The queue holds sixteen sealed windows, which is a memory bound only
while a window is BufferSize. An entry larger than that grows its window
to fit, and the depth then multiplies straight through: sixteen queued
copies of a 100 MB window is 1.6 GB of flush data alone.
Account the queued bytes and make producers wait once they pass the
ceiling the depth was chosen for. What is charged is the pooled slab
rather than the window length, since mem.Allocate rounds up to a size
class and the queue holds the whole slab. A window larger than the whole
budget still goes through on its own, so an oversized entry is never
stuck.
Windows are admitted in the order they were sealed. A producer can now
park here for seconds, and letting a later window overtake an earlier one
would persist them out of order and walk lastFlushedOffset and
lastFlushTsNs backwards.
A window is copied into its slab under the write lock, before the
reservation is taken, so a burst of concurrent oversized writers would
each hold a full copy in hand while queueing up -- memory the budget
never sees. Large writers wait for queue headroom before they take the
lock, which throttles the burst; it does not bound it, since a writer
that passes the check still seals unconditionally.
Take any room in the queue before the shutdown escape, too: the window is
already sealed by then, so dropping it loses records the caller was told
were accepted. A shutdown that races a full queue can still drop one --
that predates this change and needs the flush loop's lifetime reworked.
Size a grown window to the entry rather than to twice it: the extra room
only bought space for a second oversized record in the same window, which
doubles the flush copy and the snapshot taken of it. The overflow guard
halved its bound for that doubled allocation, so raise it to match what
is now allocated and what maxBufferSize documents.
* volume: reject needle blob writes to read-only volumes
WriteNeedleBlob appends the blob to .dat and only then calls nm.Put. On a
read-only volume the needle map is a SortedFileNeedleMap whose Put always
fails, so the append is never indexed and never rolled back.
Nothing upstream stops this: volume.check.disk picks its targets from the
master's cached topology, which goes stale the moment a volume server marks
a replica read-only itself — a failed data integrity check at load, or an
EIO quarantine. Each sync attempt then grows the .dat of a replica that is
supposed to be frozen by one unindexed needle, and reports it as "invalid
argument", the bare os.ErrInvalid the needle map returns.
Check IsReadOnly before touching .dat, same as the upload path does.
* volume: say which needle and volume failed to index
An index write that fails surfaced as a bare errno with no volume, no needle
and no file — "invalid argument" for a read-only needle map, or a plain
ENOSPC when .idx lives on its own filesystem via -dir.idx. Both were logged
at V(4), so by default the operator saw only the errno the client got back.
* filer: write the metadata log in pieces a volume server will accept
A single oversized metadata event grows the log buffer past the volume
server's fileSizeLimitMB, and the flush of that buffer is then rejected
forever: the retry loop has no exit, so the blob at the head of the queue
blocks every later flush and the metadata feed stalls until restart.
Split the flushed buffer into BufferSize pieces, on record boundaries
where possible so each piece still decodes on its own, and retry each
piece separately so a partial success is not replayed. Log files are
already read as a chunk stream, with a whole-file fallback when a chunk
does not decode standalone, so a record may cross a piece boundary.
* log_buffer: let go of a window array grown for an oversized entry
An entry larger than BufferSize grows the window array to 2*size+4, and
window arrays cycle through SealBuffer rather than being freed. One such
entry therefore leaves every later window carrying its size, and
currentSnapshotView allocates a snapshot as wide as the array on each
window, so a few KB of metadata keeps paying for it.
Drop the array when SealBuffer hands it back. Growth is on demand, so
the next oversized entry just reallocates.
* iceberg maintenance: store merged data files as chunks, not inline
saveFilerFile had no size threshold, so compaction wrote whole merged
parquet files -- hundreds of MB -- as Entry.Content. That puts the
parquet bytes verbatim in the filer store and sends them through the
metadata change log again as one event.
Keep manifests and metadata JSON inline, upload anything larger to
volume servers in chunks, assigning through the filer so the path's
storage rules apply.
* filer: follow the file size limit the volume servers report
The starting piece size is a constant, so a cluster whose
-fileSizeLimitMB is set below it would reject every piece and wedge just
the same. The rejection names the limit, so take it from there and
re-cut the rest of the flush to fit.
Piece the buffer one at a time rather than up front, since the size can
change partway through a flush.
plugin: stop the scheduler lock test racing its own background loops
TestRunLaneSchedulerIterationLockBehavior constructed the plugin with a
cluster-context provider, which makes New start a background scheduler
loop per lane. Those loops call runLaneSchedulerIteration on the same
lane the test then drives by hand, so a loop could consume the due job —
running detection and pushing the next-detection time forward — before
the manual call observed the lock. The Default case then saw the lock
acquired zero times and failed intermittently.
Construct without the provider so no loops start, and set the provider
afterward so the manual iteration can still detect. This is the pattern
scheduler_status_test.go already uses for the same reason.
Reproduced under -race -count=100 -p 4 before, green after.
Claude-Session: https://claude.ai/code/session_01Ks16jnt4S7gdDk8cheQ3xu
* util: match transient error messages case-insensitively, expose the message form
The same condition reaches different layers capitalized differently -- a
volume server relays its idle timeout as "I/O timeout" inside a JSON string --
and the callers that grew their own substring lists all lower-case first.
Also split out IsTransientErrorMessage for the paths that carry only the text,
such as the per-file status strings in a batch delete response, and pick up
"no route to host" and "network is unreachable" from the gRPC classifier.
* filersink: classify transient network errors through util.IsTransientError
The local list caught i/o timeout, connection reset, and broken pipe but not
connection refused, no such host, unexpected EOF, the syscall errnos, or the
gRPC and S3 overload codes. Keep only the bare io.EOF case, which is transient
here -- a truncated chunk read -- but a clean stream end elsewhere.
* filer deletion: reuse util.IsTransientErrorMessage for the network patterns
Six of the sixteen patterns were already covered. Keep the ones specific to
this pipeline -- read-only volumes, lookup failures, backpressure -- and note
why context cancellation stays retryable here: it decides whether to requeue
the deletion, not whether to retry a call.
