* fix(decommission): move the free version of an expired tier version
Decommission only queued the lifecycle expiry of a transitioned version, so a source cleanup that ran first deleted the version itself and appended its tier free version after the entry's last pass. The post-check then failed the decommission until an operator re-ran it.
Apply the expiry of a transitioned version synchronously, as rebalance does. Its free version is on the source before the cleanup preflight, which reports the entry changed, and the retried entry moves the record to a target in the same pass. An expiry that cannot be applied leaves the version live, so it moves like any other version.
Refs rustfs/backlog#2873
* fix(decommission): move a delete-all expired tier version live
Applying a delete-all expiry inline made its prefix delete wait for the bucket lifecycle write lock behind the entry's own read lock until the lock timeout, blocking every write in the bucket meanwhile. Move such a transitioned version live and leave the delete-all to lifecycle on the target.
Cover a genuinely unapplied expiry and a free-version worker that consumes the source record before the entry moves it.
Refs rustfs/backlog#2873
* fix(decommission): defer source cleanup quorum failures
A quorum failure while deleting the source of an already copied entry
was fatal and marked the whole pool failed, although the copied source
is still in place and an uncommitted delete transaction is settled by
delete recovery. Carry the cleanup error typed and resolve a quorum
failure as Deferred: the entry is replayed at the end of the round, and
a deferral that persists takes the existing capacity-blocked pause path
instead of a terminal failure. Other cleanup errors and quorum errors
from other stages stay terminal.
Add set-level regressions for parent and child keys (`a/b`, `a/b/c`):
parent cleanup leaves every child file untouched, and parent cleanup
reaches quorum while the child is cleaned up or rewritten concurrently.
* test(ecstore): keep source cleanup errors typed in prefix cleanup tests
Return SourceCleanupError from the test helper instead of formatting it
into Error::other, and shrink the error(format!) ratchet baseline for
core/pools.rs, which lost one call site in the cleanup failure change.
* fix(decommission): retry cleanup quorum failures and count replays once
Retry a source cleanup quorum failure in place, up to three attempts
with the existing source-cleanup backoff, before deferring the entry.
Carry the versions an entry already counted on its deferral, so the
end-of-round replay does not add a fully counted entry to
items_decommissioned and bytes_done again.
A deferral that persists pauses the pool through the existing deferred
drain path. When source cleanup caused it, persist a reason led by
"source cleanup blocked" without the carrier error's free-drive wording,
and report waitingReason "source_cleanup" instead of "capacity".
Add a source_cleanup test fault stage and a routine test that drives a
cleanup quorum fault through deferral and replay to completion.
* fix(decommission): wait seconds between cleanup quorum retries
Give the in-place source cleanup quorum retry a fixed floor of one and
then two seconds, with the existing jitter on top, so it can ride out a
drive reconnect or health-check blip instead of retrying within a few
hundred milliseconds. The wait stays bounded and cancel-safe.
Find a source cleanup pause reason through stage wrappers, as the
capacity-blocked check does, and add a routine test where cleanup keeps
losing quorum: the decommission pauses with a source cleanup reason, the
bucket stays queued and the source stays in place.
Document the additive waitingReason value source_cleanup and its
compatibility with older nodes reading the same pool.bin.
* fix(decommission): never retry a lost source cleanup fence
The cleanup retry classifier used is_quorum_error(), which also matches
NamespaceLockQuorumUnavailable. The delete commit fence reports a lost
namespace lock or mutation fence with that variant, so a lost fence was
retried in place and only stopped later at the pre-cleanup fence check.
A lost fence means this worker no longer owns the key; it must stop
without retrying. Retry and defer only drive quorum failures:
ErasureReadQuorum, ErasureWriteQuorum, InsufficientReadQuorum and
InsufficientWriteQuorum.
---------
Co-authored-by: Hauser <housemecn@gmail.com>
* fix(decommission): let a failed start roll back its own activation
A fresh decommission start whose peer reload failed could not roll back:
the pool metadata stale-write guard from #6604 refuses any save that
clears a persisted active or queued decommission, so the pool stayed
active with no worker until the watchdog or a leader restart picked it
up.
