deny vetoes the call. PutTableBucketPolicy likewise consults the existing bucket policy for a Deny before replacing it. Resource and bucket tags populate the policy context so tag-conditioned Denies evaluate, and policy writes assert the policy authorized against is still in force, returning a conflict when it changed mid-request.
Generated with [Devin](https://devin.ai)
Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* s3tables: apply a table policy Deny to tag operations
TagResource, ListTagsForResource and UntagResource evaluated only the bucket policy; a Deny in the table's own policy is now honored when the target ARN is a table. ListTagsForResource and UntagResource also returned early when the resource carried no tags attribute, skipping the permission check entirely; the check now always runs. Policy evaluation uses the canonical resource ARN rebuilt from stored metadata, so a request ARN with a forged account or region cannot dodge the Deny.
Generated with [Devin](https://devin.ai)
Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* s3tables: regression tests for resource policy denies
Cover the case where a bucket policy grants an identity s3tables:* while the table policy denies it: the policy, tag and maintenance operations must stay denied so the protection cannot be stripped. Also cover a bucket policy that denies PutTableBucketPolicy under the open fallback, and a tag-conditioned Deny, which vetoes only while the stored tag matches.
Generated with [Devin](https://devin.ai)
Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---------
Co-authored-by: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* volume: give Volume interior locks mirroring Go's
Volume's mutable state moves behind interior locks matching Go's
per-volume lock domains, so the store can hand out shared handles
instead of requiring exclusive &mut access:
- io: RwLock<VolumeIo> owns the .dat/remote handles, needle map,
append cursor, compaction bookkeeping and write watermarks —
Go's dataFileAccessLock-adjacent mutable set; helpers take the
guard as &mut VolumeIo so nothing re-locks.
- super_block and volume_info each get a RwLock
(superBlockAccessLock, volumeInfoRWLock); dir_idx gets its own
so path building never contends with I/O.
- no_write packs Go's noWriteLock pair into AtomicU8;
last_modified_ts_seconds and compaction_byte_per_second become
atomics.
The data-file write lease (dataFileAccessLock) now covers every
mutating entry that previously relied on borrow exclusivity:
writes, deletes, blob writes, set_writable's flag/index swap,
remote-file updates, local-backend swaps, index relocation,
replica-placement rewrite, put_needle_index, close and destroy.
set_read_only_persist leases only its can_delete idx-attach branch
because mark_io_unavailable reaches it under a held lease.
commit_compact takes try_write_lock — callers hold the store lock,
so blocking would stall every volume behind one slow reader;
WouldBlock surfaces as COMMIT_WAITS_FOR_READS for the caller to
retry.
Lock order is fixed: lease -> io -> volume_info/super_block/dir_idx.
* storage: mount volumes as shared Arc<Volume> handles
DiskLocation's map owns Arc<Volume> so the store can hand a caller a
handle that stays valid across store-guard releases, instead of
borrowing &mut Volume under the store write lock.
- volumes: HashMap<VolumeId, Arc<Volume>>; find_volume returns a
cloned Arc.
- find_volume_mut and volumes_mut are gone — every Volume method is
&self now, so shared handles cover what exclusive borrows did.
- writable_volume_mut -> writable_volume(&self) -> Result<Arc<Volume>>
keeps the same admission checks (exists, disk-space-low).
- write_volume_needle/write_volume_needles/delete_volume_needle are
&self; lifecycle methods that mutate the map or mounted state stay
&mut self.
* server: run needle I/O with the store guard released
The store write lock previously wrapped needle writes and deletes
end-to-end: process_batch held it across every grouped fsync, and the
HTTP/gRPC fallbacks, tail receiver and fetch-and-write paths did the
same. One volume's slow I/O then stalled admission and metadata on
every other volume.
- write_queue resolves each group's handle via writable_volume under
a brief store read, then runs write_needles_grouped after the guard
drops.
- HTTP write/DELETE fallbacks, the gRPC batch-delete EC recheck +
delete, tail receiver and fetch_and_write local writes, blob writes
and the quarantine/sync/set_writable/remote-file paths all run
under store.read() now that every volume method they reach is
&self.
- Lifecycle calls that mutate the volume map or mounted state
(mount/unmount/delete, EC shard ops, compaction commits,
load_new_volumes, heartbeat volume actions) keep store.write().
* test: cross-volume admission proceeds during another volume's write
Regression coverage for the store-lock fix:
- write_queue's test parks a volume-1 write inside its .dat append via
a new needle_write_hook seam (mirroring needle_read_hook), then
proves the store still hands out a lifecycle writer and a volume-2
handle — while volume 1's data-file lease still excludes same-volume
writers.