* wdclient: fold the shared classifier into the volume lookup retry check
The string tail duplicated the shared list and missed the syscall errnos and
net.Error timeouts. Keep "connection" and "timeout", which are broader than
the shared classifier on purpose: a volume lookup is a cheap read-only call.
* filer.replicate: commit the kafka offset after replicating, not on receipt
The partition consumer committed the offset as soon as it handed the message
to the channel, so a sink write that failed was logged and the message was
already behind the committed offset -- never redelivered, permanently missing
from the sink.
Commit in onSuccessFn instead, and hold the committed offset behind the
oldest offset that failed to replicate so a restart redelivers from there.
* filer.replicate: delete the sqs message after replicating, not on receipt
ReceiveMessage deleted the message before the replicator had a chance to run,
so a failed sink write dropped it for good. Move the delete into onSuccessFn
and leave the message in the queue otherwise, letting the visibility timeout
redeliver it.
* util: retry transient errors, not just the ones containing "transport"
util.Retry only retried when the error string contained "transport", so a
plain "read: connection reset by peer" from S3 got zero retries. Classify
the error instead: net timeouts, connection resets, and the throttling and
overload replies S3 and gRPC return are all worth another attempt, while a
cancelled or expired context is not.
* filer sync: hold the sync offset behind a failed event
A sync job that returned an error was logged and forgotten, and the
watermark advanced past it anyway. The offset is the durable resume point,
so the event was never replayed: for filer.remote.sync that left the file
present locally, absent on the remote, with no RemoteEntry and nothing to
retry it.
Pin the watermark at the oldest failed event. Later events keep flowing,
but the persisted offset stays behind the failure, so a restart replays it.
The scaffold advertises enable_tls, ca_cert_path, client_cert_path and
client_key_path under redis2, redis2_sentinel and redis_cluster2, but only the
plain redis2 and redis3 stores ever read them, and under a different name,
enable_mtls. Sentinel and cluster setups quietly connected in plaintext.
Build the TLS config in one place and use it from all six stores. enable_mtls
still works. The CA and the client key pair are optional now, so enable_tls
alone verifies against the system roots, and ServerName is left unset so
go-redis validates each address the sentinel and cluster clients dial.
redis2_sentinel and redis3_sentinel stores only passed Username/Password
into redis.FailoverOptions, which authenticates against the Redis
master/replica servers. When Sentinel itself requires auth (requirepass
set in sentinel.conf), go-redis had no credentials to send to it,
causing a NOAUTH error before ever reaching the master.
Add sentinel_username/sentinel_password config options that map to
go-redis's SentinelUsername/SentinelPassword fields, distinct from the
existing master auth credentials.
Deleting the only object under a prefix in a versioned bucket writes a
delete marker and keeps the version history, so the filer directory
survives with nothing a current-version listing would return. A delimited
ListObjects kept reporting that path in CommonPrefixes, because the
prefixes come from the directory tree rather than from the keys, while a
listing scoped inside the prefix correctly came back empty.
Probe a directory before reporting it: one that holds entries but no key
the listing returns is neither a CommonPrefix nor a path the
trailing-slash probe answers for. Empty directories keep the meaning they
have today, and the probe only runs for buckets with versioning
configured, the only ones that can reach this state.
A maintenance snapshot carried only its own labels — merged-files,
delete-groups and friends — and no summary counters, so every engine that
reads a table's size out of the current snapshot summary reported nothing
for it: PyIceberg's inspect.snapshots, Trino's $snapshots and Spark's
DESCRIBE all read total-records, total-data-files and total-files-size
verbatim, and a table lost them the moment compaction touched it.
Accumulate the files each operation adds and removes, and render them the
way the spec defines: the added-*/removed-* counters from the files
themselves, then the totals carried over from the parent snapshot.
Carry a total only when the parent recorded it. Iceberg treats a missing
total as zero, which turns a compaction replacing two files with one into a
negative total-data-files, or a table with millions of rows into
total-records: 0. Leaving the field out lets a reader fall back to the
manifests instead of believing a made-up number.
Compaction also accounts for the delete files it consumes, so a run that
folds every delete into the rewritten data reports them as removed.
The worker assumed every file of a table sits under its catalog path, so
loadFileByIcebergPath stripped the scheme off a recorded location and joined
the remainder onto /buckets/<bucket>/<ns>/<table>. A table the REST catalog
placed elsewhere in the bucket — which is what a client gets whenever the
catalog path is already occupied — then resolves to a doubled path:
lookup /buckets/lake/source/t/lake/source/t-0cd81bca-.../metadata/snap-.avro
so the very first manifest list read fails and the job fails again on every
scan interval, indefinitely.
Resolve absolute references (s3:// URIs and /buckets paths) from the bucket
root and keep relative ones under the table's own directory; the
bucket-relative form is now the canonical key everywhere references are
compared. That directory comes from the metadata location the catalog stores,
so reads, writes and deletes all land where the table's other files are
instead of splitting it across two trees. References outside the table's
bucket are rejected rather than silently misresolved.
Rewritten position-delete files now name their data file by absolute URI,
the way the table itself names it, instead of a path relative to the table.
The dashboard charted one summed disk-usage line, so a step in the total
gave no hint which cluster moved. A new panel stacks each cluster's daily
size as its own band: the top of the stack is the fleet total, each band
is one cluster, and the clusters past the twentieth are summed into an
"other" band so the stack still adds up to the total.
The series is built from the per-cluster daily histories and served by
/api/cluster-sizes. Clusters report roughly once a day at no fixed hour,
so a day with no report carries the previous value forward — dropping it
to zero would sag the total every day as the clusters that have not
reported yet fall out from under it. A cluster that stops reporting past
the active window ends at its last sample instead of holding capacity
forever. Ranking is by the most recent day, tie-broken on cluster id so
the colors do not shuffle between refreshes.
Hover and click resolve to the band under the pointer: Chart.js's builtin
interaction modes match the nearest line, which on a stack of thin bands
is rarely the band being pointed at. Clicking one fills the per-cluster
history lookup below it.
LiveMoveVolume and the copy, tail, delete, mark, replicate, and
configure helpers around it issued every RPC on context.Background(), so
a caller had no way to bound or abort a move once it started. They now
take a context, which the exported LiveMoveVolume in particular needs:
callers outside the shell drive long moves and want to stop them.