The rollback save now carries the activation that start committed. The
guard exempts a pool only while its persisted record still equals that
activation field for field, so a worker checkpoint or a later start
keeps the rejection. A canceled record is not an active decommission
and is cleared by the rollback as well. Every other save path passes no
activation and is guarded as before.
A peer may already have reloaded the activation and started a worker
before the rollback. Saves are now fenced by decommission generation: a
write carrying a start_time is rejected when the persisted record
belongs to a later start, or has no decommission state and was updated
after that start_time, so such a worker cannot make the pool active
again. Its queued promotion and progress checkpoint also adopt the
persisted record and stop the worker.
Refs rustfs/backlog#2873
* fix(decommission): fence terminal writes of a withdrawn decommission
Terminal records carry no start_time, so the generation fence did not
cover them: after a start rollback, a peer worker of the rolled back
generation that failed (for example on a bucket completion) could still
persist a failed record over the restored pool. The merge now rejects a
failed, canceled or complete record over a persisted record without
decommission state; every attempt persists decommission state first, so
no legitimate path writes a terminal record there.
Peers also drop a withdrawn decommission sooner. A pool metadata reload
replaces a running decommission with a newer record that withdrew it,
even under an active worker; bucket completion adopts such a record like
promotion and progress checkpoints do; and a withdrawn progress
checkpoint stops the worker instead of logging a failed save and moving
more entries.
Tests cover the peer failure mark, a rollback retry that sees a worker
checkpoint land between attempts, and the reload withdrawal. A heal
fixture that wrote complete records over a fresh pool.bin now persists
the active attempt first.
Refs rustfs/backlog#2873
* fix(decommission): fence peer writes after a retry start rollback
A retry start replaces a failed or canceled record that keeps its
unresolved ledger, so its rollback restores that terminal record rather
than clearing the pool. The withdrawal checks only recognized a record
without decommission state, and the generation fence ignored a persisted
record without a start_time. A peer worker of the rolled back retry
could therefore checkpoint the pool back to active.
The generation fence now compares the persisted last_update whenever
the persisted record has no start_time of its own (cleared, terminal or
queued). The progress checkpoint also checks the generation against the
record read under the pool metadata lock, as bucket completion already
does, and the save-path withdrawal check treats a terminal record newer
than the running view as a withdrawal, so the peer adopts it and stops.
A reload still never lets a newer terminal record replace an active
worker's view, so after a retry's rollback a peer worker stops at its
next bucket completion or progress checkpoint rather than on reload.
Tests cover the retry rollback against a peer checkpoint and direct
write, and a peer bucket completion after a fresh start's rollback.
Refs rustfs/backlog#2873
* fix(decommission): keep a retry rollback's restored record intact
After a retry start rolls back, the restored record is failed, so a
peer worker's own failure mark passed every merge guard and replaced
it. A failure mark now saves through save_current_pool_meta_checked,
which runs the withdrawal check against the record read under the pool
metadata lock and adopts that record instead of writing.
The withdrawal check compared the restored record with the worker's
in-memory last_update, which a rejected save had already raised, so the
peer never adopted the rollback and the watchdog kept resuming it. It
now compares against the running generation (start_time), and a
rejected unresolved-entry save undoes its in-memory change.
save_current_pool_meta_for_decommission_start is also compiled for the
test-util feature again; a test-util helper still calls it, which broke
the workspace clippy and Connect builds.
Refs rustfs/backlog#2873
* fix(decommission): keep an unresolved entry when its own save fails
The rollback of a rejected unresolved-entry save also dropped the entry
on transient failures such as a quorum or lock-fence error, so the
following failure mark persisted a failed record without it and an
operator could clear a decommission that skipped an unreadable object.
Keep the entry and undo only the last_update bump, which is all a
withdrawn worker needs to still see the withdrawal.
Refs rustfs/backlog#2873
* fix(ecstore): leave no source free version for a moved tier version
Data-movement source cleanup deletes each source version separately and
never set skip_tier_free_version, so every transitioned version that
decommission or rebalance had already copied to a target pool left a new
tier free version on the source. Decommission then failed its post-check
for any pool with tiered objects; rebalance left the record behind. The
record claims the remote object the target copy still reads and stays
blocked behind that live reference.