- tests/store_lock_repro.rs exercises the public API: concurrent
read-guarded writes on two volumes, and a stale Arc<Volume> handle
failing closed after unmount instead of writing into torn-down
files.
* test: drop store write guards that shared handles made read-only
Test setup sites held store.write() purely to reach find_volume_mut;
with Arc<Volume> and &self methods the guard is read-only (clippy
readonly_write_lock). The volume lease still serializes the writes
inside.
* volume: drop a windows-only read_exact_at that shadows the io helper
cargo's windows build failed E0255: the free fn copied with the
interior-lock conversion collided with the unconditional
crate::storage::io::read_exact_at import, which already carries the
platform branches internally.
* server: release the store read guard before needle I/O
The read guard still ran through write_needle_blob_and_index, the tail
receiver write, the fetch-and-write local write, and the HTTP write and
DELETE fallbacks, so a slow append blocked mount/unmount writers (and,
on writer-preferring locks, readers on other volumes). Resolve the
Arc<Volume> under a short guard and write through the handle after it
drops — the same scoping process_batch already used.
* volume: serialize the read-only mark and its .vif save under the lease
set_read_only_persist only leased the idx-attach branch, so concurrent
mark-readonly and mark-writable calls could interleave flag updates and
.vif writes — including two save_vif calls truncating and writing the
file at once. The lease now spans the whole flag change and save, with
a _held variant for mark_io_unavailable, which reaches this from write
paths already holding the lease (Go: markVolumeReadonly runs under
dataFileAccessLock).
* volume,server: span the tier-down transition under one write lease
update_remote_files cleared the remote-only write restriction and
released its lease before open_local_dat_backend attached the local
file, so a write could slip into the gap — pass check_writable, then
fail appending to a missing dat_file. The handler now holds the lease
across the reference trim, .vif save, and backend swap (and the rollback
re-insert), via _held variants; the store read guard is scoped to the
handle lookup. The swap still lands after the .vif trim — the ordering
that keeps a crash from pointing the volume at a deleted remote object.
* server: run a batch's volume groups on separate scoped threads
A single batch can hold writes for several volumes; running the groups
sequentially made every later volume wait behind a stalled append on an
earlier one — the same head-of-line blocking the store guard release
fixed, one level up. Each group now gets a scoped thread, preserving
order within a group while independent volumes progress in parallel.
* storage: close a volume handle a mount slot replaces
With shared handles, inserting a new Volume over an existing map entry
left the old handle alive in callers — two Volume instances could then
write the same files with separate in-memory indexes. set_volume and
create_volume now close() the replaced handle so its writers fail
closed, matching what unload_volume already does.
* test: key the write hook off a reserved needle id
The hook registry is process-global and keyed by needle id, and other
parallel tests write NeedleId(1) — an unrelated write could enter the
barriers and leave the test's real writer parked forever.
* server: wait out readers for vacuum commit without the store guard
commit_compact try-locked the write lease because callers held the store
write lock, returning COMMIT_WAITS_FOR_READS while a download was in
flight — and the master reads that as a failed vacuum, discarding
completed compaction work. The commit RPC now resolves the handle under
a brief store read and blocks on the volume's lease, matching Go's
CommitCompact under dataFileAccessLock: a slow reader stalls only this
volume's vacuum. The try-lock path stays for callers that still hold the
store lock.
* server: scope the remaining store read guards to volume lookup
Same class as the write-path fix: the batch-delete cookie check and
tombstone, the query read, volume_needle_status, and the local chunk
read all kept the read guard through the needle operation, so slow I/O
there could still stall store writers. Each now resolves the Arc<Volume>
under a short guard; the EC rechecks keep their guards.
* seaweed-volume: fail closed on detached volume handles
A handle resolved under a brief store read can outlive its mount: an
unmount+remount between lookup and lease leaves the Arc pointing at a
closed Volume while a new instance owns the files. Mark the handle in
close() and reject it under the write lease in commit_compact_held,
delete_needle, check_writable, set_read_only_persist and
update_remote_files_held; the tier-down handler also checks the flag
before deleting a remote object its replacement still reads.
* seaweed-volume: drop the store guard before mark-readonly waits on the lease
mark_volume_readonly held a store read across set_read_only_persist, which waits out the volume's leased readers; a slow download would stall every lifecycle operation behind the guard. Resolve the handle under a short read, then mark off the guard; a detached handle now maps to not_found.
* seaweed-volume: retry batch deletes that raced an EC mount
The store guard now drops before delete_needle, so ec.encode can mount shards between the pre-delete recheck and the tombstone; the encode then removes the .dat and the needle resurrects from pre-tombstone shards. Re-check has_ec_volume after each successful delete and return the retriable 503 so the filer requeues onto the EC path.