The deferred restore in copyVolume runs on a detached, bounded context
rather than the caller's. Marking the source writable again is cleanup,
and cancelling the copy must not skip it and leave the volume readonly —
the same guard balance_task.go already applies for the same reason.
Shell commands pass context.Background(): their Do signature carries no
context, and changing it would touch every command in the package.
Claude-Session: https://claude.ai/code/session_01Ks16jnt4S7gdDk8cheQ3xu
* filer: stamp a log position on lookup and remote-cache responses
Metadata events are logged after their store write and stamped with the
filer clock. Reading that clock before serving an entry therefore gives
a timestamp with a causal guarantee: every event at or below it is
reflected in the returned entry. Clients caching filer state can use it
as the entry's version to order the response against subscription
events, including events committed before the call but delivered after
it.
* mount: version open file handles by filer log position
A subscription event refreshing an open handle did a second lookup; a
transient failure left the handle pinned to its old entry with no
retry, since the subscription cursor had already advanced. The deeper
problem is ordering: the handle is a cache written by three unordered
channels — the async invalidation worker, local mutation acks, and
open-time lookups — and overwriting cached state safely requires
knowing which write is newer.
The filer log timestamp is that order, and it now travels with every
value instead of being derived out of band. Events carry it natively;
lookup and remote-cache responses carry the log position stamped before
the serving read; mutation acks carry it in their returned event; and
the local store pairs each read with a version cursor advanced under
the same lock as the store write. Each handle records the version its
entry reflects, and one rule replaces the per-site reasoning: state at
or below the handle's version is old news and must not be installed.
The invalidation itself applies the event's own entry — no lookup, so
no transient-failure window — except under a cached parent, where the
store entry is the ordered merge of the event and anything applied
since, and its version outranks the event's. An uncached parent
receives no store writes, so a hit there would be a stale leftover
masking the event. A vacated path (delete, rename away) keeps the last
entry so unlinked-but-open reads still work. Directory builds version
the completed directory at the listing snapshot and re-invalidate
buffered events at that version, since their mid-build refresh ran
against an incomplete store.
The tests replay every race this replaces machinery for: rollback of a
newer local flush (queued, cached, and read-through), stale leftovers
under uncached parents, the build window including abort, handles
opened after an event was queued, events landing mid-lookup, and
undelivered events at remote-cache time across a filer failover.
* filer: serialize the log position fence with mutations, stamp mutation acks
The fence stamped before an unlocked entry read could precede state the
read returned: a mutation writes storage first and assigns its event
timestamp only at notify time, so a lookup racing that window handed
the mount an entry newer than its fence, and the event's later delivery
looked like fresh news — destroying dirty pages for a change the handle
already had. The mutation handlers already hold an exclusive per-path
lock across read, write, and notify; the lookup and remote-cache reads
now take it shared around the stamp and the read, making the fence
exact: everything at or below it is in the entry, nothing above it is.
A no-change update returns success without an event, leaving the mount
nothing to fence with even though the response confirms current state.
Create and update acks now carry a log position stamped under the same
lock, and the mount falls back to it whenever the ack has no event.
Also regenerate the VT marshalers, which the earlier generation missed:
without them a VT round-trip silently zeroed every log position.
* java: sync filer.proto
* mount: scope store versions to what they vouch for; atomic handle install
The store's version cursor claimed too much. Advanced by local mutation
acks and directory listing snapshots, it inflated the version of store
reads for unrelated paths whose events the subscription still owed, and
those events were then fenced out permanently. The cursor now tracks
subscription progress only — events arrive in log order, so everything
at or below it has been delivered for every path — and a completed
listing records its snapshot as a per-directory floor instead of a
global claim. Local acks never touch it: they version their own handle
directly. Buffered build events advance the cursor at delivery, since
their store write may never happen (abort) while their invalidation is
already queued; their read-through directory pairs no store read with
it, and rename fragments are applied first.
Concurrent first opens raced: a slower opener's older lookup could
overwrite the newer entry a faster opener had installed, while the
monotonic version kept the newer timestamp — an old entry fenced at a
new version, immune to every correcting event. Entry and version are
now installed as one decision under the handle map lock, and an install
that does not outrank the handle's version is dropped.
The remote-cache commit also escaped the fence: it wrote storage and
notified without the path lock, so a lookup's shared-locked fence and
read could land between the two and hand out the cached state
under-versioned. The commit now re-reads and writes under the exclusive
path lock, and backs off entirely when the entry changed during the
download — the concurrent writer supersedes the cached content.
* mount: floors gate store applies; installs respect handle users; renames join the fence
A directory floor certifies the listing state as of its snapshot, but a
delayed event at or below the floor was still applied to the store —
rolling the content back to pre-snapshot state while the floor kept
claiming the snapshot version, so the correcting events were fenced out
of every future read. Events are now gated against the affected
directory's floor, each half of a rename independently.
Fences are lower bounds: a listing or lookup can include a mutation
whose event has not been delivered yet, and that event later passes
every gate carrying state the handle already holds. Such a re-delivery
now advances the version without destroying dirty pages or reinstalling
the entry — invalidating local writes over a no-op was the real damage
in every remaining under-fence window, including the unlocked listing
snapshot, which no per-path lock can serialize.
The concurrent-open install moved from the map lock to the handle lock
every reader, writer, and invalidation synchronizes on, and rejects
what cannot improve the handle: dirty state (local writes would be
lost), unversioned lookup responses (they cannot outrank anything, and
two zero-version opens must not overwrite each other), and anything not
strictly newer. New handles are still fully initialized before the map
exposes them.
Renames committed metadata and emitted events with no path lock, so a
lookup could read the renamed state under a fence preceding its events.
Both rename handlers now hold the source and destination locks, ordered
by path, across commit and notification; descendants of a renamed
directory are not individually locked and rely on the no-op re-delivery
handling above.