Decommission and rebalance now pass the identities of transitioned
versions whose copy committed on a target, and the cleanup skips the
free version for exactly those. A transitioned version that did not
move, such as one lifecycle expired during the move, still gets its free
version because that record is its only remote-delete owner.
Refs rustfs/backlog#2873, rustfs/backlog#2872
* fix(ecstore): keep the free version of a superseded moved source
Data movement treats a source null version as complete when a newer
unversioned object already sits on the target, without copying it.
That outcome returned the same Ok as a committed copy, so a restored
transitioned null version superseded this way was recorded as
transferred and lost its source free version, leaving its remote object
with no owner.
The copy paths now return DataMovementCopy::Committed or ::Superseded,
and decommission and rebalance record only committed copies of
transitioned versions. A superseded source keeps its free version.
Tests cover the superseded restored null version for rebalance and
decommission, a restored and a null transitioned version moving by
rebalance, an entry with one lifecycle-expired and one moved version,
the per-version skip with two versions of one key, and a directory
object. The tiered stale-target routine test from #8724 now expects the
decommission to complete. The ILM ownership contract and the
decommission notes describe the moved, superseded, expired and legacy
cases.
Refs rustfs/backlog#2873, rustfs/backlog#2872
* test(ecstore): build the decommission ILM fixtures without s3s
The superseded and expiry-mix decommission tests named s3s request and
lifecycle types directly, which added pools_test.rs to the ecstore
files that reference s3s and failed the footprint ratchet. Set the
restore request through ObjectOptions, and store the lifecycle rule as
bucket metadata XML and load it with get_expiry_configs, as the
decommission routine does.
* test(ecstore): pin restored decommission copies and multipart supersession
Decommission recorded a committed data copy of a restored transitioned
version without any test failing if the record were dropped; add a
routine-level decommission of a restored version that checks the post-
check passes, the source keeps nothing, and the target keeps the full
remote tuple. Add the multipart variant of the superseded restored null
version (rebalance), and a unit test that separates committed,
superseded and mismatched targets in the overwrite resume adjudication.
* fix(decommission): run the leader-side start to completion
A forwarded decommission start awaits `ECStore::decommission` inline in
the RPC handler. When the forwarding node's RPC times out, the stream is
reset and the handler future is dropped wherever it is. Dropped after
`pool.bin` records the pool as active but before the peers reload and
the worker is spawned, the start left the pool active with no worker and
peers on the old metadata, and a retry was rejected as already active.
An admin client that disconnects can do the same.
Run the start in its own task and let the caller await only its result.
The start then always reaches its end: it fails before persisting
anything, or persists the pool, has the peers reload and spawns the
local workers, rolling back if the reload fails. A retry still cannot
start twice: it waits on the start gate held by the running start and
either reports the busy gate or is rejected because the pool is active.
`decommission` now takes `self: &Arc<Self>` and spawns workers on that
store instead of looking up the global object store, which is the same
store in production and makes the path testable; the unused
`require_decommission_store` helper is removed.
Refs rustfs/backlog#2873
* fix(decommission): detach the leader-side start from its caller
9ac4c10de carried only the reproduction (the testable entry point, the
peer-reload pause point and the proof test). This adds the fix it
describes: `ECStore::decommission` spawns the start as its own task and
awaits only its result, so a caller that stops waiting, such as a
forwarded start whose internode RPC timed out, can no longer drop the
start between persisting the pool as active and having the peers reload
and the local worker start.
Also prove a retry while the first start runs cannot start the pool
twice, update the operator and architecture docs for the new behavior,
and shrink the error(format!) ratchet baseline for the removed helper.
Refs rustfs/backlog#2873
* fix(decommission): single-flight start, recheck pool state under gate
Review of 3a5276c5b:
- A start re-checks the pool state under the start gate before it
cancels routines, so a start racing another one can no longer cancel
the worker the other start just spawned.