* seaweed-volume: one write-queue worker per volume
A stalled batch kept later batches for other volumes queued behind it: the single worker awaited process_batch before receiving again. Go's volume_write.go keeps one batching worker per volume, so give each volume its own FIFO worker spawned lazily, bounded by a global concurrent grouped-I/O semaphore and a global queued-request bound; workers exit only when their channel closes, and submit refuses unknown volume ids before spawning.
* seaweed-volume: move the queued-write permit inside the request
submit held the global queued-slot permit across the send and the response wait, so a canceled submit freed the slot while the worker still held the needle; repeated cancels could overflow the bound. The request now owns the permit until the worker drops it.
* seaweed-volume: retire write workers whose volume left the store
worker_for kept a sender per volume forever, so servers that mount and retire volumes accumulated idle workers. A worker now polls while idle and retires once its volume leaves the store: close the channel so racing senders fail fast, drain accepted writes, then release the map slot — a replacement worker can never start while this one still holds queued writes, so same-volume FIFO survives retirement.
* design: durable volume-id allocation boundary (#11763)
* master: mirror max volume id to a file outside raft (#11763)
The allocation bound lived only in raft state, so a master that lost it
rebuilt the bound from whatever volume servers were connected and could
reissue ids held by offline servers.
NextVolumeId now persists the bound to <mdir>/max_volume_id before the
raft apply, and both raft implementations persist on apply and recovery
so follower mirrors stay current. The file also carries the topology id,
which survives raft state cleanup the same way the snapshot recovery did.
* master: serialize and fully sync the max volume id mirror (#11763)
* master: rewrite the mirror when the topology id changes at the same bound
A same-bound raft apply carrying a new topology id was skipped by the
lastPersistedMaxVolumeId check, leaving the mirrored topology id empty
until the next allocation. Track the mirrored topology id too and
rewrite when either value differs.
* drop the design doc from the PR stack
The reverse index walk dedupes seen NeedleIds with a BoomFilters bloom
filter. Its default kernel is FNV-1, whose final step only XORs the last
input byte into the hash's low bits, so keys sharing the high 7 bytes of
a big-endian NeedleId land on correlated probe positions and false
positives arrive in bursts of hundreds instead of the ~0.001 target
rate. Each false positive hides a live-latest key and inflates the
rebuilt deletion counters, which is what made
TestFastLoadingNeedleMapMetrics flake past its +256 bound in CI
(observed DeletedCount drift of 267).
Swap the kernel for xxhash, which fully diffuses the key bytes; the
observed false-positive count in 300 local runs of the test drops to
max 2 (avg 0.23).
* ec: compare live idx NeedleId sets before encode
Refuse encode when replicas' live .idx NeedleId sets differ. Walk only
live rows (skip deleted and Size<=0) and compare NeedleId, ignoring Size
so MIME/flags drift does not abort a matching id set. Mark the volume
read-only for encode and reject writes after that.
* fix(ec): compare all replicas then pick capacity source
Log matching live idx sets with "replicas are consistent" for the
shell CI check. After readonly, compare every replica's live NeedleId
set, then require the encode source to have FreeVolumeCount>=2.
Empty live-idx vs non-empty refuses encode instead of skipping
superblock stubs. Drop the FileCount rebind after the idx gate.
* volume: mark readonly under the write lock so checked appends finish first
* ec: fix readonly deadlock and restore worker-host source preference
MarkVolumeReadonly held dataFileAccessLock while calling
reopenIdxForWrite, which re-acquires the same mutex and deadlocks
TestMarkVolumeReadonlyCanDelete. Split out reopenIdxForWriteLocked for
callers already holding the lock, matching the existing Locked pattern.
The worker task lost its worker-host source preference when the idx gate
replaced the replica union. Late-bind the source to a replica on the
worker host using only the checked location set, with no FileCount
rebinding. Drop the now-unused helpers, trim comments added by the
earlier commits, and rename tests whose names still claimed convergence
while asserting refusal.
---------
Co-authored-by: Chepurnov Simon <195125934+CHEPURNOV-S-S@users.noreply.github.com>
Co-authored-by: Devin <devin-ai-integration[bot]@users.noreply.github.com>
Co-authored-by: Chris Lu <chris.lu@gmail.com>
A restarted goraft master wipes its raft state and only learns the
TopologyId after the leader re-admits it and replays the entry. That
trails HTTP readiness by several seconds, so the one-shot check in
assertTopologyIdConsistent raced the propagation (CI run 39199033687:
TopologyId applied on the follower ~2s after the test gave up).
Chris Lu
github-actions[bot]
