Chain of fixes needed to get --force-snapshot soak to a real live PASS
(raft-ha-core-safety task 5.10a / ha-entity-snapshots task 7.1):
1. WAL segment rotation was gated on a hardcoded 64MB compile-time
constant; trim_wal_segments() never deletes the active (currently-open)
segment. At soak job counts a segment never rotated, so nothing was
ever physically trimmable and ha_evaluate_repl_action always chose
STREAM_WAL over SEND_SNAPSHOT for an empty-rejoining node. Fixed by
threading a wal_segment_bytes override through persist_config_t
(mirrors the --max-wal-records pattern: config.h default, persist.h
setter, persist.c 3 usage sites + minimum-floor enforcement, 6-point
CLI wiring through main.c/myconf.h/myapp.c/cluster.c) plus a new
--wal-segment-bytes CLI flag. run-ha-soak.sh's --force-snapshot mode
now sets WAL_SEGMENT_BYTES=4096 to force real rotation at soak scale.
2. This unmasked a genuine heap-use-after-free in INVOKE_MSG_HANDLER
(include/netio.h): the macro calls ep->msg_handler(ep) synchronously
(not via the vtask dispatcher, so the bug-1298 vtask ref/unref pairing
doesn't cover it) then writes ep->fsm_in.pkt = NULL. A stale vote-reply
handler chain (raft_on_vote_reply -> raft_cb_become_follower ->
endpoint_unref) could drop the endpoint's last ref and free it inside
the handler call, so the trailing write hit freed memory. This is the
first-ever live exercise of a node receiving a snapshot transfer while
mid-election. Fixed by holding a ref across the synchronous call, same
pattern task 5.8 established for the async path.
With both fixes: --force-snapshot soak now PASSES live (committed
divergence=0, job_loss=0, confirmed real "HA: snapshot" install on the
restarted node) — first genuine live proof of this scenario. 3-round
--chaotic regression clean. Full 217-test controller suite green
before/after every fix in the chain.
Builds on the delaymgr starvation fix (b9ad018) and walcompare
checkpoint-floor fix (f3c86a6) from earlier in this investigation.
Root-caused via raft-ha-core-safety 5.10a investigation: schedule_checkpoint()
(persist.c) fires correctly once wal_record_count crosses its threshold, but
its delay_task()-scheduled checkpoint_task_fn never executed in a live
multi-node deployment -- confirmed live via temporary diagnostic logging
(wal_record_count=100 max=100 logged, but the task itself never ran).
Root cause: delay_task_t carried TWO incomparable time representations --
`nextrun` (wall-clock time_t seconds, used by delay_task()) and `nextrun_ts`
(CLOCK_MONOTONIC timespec, used by delay_task_ms()). task_cmp()'s mixed-type
fallback compared them directly: an ms-task's nextrun is always 0, a
seconds-task's nextrun is a real epoch value (e.g. 1785600123) -- so 0
always sorted "less than" any seconds-task, regardless of actual due time.
ordq (src/lib/ordq.c) keeps its queue as an ascending sorted linked list by
this comparator, and delay_worker() only ever inspects the list HEAD
(ordq_peek with no matchfn returns q->head immediately, no scanning). Since
cluster_t.delaymgr is shared with ha.c's ~10 recurring delay_task_ms() call
sites (heartbeats, election timers, RDHUP polling), the head was
structurally almost-always occupied by an ms-task -- any delay_task()
(seconds) caller sharing that delaymgr, including schedule_checkpoint's
checkpoint_task_fn, could be starved indefinitely.
Fix: eliminate the dual representation entirely. Every task -- whether
scheduled via delay_task() (seconds) or delay_task_ms() (milliseconds) --
now stores one flat uint64_t CLOCK_MONOTONIC millisecond deadline
(nextrun_ms). delay_task() becomes a thin seconds->ms wrapper around
delay_task_ms(). task_cmp() and task_due() collapse to a single integer
comparison; delay_worker()'s reschedule is a flat integer add (no more
tv_sec/tv_nsec carry arithmetic, no more is_ms_task branching). Net -42
lines. Live-verified: after this fix, a forced checkpoint round produces a
real installed snapshot file (previously: schedule_checkpoint fired but the
task silently never ran).
controller_unit_test: full 58-group sweep, 0 Failures, 0 ASAN reports
(before and after this change) -- no regression, including every
delaymgr-touching group (test_demotion_convergence, test_ha_start_election_async,
test_ha_core_safety, test_ha_startup, etc).