* mount: per-entry store versions replace the cursor and directory floors
The store's aggregate versions — a global subscription cursor and
per-directory listing floors — were versions at coarser granularity
than the values they described, and every over-claiming bug in this
series traced to that gap: an aggregate vouching for state its source
never saw. Each store entry now carries the filer log position of the
write that produced it — the event that applied it, or the listing
snapshot that inserted it, recorded in the store's key-value space
under the same lock as the entry write. The store becomes what the
handle already is: a last-writer-wins register with one rule, install
only what outranks the current claim.
The cursor, the floors, their advancement rules, the pairing ordering
constraint, and the floor gating all collapse into that rule. Applies
are gated per entry, each half of a rename independently; an
unversioned local write clears the claim its content no longer proves;
version records lingering after a bulk folder wipe cannot fence a
recreate, since a claim only blocks while its entry exists. Listing
inserts are stamped at build completion, before the buffered replay so
newer replayed events override the stamp.
Filer side, the fence dance every versioned read must perform is now a
single choke point, fencedFindEntry, so a future read RPC gets the
lock-serialized stamp by construction rather than by convention.
* mount: judge no-op re-deliveries against an immutable base, not the live entry
The equal-state skip compared the incoming event to the live handle
entry, but local writes mutate the live entry — size, timestamps,
chunks — so a delayed event re-delivering the base the handle was
opened with no longer matched, and the installer destroyed the dirty
pages and rolled the entry back over nothing new. The handle now keeps
an immutable snapshot of the filer state it last installed or
acknowledged, refreshed at every install and mutation ack (flush acks
snapshot the request entry before the id mapping mutates it), and the
no-op judgment runs against that base: an event carrying the base
brings nothing, whatever the live entry has diverged to since.
* mount: tombstones for versioned deletes, absence floors, copy enrollment
Four gaps in the per-entry version protocol, all the same shape: a
versioned fact with nothing carrying its version.
A deletion is a fact about a path with no entry left to hold it —
clearing the record let a delayed older event resurrect the deleted
path, permanently, since the deletion's own redelivery is
dedup-suppressed. Versioned deletes now leave a tombstone record that
fences without an entry; renames tombstone their source the same way.
Plain records still only block while their entry exists, so records
lingering after a bulk folder wipe cannot fence a recreate.
A completed listing proves absences as well as presences: a name it
omitted was deleted as of the snapshot, and a delayed create below the
snapshot re-creates it. The snapshot is kept per directory strictly as
an absence fence, consulted only when a path has neither an entry nor
a version record — present entries carry their own versions and never
touch it, which is what separates this from the over-claiming floor it
replaces.
A rebuild against a pre-upgrade filer returns no snapshot; stamping
now clears the children's records in that case, so a reinserted entry
cannot reactivate the stale claim its previous incarnation left
behind and reject valid events below it.
Server-side copies installed the copied entry without enrolling in the
base protocol, so the copy's own event differed from the stale
pre-copy base and destroyed writes made to the destination after the
copy. The install now refreshes the base and takes its version from
the fenced readback.
* mount: deletion facts outlive the cache's knowledge of the entry
A versioned delete of a path the store held no entry for recorded
nothing, so a delayed older event recreated the path — permanently,
with the deletion's redelivery dedup-suppressed. The tombstone is now
written whenever a versioned event vacates a path: the deletion is a
fact about the path, not about what this cache happened to hold.
For an absent entry, the listing's absence floor now speaks whatever
older record remains: a tombstone at one position does not exhaust
what is known about the path when a newer snapshot has confirmed the
name still absent, and an event between the two was slipping past
both.
A committed copy whose readback failed installed a synthesized base
with local timestamps; the copy's real event legitimately differs from
it, and was read as foreign state — destroying writes made to the
destination after the copy. The handle now marks that its own event is
en route and adopts that event's state as the base without touching
the live entry or the dirty pages; the adoption is one-shot, so a
genuinely foreign event still invalidates.
* mount: authoritative acks cancel pending event adoption; tombstones scoped and pruned
The copy-event adoption flag could outlive its purpose: a flush after
the failed readback installs a newer base and advances the version, the
copy's own event is then version gated without consuming the flag, and
the next genuinely foreign event was silently adopted — base advanced,
live entry and dirty pages untouched — leaving the mount to later
overwrite that remote change. Every local acknowledgment now installs
its base through one helper that also cancels any pending adoption: the
ack supersedes the mutation the adoption was waiting for.
Tombstones were written for every versioned delete under the mount and
survived directory eviction by design, growing LevelDB with historical
deletions on delete-heavy mounts. They are now scoped to directories
whose cached state the fence actually protects — an uncached parent
never serves from the store nor applies the resurrecting insert — and a
completed listing prunes the direct-child tombstones its absence floor
supersedes, leaving only those above the snapshot. The store gains a
key-prefix visitor for the sweep.
* mount: acked saves install their value; trailer snapshots; direct-child prune range
A version must never advance without its value. saveEntry stamped any
open handle with the acknowledgment's version, but a handle opened
while the save was in flight holds the pre-mutation entry — stamping it
fenced out the events carrying the state it lacked, permanently, with
the local apply performing no invalidation and the redelivery
deduplicated. The acknowledged entry is now installed together with its
version, through the same guarded install the racing-open path uses:
under the handle lock, only when it outranks the handle, never over
dirty local writes.
Empty listings return no in-band snapshot — a snapshot-only response
would be read as an entry by older consumers — so directories that end
empty gained no absence floor and their tombstones were never pruned.
The filer now sends the snapshot in the stream trailer, which older
clients ignore, and the client reads it when no in-band snapshot
arrived. Empty directories get real floors, their tombstones prune,
and their buffered replays gain the snapshot filter instead of the
replay-all fallback.
Version records now encode the parent directory and name separated by
a NUL, making a directory's direct children one contiguous key range:
the tombstone prune scans exactly them under the cache lock, instead
of walking every descendant record — the whole store, for root.
* mount: fix dirty-page loss, uid/gid base, download race, copy adopt, leak; dedup
Correctness fixes from the versioned-invalidation review:
- A foreign delete/rename-away of a file held open with unflushed local
writes destroyed the dirty pages unconditionally. A process may keep
writing to an unlinked-but-open file and those writes were already
acknowledged; preserve the pages when the handle is dirty.