- Only one start runs per instance at a time: another request, including
the retry of an operator whose request timed out, is rejected as
already running instead of stacking a second detached start and its
preflight heals. The start runs through `run_detached_mutation`, and a
failure nobody waits for any more is logged.
- Tests: the proof test pins its drop after the save and requires a
worker that persists a cancellation; the retry test really drops the
first caller; new tests cover the under-gate recheck and a peer reload
failure after the caller is gone. That last one shows the start
rollback cannot clear a fresh start because the pool metadata
stale-write guard refuses it; the gap is asserted and documented and
tracked in rustfs/backlog#2873.
- The start tests run on their own large-stack thread: the io_uring lane
(`uring_`, which also selects `during_`) runs them with libtest's 2 MiB
default stack, where the proof test overflowed.
- Docs: single-flight starts, status and cancel during preflight, the
rollback gap and its recovery, and the release requirement.
Refs rustfs/backlog#2873
* fix(decommission): turn away local start retries at the start slot
Second review of 29253430a:
- The leader-local admin pre-check refreshed rebalance metadata under
the start gate before the start slot was checked, so a local retry
waited out its 20 s gate budget behind a running start (which holds
the gate through its peer reload) and then got a 503. The pre-check
now checks the slot first and returns `DecommissionAlreadyRunning`,
which the handler reports as a 400 that keeps the reason.
- Prove the reload-failure path really ran in the detached start (the
injection is consumed), add a test that a fresh start's failed
rollback reports both the reload error and the stale-write refusal,
and assert an abandoned start stays tracked for shutdown.
Refs rustfs/backlog#2873
* fix(ecstore): bound decommission target gate contention
Target capacity gate contention during pool decommission escalated a
per-object transient into a durable bucket pause: each contended object
failed its bucket entry, which re-ran the whole bucket listing and amplified
attempts on the same objects.
- centralize the decommission capacity failure classification so gate
contention, benign contention and fatal failures are decided once
- retry target-gate contention inline (bounded, jittered) before a mutation
is admitted, covering put, part, complete, new-multipart and abort
- defer contended entries to the end of the round instead of failing the
bucket entry, and require the deferred set to drain before a set completes
- treat a missing object or version, an overwrite and a superseded upload id
as benign contention that is neither counted as a failure nor escalated
- use full-jitter exponential backoff, capped, for decommission retries
- expose the capacity pause reason, the waiting reason and a cumulative
pause count in the admin pool status, plus gate-retry and per-object
attempt metrics
Related: rustfs/backlog#2644
* fix(ecstore): correct deferred replay and metadata compatibility
Agent-instruction and architecture docs had drifted from the code:
- CLAUDE.md: slim to commands + pointers; fix wrong claim that
`make pre-commit` is the full pre-PR gate (that is `make pre-pr`);
drop stale pre-#3929 file paths and merged bug narratives
- AGENTS.md: drop dead `rust-refactor-helper` skill rule and the
hand-maintained (already stale) scoped-AGENTS index; link
architecture docs from Sources of Truth
- .github/AGENTS.md: replace the outdated copied CI command matrix
with a pointer to ci.yml
- crates/AGENTS.md: merge duplicated Testing sections
- ARCHITECTURE.md: resolve the utils->config contradiction (edges are
removed), mark volatile counts as snapshots, fix a bad path
- docs/architecture: add README router; move one-shot plans/trackers
(rebalance-decommission phases, migration-progress ledger, PR
template) to docs/superpowers/plans with archive headers; fix stale
source paths in kept inventories (core/sets.rs, core/pools.rs,
store/mod.rs, startup_* split from #3671)
- docs/operations/tier-ilm-debugging.md: extracted tier debugging
playbook with corrected paths
- scripts/check_doc_paths.sh: new guard failing pre-commit/pre-pr when
instruction/architecture docs reference nonexistent file paths
- .claude/skills: add tier-debug and arch-checks repo skills;
.gitignore now keeps .claude/skills and docs/operations committable
Verification: ./scripts/check_doc_paths.sh,
./scripts/check_architecture_migration_rules.sh (both pass)
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Chris