Records this session's investigation: walcompare's per-member floor
fix (f3c86a6) and the --force-snapshot soak scaffolding (9f4292d) were
built specifically to close 5.10a with direct live evidence, but hit
a new, separate blocker -- schedule_checkpoint() fires correctly but
its delaymgr-scheduled checkpoint_task_fn never executes in a live
deployment, so no snapshot is ever installed to test against. Task
5.10a stays open pending a dedicated delaymgr investigation.
Adds --max-wal-records (mirrors the existing --wal-sync-ms pattern:
main.c argp option -> myconf_t -> persist_config_set_max_wal_records())
so a deployment can force the WAL-record-count checkpoint trigger to
fire on a job count a soak can actually reach, instead of the
production default of 10000.
Adds a new `--force-snapshot` soak mode (run-ha-soak.sh): submits
enough jobs to cross the configured threshold, kills a non-primary
node, wipes its data dir (a genuine empty rejoin), restarts it, and
asserts the cluster reconverges with zero walcompare divergence and
zero job loss. Fails loud (not a false PASS) if the restarted node's
log shows no actual snapshot install, so the round can't silently
report success without exercising the real scenario.
INVESTIGATION FINDING (not yet resolved, this commit does not close
5.10a or ha-entity-snapshots 7.1): live-tested with temporary
diagnostic logging (since reverted). schedule_checkpoint() correctly
fires once wal_record_count reaches the configured threshold
(confirmed: "wal_record_count=100 max=100" logged), but the delaymgr-
scheduled checkpoint_task_fn never executes afterward -- no checkpoint
ever completes, no snapshot file is ever written, in a live multi-node
deployment. This is a separate, deeper issue in the delaymgr scheduling
path (or its wiring for this specific call shape), not a bug in this
commit's CLI plumbing or soak scaffolding -- both are built, tested,
and confirmed working up to the point where the async checkpoint
should run. Needs its own dedicated investigation (likely src/lib
delaymgr, not persist.c) before --force-snapshot can produce a PASS.
This also has production implications beyond the test scenario: if
confirmed, WAL checkpoints may never fire automatically today, meaning
WAL segments are never trimmed via this path in a live deployment.
controller_unit_test: test_ha_startup 34/0, test_ha_core_safety 28/0,
test_persist_checkpoint 5/0, test_walcompare 17/0 (no regression).
Soak --selftest 17/17.
walcompare previously had no concept of a per-member compaction floor:
committed_lsn only capped the HIGH end (skip LSNs a member hasn't
caught up to yet), but the LOW end unconditionally expected every
member to hold raw WAL from LSN 1. A member that legitimately caught
up via a full-checkpoint snapshot holds no raw WAL at or below its own
checkpoint LSN by design (that range was entity-restored, not
WAL-replayed) -- so it was always flagged MISSING/divergent there,
regardless of whether the restore was correct. This is the root cause
ha-entity-snapshots task 7.1 diagnosed (committed_divergence=11 on an
otherwise-healthy soak) and raft-ha-core-safety task 5.10a's residual.
Adds member_snapshot_floor(): each member's own snap/ sibling
directory (matching the deployed <data>/wal + <data>/snap layout) is
scanned for the highest installed snap-<lsn>.{capnp,bin} file. Any LSN
at or below that member's own floor is excluded from its comparison
the same way an LSN above its own committed_lsn already is -- never
uniform/shared across members, since compaction is a per-member fact.
wc_status() gains a new "compacted" token distinct from "MISSING" so
the machine-readable report can tell the two apart.
Two new regression tests prove: (1) the floor correctly suppresses the
compacted range from comparison, and (2) it does NOT mask a genuine
divergence within the range a member still claims to hold. Caught and
fixed one bug myself before landing: the floor comparison was
off-by-one (< instead of <=), which the first test run surfaced
immediately (Expected 0 Was 1).
test_walcompare 17/17, test_ha_core_safety 28/0, test_config_repl
66/0, test_waldump 7/0 (all no-regression). Soak --selftest 17/17.