- downloadRemoteEntry stored the handle's base with filer-side uid/gid
while every candidate it is later compared against is in local form,
so under a non-identity UidGidMapper an unchanged re-delivery looked
foreign and force-destroyed dirty pages. Map the base to local.
- downloadRemoteEntry wrote the entry/base/version triple under only the
handle's shared lock, so two concurrent reads of the same remote-only
file could tear it. Serialize the install with a dedicated mutex
(invalidation is already excluded by the exclusive handle lock).
- A committed server-side copy whose readback failed adopted the FIRST
event past the version gate as its base; a foreign write delivered
first was silently swallowed. Adopt only an event whose content
matches the synthesized base — the copy's own event — and install any
other normally.
- The deferred-create path relied on AcquireFileHandle installing the
passed entry on a pre-existing handle, which the version rework
dropped. Restore that install in the compat wrapper; the versioned
open path keeps its gated install.
Growth and hot-path cost:
- Per-entry version records and tombstones leaked when a directory was
evicted or read-through without a rebuild. An uncached directory
gates its own inserts, so its records fence nothing; clear a
directory's child version records when it is wiped for eviction.
- FindEntry paid for the version KvGet on every lookup/getattr cache hit
and threw it away. FindEntry now reads only the entry; the hot
lookupEntry cache-hit path skips the version entirely.
Cleanups:
- Extract ackVersionTsNs over the shared response interface, replacing
the metadata-event-else-log-ts snippet copy-pasted at four ack sites.
- Extract acquireRenamePathLocks, replacing the verbatim sorted
two-path lock fence in both rename handlers.
* mount: no resurrection on foreign delete, version no-event acks, gate downloads, tighten copy adopt
Follow-ups to the review patches:
- Preserving dirty pages on a foreign delete let the next flush pass the
isDeleted guard and CreateEntry, resurrecting the remotely-unlinked
name. Mark the handle deleted in the vacate branch: the open fd can
still read its buffered writes, but a flush no longer recreates the
file.
- A no-event acknowledgment (log fence only) synthesized a metadata
event with TsNs 0, so the cache stored the entry unversioned and an
older subscriber event rolled it back. Stamp the synthesized event
with the ack's log position at all four ack sites.
- downloadRemoteEntry serialized its install but did not check the
version, so an older response arriving last overwrote the entry/base
while the monotonic version kept the newer value, fencing corrections
out. Install only when the response is at least as new as the handle.
- sameEntryContent compared only size and chunks, so a foreign chmod
with unchanged content was adopted as the copy's own event. Compare
everything except server-assigned timestamps, so a metadata-only
foreign change installs instead.
* mount: trim comments to the non-obvious why
The versioning work accumulated multi-line comment blocks restating what
the code says. Keep the constraint a reader cannot derive — why a fence
is exact, why a version must not advance without its value, why an
uncached parent's records fence nothing — and drop the rest.
* mount: distinguish rename from delete, tighten the download and adopt gates
- A rename emits a nil old-path invalidation just like an unlink, so the
vacate branch marked the handle deleted and later writes through the
already-open descriptor were skipped instead of persisted. Carry the
delete/rename distinction on the invalidation and mark only an actual
delete.
- The remote-download install accepted an unversioned response
regardless of the handle's version, so during a rolling upgrade a
delayed response could install stale content under a newer version.
Require the response to be at least as new, with one exception: a
handle still lacking local chunks takes the content anyway — it cannot
read without it — but does not claim the response's log position.
- Copy-event adoption returned without installing, so a foreign touch
arriving before the copy's own event lost its timestamps. Content is
unchanged either way, so the dirty pages stay valid; a clean handle now
takes the entry, while a dirty one keeps its diverged version.
* mount: one directory floor instead of a record per child; agree on TTL
Review feedback:
- Build completion wrote one KV record per direct child inside the cache
write lock, so a large directory stalled every other cache operation
for O(children) store writes. The directory's listing snapshot already
covers every child it saw; make that floor the version for any child
without a record of its own, and a child earns a record only when a
later event touches it. One map write per build replaces the per-child
writes, with the same fencing.
- The presence probe read the store directly and so counted a
TTL-expired entry as present, judging the path by a record describing
content that has logically vanished. It now applies the same expiry
the read path does, and an expired path falls back to its directory
floor.
- Preserve ErrNotFound identity when the commit-time re-read finds the
object deleted, so callers still surface a 404.
- Assert the rename-away source fence timestamp in the invalidation test.
Also record the tombstone ceiling: distinct deleted names in a cached
directory accumulate until it is rebuilt or evicted, which prunes
everything at or below the new snapshot.
* mount: pin the fence's clock domain instead of letting skew decide
A log-position fence is stamped by one filer's clock under that filer's
in-process lock, so comparing it to an event another filer logged is
comparing two unrelated clocks. The two error directions are not equally
costly: applying an event the fence already covered is a re-apply the
base-equality check absorbs, while skipping one it does not cover leaves
the handle holding exactly the state the event was meant to correct,
with the subscription cursor already past it — the unhealable staleness
this whole PR exists to remove.
So refuse to guess. Fences now carry the signature of the filer that
stamped them, and a handle records it alongside the position. An event
is only fenced out when the filer that logged it is the one that stamped
the fence — the logging filer appends its own signature, so its presence
identifies the clock domain. Events from any other filer are applied.
Positions taken from events keep comparing as before; the subscription
already delivers those in order.
The invalidation callback takes a struct now: it carries the path,
entry, position, delete/rename distinction, and signatures, and was
about to need a fifth positional parameter.
* mount: follow a foreign rename; key page invalidation on content, not equality
- A rename's old-path invalidation now carries the destination, and the
handle follows the file there: an open fd tracks the inode, and leaving
it on the old path made its next flush recreate that name instead of
updating the renamed file.
- Dirty pages overlay content, so only a content change invalidates them.
Keying that on exact equality meant any timestamp-only event destroyed
them, which the copy-adoption marker existed to paper over — a foreign
touch could consume the marker and leave the copy's own event to drop
the post-copy writes. Comparing content instead makes the marker
unnecessary, so it is gone: a metadata-only event keeps the overlay,
and a dirty handle keeps its diverged entry unless foreign content
supersedes it.