Confirms the live chaotic soak eliminates the divergence this spec's
Headline Finding hypothesized (LSN-9 committed-prefix divergence,
round-4 election stall) -- not by synthetic unit tests alone, but by
the task 7.2 acceptance soak (3 independent PASS runs, zero divergence,
zero job loss, 6/6 clean elections).
Documents the full resolution chain (Phase 2 payload loss -> Phase 5
election liveness, 5 layers -> Phase 3 snapshot boundary -> Phase 6
end-to-end + single-seam verification -> Phase 7.1 full regression
sweep) and reviews the final behavior against all four raft/raft.tla
safety invariants (Election Safety, Leader Completeness, State Machine
Safety, Commit Ordering/Durability). No residual divergence remains
open under this spec; two items explicitly routed to their owning
specs rather than closed here (dormant heartbeat-driver path, soak
harness's transient committedLsn=0 read).
This closes the last task in raft-ha-core-safety (18/18 sub-tasks).
The chaotic soak's walcheck()/walcheck_count() previously forced a
static `1 + 2*njobs` boundary onto walcompare, which doesn't account
for per-election NOOP WAL entries and overshoots the real tip --
producing spurious committed_divergence counts on an otherwise-clean
cluster (verified via waldump: byte-identical, gap-free WALs across
all 3 nodes despite a reported divergence=7).
Replaces the static estimate with the LIVE, authoritative committedLsn
(via the existing probe_primary client probe), applied as an explicit
--committed boundary only when the cluster is genuinely converged
(cluster_committed_converged) -- falling back to walcompare's own
auto-derive (min CRC-intact tip) when not, so a legitimately-lagging
just-restarted standby never gets a false "MISSING" verdict mid-chaos.
Pure auto-derive alone can't catch a member that permanently lost a
committed entry (its own low tip just becomes the shared boundary),
so this restores exact-boundary detection at every genuinely-settled
instant while keeping the auto-derive fallback for in-flight kills.
Also fixes a live-caught edge case in the same mechanism: a transient
ctrlstatus race can report committedLsn=0 for the primary right after
reconvergence; live_committed_boundary() now rejects a bare "0" as
untrustworthy and falls back to auto-derive, instead of handing
walcompare a vacuous --committed 0 boundary that would check nothing.
Selftest 17/17. Three independent fresh 6-round chaotic soaks (own
run + reviewer's reproduction + post-guard-fix re-run) all PASS:
committed_divergence=0, committed_job_loss=0, 6/6 elections, 21/21
jobs present throughout. No production C code touched.
A full 69-group sweep of controller_unit_test (this spec's harness had
never been run end-to-end before) found 2 deterministic failures,
both test-only staleness against already-shipped, already-approved
production behavior from earlier in this same spec:
- test_ha_vote_gate: test_pre_vote_catch_up_denies asserted the
pre-task-5.3 unconditional in_catch_up PreVote deny. Task 5.3
(bf88681) deliberately gated that deny on primary_ctrl_fd>=0 &&
ha_heartbeat_is_fresh to fix a real round-4 election-stall — this
test's exact scenario (stale heartbeat, primary_ctrl_fd=-1) is the
case 5.3 intentionally flipped to grant. Split into
test_pre_vote_catch_up_grants_when_primary_unreachable (matches the
current contract) and test_pre_vote_catch_up_denies_while_primary_live
(preserves the original protective intent, correctly scoped).
- test_egress_filter: test_live_submit_tags_submit_lsn called the
JOB_NEW persist_job_hook_fun AFTER_EVENT hook directly, but task 5.9
(15be5d9, bug-1306 fix) made that hook's WAL_MODE_WAIT branch
notify-only, moving the real WAL append + commit barrier into
persist_submit_wal_and_commit(). 5.9 relocated 4 equivalent tests in
test_persist_hooks.c but missed this one. Now drives
persist_submit_wal_and_commit() directly, mirroring that pattern.
No production code changed. Full 69-group suite: 0 Failures, 0 ASAN
reports (independently reproduced 3x: initial sweep, implementer,
reviewer).
The term-blind ha_apply_single_entry direct-persist branch (carried
forward from task 2.4) is unreachable with a real non-zero term in any
current production path: its only trigger, raft_send_append_entries,
has zero production call sites (library simulator only). Both real
senders (repl_broadcast, ha_stream_wal_to_standby) always emit a
non-zero term for a present LSN. Documents this in ha.c and design.md
rather than structurally retiring a dormant path (YAGNI), with an
explicit flag for future re-evaluation if a heartbeat-driver task ever
wires up raft_send_append_entries.
Adds two regression pins: restart preserves (term,lsn) identity
(replay_wal already reconstructs the term index via record_entry_term),
and real senders never produce the term-blind wire shape.
test_ha_core_safety 28/0, no regression, no ASAN reports.
Integration test proves the guarantee tasks 2.3/2.4 hardened in
isolation (synthetic fault injection on the streamer) also holds
end-to-end over two real persist_t/WAL instances through the real
production ingest seam (ha_apply_wal_catchup_entry). Covers live
in-term replication and cross-term catch-up: no term==0 sourced for a
live entry, no non-empty entry ever lands as PLEN=0, standby persists
the entry's stored term (not the AE header's leader-current-term,
the bug-1067 split), and walcompare reports a clean committed prefix.
test_ha_core_safety 26/0, no regression, no ASAN reports.
Adds checkpointTerm (@6, append-only) to the SnapshotTransfer wire
message. The leader stamps it via persist_entry_term(p, checkpoint_lsn)
when sending a snapshot; the standby reconciles it into its own
per-LSN term index (persist_reconcile_checkpoint_term) right after
persist_load_snapshot succeeds, before applied_lsn advances. This
authoritatively overwrites any stale local term a divergent
pre-snapshot entry left at the checkpoint LSN, closing the gap task
3.1's RED test proved (Expected 7 Was 3 -> now passes). Transport
stays persist_t-free; the reconcile lives in persist.c per the
existing dependency direction. Decode is auto-generated via capnp
mapname; only the hand-rolled encode path (request.c/typemap.h, needed
for the raw snapshot payload) required a matching manual update.
test_ha_core_safety 25/0, test_config_repl 66/0, no ASAN reports.
ha_handle_snapshot_chunk/persist_load_snapshot advance checkpoint_lsn,
current_lsn, and applied_lsn on snapshot install but never touch the
per-LSN term index, so a divergent pre-snapshot entry's stale local
term survives the install unreconciled. New test proves the gap
(Expected 7 Was 3); fix lands in task 3.2.
persist_recover's local snapshot load passed NULL for out_entities_restored,
so a rolling-upgrade restart against a pre-feature checkpoint had no
observability parity with the wire receive path's (task 6.1) INFO log.
Mirror the same conditional logging here; no behavior change, WAL replay
in Step 3 still applies unconditionally either way.
Live-verified in a 3-node cluster: planted a hand-crafted 16-byte legacy
sentinel as ctrl-1's own checkpoint file, restarted all 3 — no crash, INFO
log fired, ctrl-1 was elected primary serving the full 37-job list, and
walcompare confirmed all 3 members' committed prefixes (LSN 1..49) are
byte-identical.
10-round chaotic soak: committed_job_loss=0 (core goal confirmed fixed),
committed_divergence=11. Root-caused to walcompare having no per-member
compaction-floor concept (uniform --committed N for all members), which is
exactly raft-ha-core-safety task 3's unimplemented snapshot-boundary
reconciliation, not a gap in this spec's boundary. Also: this run never
exercised an actual snapshot install (start_node doesn't wipe data dirs) —
forcing a real empty-rejoin round is still outstanding for a true signal.
ha_handle_snapshot_chunk now logs whether a received checkpoint was
fully entity-restored or was a legacy position-only file (INFO level,
framed as an expected rollout state -- never an error, per design.md).
The rc-gates-applied_lsn control flow is byte-for-byte unchanged from
task 5.2 (confirmed by review) -- this task is purely additive
observability, matching design.md's own note that this call site
"needs no change" once persist_load_snapshot's gate is correct.