- A remote download response that is merely older is now refused even
when the handle still lacks chunks; only an unversioned one is taken
(and claims no position), since an older response's content predates
what the handle reflects.
- A refused or unversioned download no longer publishes to the metadata
cache, where a zero-position event would clear the entry's version and
let an older subscriber event roll the cache back.
* mount: page invalidation keys on content alone; unversioned writes claim no position
- sameEntryContent compared everything but timestamps, so a foreign
chmod, chown, or xattr change counted as a content change and
destroyed the dirty-page overlay. It was strict only to serve the
copy-adoption marker, which is gone; its one caller now asks the
question it actually needs — did the bytes change — so metadata-only
events leave the overlay alone.
- A rename over an existing file destroys that file, but its open handle
was left live and still pointed at the name the renamed source now
occupies, so its flush could overwrite it. MovePath already reports the
displaced inode; mark that handle deleted.
- An acknowledgment was refused whenever its position was numerically
lower, even when a different filer stamped the fence it lost to. Two
known, differing signatures mean unrelated clocks, so the comparison no
longer applies there; unknown signatures still compare as before.
- A local write with no log position behind it now records that
explicitly instead of deleting its version record. Absence means the
directory listing covers the path, which is why the snapshot floor
applies; local content the listing never saw must not inherit it, or
the events that would correct it are fenced out.
* mount: widen the existing lookup functions instead of forking WithVersion twins
The versioning work grew a parallel function for every accessor that
needed to return a log position — lookupEntryWithVersion beside
lookupEntry, maybeLoadEntryWithVersion beside maybeLoadEntry,
FindEntryWithVersion beside FindEntry, AcquireFileHandleWithVersion
beside AcquireFileHandle, advanceEntryVersion beside
advanceEntryVersionTsNs, plus a getPbEntryWithVersion wrapper and an
InsertListedEntriesForTest hook. Two names for one operation is two
places to keep in step, and the split let callers pick the one that
happened to compile.
Each pair is now the single original name carrying the position, with
callers that do not want it discarding it. filer_pb.GetEntry returns the
fence its response already carried rather than a mount-side wrapper
re-issuing the lookup, and InsertEntry takes the position its content
reflects rather than a test-only twin that inserted without one.
The one behavioural knot the merge exposed: AcquireFileHandle had been
installing the entry on a pre-existing handle only in its unversioned
form, which conflated 'the caller is authoritative' with 'the lookup had
no version'. Deferred create is the only caller that means the former,
so it now installs explicitly and the map function just acquires.
On renewal failure the renew goroutine stored isLocked=false before its
deferred renewGoroutineRunning.Store(false) ran. A concurrent RequestLock
interleaving there reacquires the lease (sees isLocked=false), sets
isLocked=true, then its CompareAndSwap on renewGoroutineRunning fails because
the old goroutine's flag is still set — so no replacement renewer starts. The
lock is then held locally with nothing renewing it, and silently expires on
the master after the lease TTL, admitting a second holder.
Clear renewGoroutineRunning before isLocked on the failure path so the
reacquire path always starts a fresh renewer. Builds and vets clean; no
behavior change on the success path.
Claude-Session: https://claude.ai/code/session_01Ks16jnt4S7gdDk8cheQ3xu
The copying and tagging tests force-drop each bucket's collection at the
master so volume slots are freed deterministically between tests. But the
copy-tests CI job runs its master on 9336 and the tagging Makefile on 9338,
while the tests default to 9333 — the cleanup dialed a dead port and quietly
no-oped. Each test bucket then grows 7 volumes against -volume.max=100, and
whenever async deletion lagged the data node ran out of slots and PutObject
500ed with "No writable volumes and no free volumes left".
Set MASTER_ENDPOINT where the master port is non-default: the copy-tests
workflow step, and the copying/tagging Makefiles (derived from MASTER_PORT).
* telemetry: validate reports on the collect endpoint
/api/collect is anonymous, so reports can't be authenticated, but a
real master can't produce a non-UUID topology_id, a version outside
N.NN(-enterprise), an unknown GOOS/GOARCH, or absurd counts — reject
those to keep casual junk out of the collected data, and cap the
request body at 4 KB.
* telemetry: integration test fixtures pass collect validation
The test's topology id and version were exactly the junk shapes the
new validation rejects; use a UUID and a plain version number.
telemetry: confirmed-cluster stats
Count a cluster as confirmed once it has reported on >=2 distinct UTC
days (per-cluster history makes this a length check). Version/OS
distributions in /api/stats are computed over confirmed clusters, so a
one-shot injected report can't appear in them; falls back to all active
clusters while no confirmed ones exist (fresh server). Adds the
seaweedfs_telemetry_confirmed_clusters gauge and a dashboard card.
telemetry: per-cluster usage history
Keep one compact sample per cluster per UTC day (disk bytes, volume
count, volume servers), retained for -max-age and persisted in the
state file. Serve it at /api/history?cluster_id=...&days=90 and add a
per-cluster lookup with disk/volume charts to the built-in dashboard.
* telemetry: persist server state across restarts
The telemetry server kept the instance map and Prometheus gauges only
in process memory, so every deploy or restart reset all collected
metrics until clusters re-reported over the next 24h.
Snapshot the instance map to a JSON state file (atomic tmp+rename) on
a debounced interval and on SIGTERM, and restore it on startup,
preserving received_at so the cleanup and active-cluster windows stay
correct. Defaults to data/telemetry-state.json, which the deployed
systemd unit's WorkingDirectory already provides; -state-file=''
disables.
* telemetry: keep instances for 90 days by default
With state now persisted across restarts, a longer retention default is
meaningful; raise -max-age from 30 to 90 days so per-cluster data
survives long enough for quarterly views.
ci: build telemetry server from its own module in deploy workflow
telemetry/server has had its own go.mod since #9924, so building
./telemetry/server/main.go from the repo root fails with 'no required
module provides package'. Build from within the module instead.
telemetry: key value gauges by cluster_id only
The value gauges were labeled {cluster_id, version, os}, so a cluster
reporting back after an upgrade started a new series while the old one
kept its last value forever: sum() double-counted every upgraded
cluster, and per-cluster history broke at every version change.
Key the five value gauges by cluster_id alone so each cluster keeps one
continuous series across upgrades; version/os metadata stays on
cluster_info (deleted and re-set on change), available to value queries
via 'on(cluster_id) group_left' joins. Update README accordingly.
telemetry: aggregate /api/metrics per day so dashboard charts render
The dashboard expects {dates, server_counts, disk_usage} as parallel
arrays, but GetMetrics returned per-instance {date,value} lists under
different keys, so the two over-time charts never rendered.
* Add GitHub Actions workflow for codespell on master
* Add rudimentary codespell config
* Tune codespell config: skip generated code, ignore camelCase, whitelist domain terms
Add camelCase/PascalCase regex to ignore common Go/Rust/JS identifiers
like allLocations, publishErr, ReadInside, FlushInterval. Also skip
templ-generated *_templ.go files, and whitelist a handful of
short/domain-specific words (visibles, fo, te, ser, bject, unparseable,
keep-alives, tread, anc, ue) that show up as false positives across the
tree.
Co-Authored-By: Claude Code 2.1.217 / Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* Fix ambiguous typos and protect false positives
Fixes typos that codespell reports with multiple candidate suggestions
(so `codespell -w` cannot auto-apply them), plus one inline pragma and
one config entry to protect legitimate identifiers.
Manual fixes (single correct answer chosen from context):
- pattens -> patterns (5x) in filer/upload/shell flag help strings
- finded -> found (2x) in tarantool storage.lua comment
- spacify -> specify (2x) in helm chart values.yaml comment
- wether -> whether in skiplist.go docstring
- simpe -> simple in mq schema test case name
False-positive protection:
- Add `//codespell:ignore` next to `source GET's` (possessive of HTTP
verb) in s3api_object_handlers_copy_stream.go
- Whitelist `auther` in .codespellrc — it's a local variable meaning
"authenticator" in weed/security/tls.go, not a typo of "author".
Co-Authored-By: Claude Code 2.1.217 / Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* Extend codespell ignore list: .git-meta path and thirdparty groupId
Also skip `.git-meta` (scratch dir for commit messages that may contain
typo words verbatim) and whitelist `thirdparty` — it appears as the
literal Maven groupId `org.apache.hadoop.thirdparty` in hdfs3 poms
and cannot be renamed.
Co-Authored-By: Claude Code 2.1.217 / Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* [DATALAD RUNCMD] Fix non-ambiguous typos with codespell -w
Auto-applied fixes to the 44 remaining single-suggestion typos across
docs, comments, log messages, tests, config, and one Java pom.
=== Do not change lines below ===
{
"chain": [],
"cmd": "uvx codespell -w",
"exit": 0,
"extra_inputs": [],
"inputs": [],
"outputs": [],
"pwd": "."
}
^^^ Do not change lines above ^^^
* Revert breaking codespell fixes; whitelist unknwon and atleast
Two of the auto-applied `codespell -w` fixes were false positives that
would break the build/tests:
- go.mod: `github.com/unknwon/goconfig` is a real Go module path — the
upstream author's GitHub handle is literally `unknwon`. Renaming to
`unknown` would fail dependency resolution.
- test/benchmark/fuse_db/bin/{sqlite_verify.py,run_mysql.sh,run_sqlite.sh}:
`atleast` is a literal CLI mode value (a string constant compared and
passed as a positional argument). Rewriting to `at least` splits it
into two arguments and breaks the mode check.
Reverted those files and whitelisted both words in .codespellrc so
future runs won't re-suggest the same broken fixes.
Co-Authored-By: Claude Code 2.1.217 / Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Code 2.1.217 / Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(admin): implement ApplyPluginConfigFromToml to propagate settings to plugin config store
* Update weed/admin/dash/config_toml.go
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
* admin: overlay admin.toml onto plugin configs through bootstrap defaults
Creating a config from scratch at startup skipped the descriptor-defaults
bootstrap, so a job type with only worker keys in admin.toml persisted
Enabled=false and RetryLimit=0 and silently stopped running. Overlay
existing configs at startup, and apply the same overlay in
enrichConfigDefaults when the plugin bootstraps a fresh config from
descriptor defaults.
Also place collection_filter in the admin values where workers read it,
map preferred_tags as a string list, and stamp UpdatedAt.
* admin: trim the admin.toml help text and call-site comment
* admin: clamp toml retry values to the int32 range
* admin: fail startup when admin.toml cannot reach the plugin config
The legacy overlay already aborts startup when declared settings cannot
persist; continuing here would let workers bootstrap with stale values.
* admin: fix the retry clamp test on 32-bit
A 32-bit int cannot hold the oversized toml value, so viper returns 0
before the clamp runs.
---------
Co-authored-by: baracudaz <baracudaz@users.noreply.github.com>
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
Co-authored-by: Chris Lu <chris.lu@gmail.com>
* volume.fix.replication: add a well-placed replica before deleting a misplaced one
A misplaced volume with no surplus replica (replica count == copy count) used to have its misplaced replica deleted first, and the replacement copied only on the next pass. That drops the volume below its intended durability, and permanently so when no destination can accept the replacement copy. Now such volumes get a well-placed replica first, and the misplaced one is trimmed as surplus on a later pass, following the same fulfill-before-delete principle as volume.tier.move (#8950).
The classification switch is extracted into classifyReplicaSet so the add-wins-over-trim priority is unit-testable.
Also stop the fix loop when an -apply pass makes no progress (nothing copied, nothing deleted); previously an unplaceable under-replicated volume made the loop spin forever, re-collecting the topology every 15 seconds. And propagate ewg.Wait() errors instead of swallowing them.
* volume.fix.replication: drop unused allLocations from deleteOneVolume
---------
Co-authored-by: Chris Lu <chris.lu@gmail.com>
* helm: generate the SFTP host key per install
The SFTP secret template shipped one fixed ed25519 host key, so every
install that did not override it presented the same host identity.
Generate the key at install time instead, following the
getOrGeneratePassword pattern: an existing secret keeps its key across
upgrades, except the previously bundled one, which is replaced with a
freshly generated key on the next upgrade.
* helm: create the SFTP host-keys secret the deployments mount
Both the sftp and all-in-one deployments mount /etc/sw/ssh from
<fullname>-sftp-ssh-secret, but no template created it, so a default
install could not start its pod and host keys only reached the server
when enableAuth happened to mount them elsewhere. Create the secret
with a generated ed25519 key, keeping whatever keys an existing secret
already holds. The sshPrivateKey default becomes empty: the file it
pointed at only exists when enableAuth mounts /etc/sw, and a configured
but missing key file is fatal to the server, while hostKeysFolder now
always has a key.
* helm: test SFTP host key generation and secret lifecycle
Template checks: keys render into the secret the deployments mount,
parse as PKCS#8 ed25519, differ between installs, and the render
carries no key material from the chart itself; existingSshConfigSecret
and all-in-one wiring covered. On the kind cluster, exercise the
secret lifecycle: a generated key survives upgrades, the key earlier
chart versions bundled is replaced, and operator-managed keys are kept
untouched. chart-testing now also installs with sftp enabled, where
the pod only becomes ready if the server loads the generated host
key.
* helm: treat a whitespace-only stored SFTP host key as missing
A whitespace-only secret value skipped regeneration and then rendered
an empty key file.
* helm: mount the SFTP host keys secret at the configured hostKeysFolder
The secret was mounted at a fixed /etc/sw/ssh, so a custom
sftp.hostKeysFolder pointed the server at an empty directory. Mount at
the configured path in both the sftp and all-in-one deployments, and
pin flag/mount agreement in the rendering tests.
* s3: track manifest blob ownership through multipart completion
Manifest blobs made three orphan paths. A partial fold that failed midway
kept its earlier batches on volume servers while the write fell back to
flat chunks; the fold now records each saved blob and deletes them on
error. A completion that failed after preparing left its fresh manifests
behind on every retry; the completion state now owns them and deletes
them unless a failed rollback left the version entry still holding them.
And a completed upload removes its parts metadata-only, which stranded
the part-manifest blobs superseded by flattening; those are collected
during flattening and deleted once the completion commits.
Two shared-chunk hazards nearby: the version-file rollback deleted its
data, destroying the still-registered parts (worse once manifests
resolve to inner chunks), and the idempotent-replay cleanup data-deleted
leftover parts whose chunks the live object references. Both are
metadata-only now.
* s3: trim chunk manifest comments
* s3: test manifest fold rollback and part range selection
The fold-with-rollback and the boundary-to-byte-range logic were only
exercised by hand against a live server; give both an injectable seam
and cover the fold, the below-threshold and SSE no-ops, the midway
failure deleting the blobs it saved, and offset-vs-legacy-index range
selection including indexes that no longer address the chunk list.
* erasure_coding: WriteDatFile takes the encode-time dat size for the shard block layout
* volume server: derive EC decode layout from the encode-time dat size, not the live extent
* erasure_coding: test decode after tail deletions shrink the live extent below a large-block row
* seaweed-volume: write_dat_file_from_shards takes the encode-time dat size for the shard block layout
* seaweed-volume: derive EC decode layout from the encode-time dat size, not the live extent
* seaweed-volume: test decode after tail deletions shrink the live extent below a large-block row
* erasure_coding: reject decoding with no data shards
* worker: record the encode-time dat size in the .vif
* erasure_coding: fall back to the shard-derived layout only when the encode-time dat size is missing
* erasure_coding: reject an ambiguous shard-derived block layout
* seaweed-volume: fall back to the shard-derived layout only when the encode-time dat size is missing
* seaweed-volume: reject an ambiguous shard-derived block layout
* filer: accept a WriteCondition on UpdateEntry, under the per-path lock
UpdateEntry was a bare read-modify-write: the precondition check, the
chunk garbage diff, and the store write could interleave with a
concurrent update to the same path. Take the per-path lock CreateEntry
already holds, and evaluate an optional CreateEntry-style WriteCondition
under it, failing with FailedPrecondition like expected_extended.
* filer: IF_CHUNKS_EQUAL write condition compares the stored chunk fid set
A chunk-preserving read-modify-write (tagging, setattr, copy-in-place)
races UpdateEntry's garbage diff: if a concurrent update empties the
chunk list first, the stale writer's commit resurrects fids that are
already queued for deletion, stranding the entry on a dead needle once
vacuum reclaims it. The reverse also holds: a writer that read an empty
chunk list can wipe chunks a concurrent update just added.
IF_CHUNKS_EQUAL guards both: the stored chunk fid multiset must still
equal what the caller read, order-independent, with an empty fids list
expecting no chunks. Absent entry counts as no chunks for CreateEntry
overwrites and transactions.
* filer: delete and append serialize on the entry path lock
DeleteEntry queues the entry's chunks for deletion and AppendToEntry
rewrites the chunk list, but neither held the per-path lock, so either
could interleave with a conditional update between its precondition
check and its write — a passed IF_CHUNKS_EQUAL would then resurrect
fids already on the deletion queue, or clobber a freshly appended
chunk. AppendToEntry keeps the cluster lock for cross-filer append
serialization; the path lock covers the local read-modify-write.
* filer: reuse lockPath in UpdateEntry lookup
s3: fold large chunk lists into manifest chunks on the direct write path
The S3 gateway uploads chunks itself and hands the filer a fully prepared
entry. On the routed write path (ObjectTransaction) the filer stores that
entry as-is, so a large PutObject or CompleteMultipartUpload persisted its
whole flat chunk list - a 900GB object carries 120k chunk references in one
entry. Manifestize on the gateway before the entry is written, the same way
mount, WebDAV, and filer.copy prepare theirs.
Multipart part boundaries also record byte offsets now: the stored chunk
indexes stop matching the entry once the list is folded, and partNumber
reads plus GetObjectAttributes prefer the offsets. Legacy index-only
records still work, with bounds checks instead of a possible panic.
Copy paths resolve a manifested source into data chunks before their
per-chunk copy loops - copying a manifest chunk raw would store its blob
as object data still pointing at the source - and a large copied list is
folded again on the destination. Completion likewise resolves manifest
chunks a part entry may carry (the filer folds an oversized UploadPartCopy
range) before rebasing part offsets.
Chris LuandGitHub
Lars LehtonenandGitHub
TJDawson10andGitHub
github-actions[bot]