New test drives the legacy path end-to-end through the real production
handler (rc==0, file renamed, applied_lsn advances to the legacy LSN).
This completes the integration layer -- write path (jobs+queues),
restore-before-advance read path, and the production wiring are all in
place and independently reviewed. Only validation tasks remain (7.1
chaotic soak, 8.1 rollback compatibility).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JDYFbrgyDHbvNYcxApKhrX
Replaces the naive task-5.1 stopgap with the real restore contract:
detects legacy vs entity-bearing checkpoints by remaining file size;
restores every job then every queue status via restore_job/
restore_queue_state, in file order, with bounds-checked length-prefixed
reads; only after every entity restores successfully does
current_lsn/checkpoint_lsn commit to the checkpoint's position. Every
failure path (bounds check, decode, restore error, unrecognized
format_version) returns before that single commit point -- position is
provably untouched on failure, never partially advanced.
Also fixes the pre-existing restore_job decode leak (job.c): job_init
deep-copies submit (never nulled) and MOVE_PTRs allocs/pidinfo/
pend_reasons (nulled on the one path that runs) -- free_Job(&j) is safe
unconditionally on every reachable return path (EEXIST early-return,
create_job success, create_job failure), verified against the exact
generated free_*_ptr null-check guards.
Two pre-existing test_config_repl.c wire-protocol tests needed their
synthetic garbage snapshot bodies updated to a well-formed empty-entity
tail, since load_snapshot now genuinely parses past the legacy header
(previously ignored); their actual assertions (chunk-reassembly
correctness) are unchanged.
This is the exact function/contract that caused task 2.2's rejection
(position advancing before restore validated). Independent review
traced the memory-safety fix through the generated capnp code field-by-
field, confirmed the test_config_repl.c change doesn't weaken any
assertion, and confirmed restore_queue_state's ENOENT-aborts-the-whole-
checkpoint behavior is faithful to design.md's unconditional contract
(flagged as a watch-item for task 7.1's soak, not a defect).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JDYFbrgyDHbvNYcxApKhrX
Extends persist_load_snapshot's public signature with an
out_entities_restored bool*, per design.md's documented contract. The
static load_snapshot function -- the real restore logic -- is completely
untouched (still today's watermark-only implementation, byte-identical
to before this task); only a thin wrapper was added that naively maps
*out_entities_restored = (rc==0), deliberately scaffolding so the 3 new
RED tests fail for genuine reasons rather than accidentally passing.
ha_handle_snapshot_chunk's sole call site gets a mechanical
declare-and-pass update -- confirmed inert, never reads the new
out-param (task 6.1's job).
Three tests against persist_load_snapshot directly (unlike prior tasks'
tests, which deliberately stayed on the raw-byte-parsing side to avoid
this exact function): restore-before-advance (fails at the job-lookup
assertion -- nothing is ever restored today), corrupted-entity-leaves-
position-unchanged (fails at rc!=0 -- today's code ignores the corrupted
section entirely, since load_snapshot never reads past the legacy
header), legacy-file-reports-position-only (fails at entities_restored==
false -- the naive wrapper claims restoration for every rc==0 case,
including legacy files).
This is the exact function/call site that caused task 2.2's rejection
(a live position-advances-before-restore-validates hazard). Independent
review verified load_snapshot's body is byte-identical to commit
d52eb94, the wrapper does nothing beyond the naive mapping, and ha.c's
change never reads the new out-param anywhere -- no repeat of that
mistake here.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JDYFbrgyDHbvNYcxApKhrX
rotate_snapshots now parses each candidate file's LSN and skips deletion
via a new snapshot_lsn_in_flight() check against
p->cluster->ha.snapshot_inflight_lsn[] (task 4.1) before unlinking --
protected files are retried on the next rotation pass once the transfer
completes and the LSN clears.
Checks LSN directly (nonzero + matches) rather than gating on
snapshot_inflight[]==1 first, per task 4.1's own note: the LSN clears
before the flag in every completion/failure path, so a stale flag=1 can
momentarily coexist with an already-cleared LSN=0. A checkpoint LSN is
never 0, so this is race-free without needing the flag at all.
p->cluster==NULL (standalone lib-level tests) means no HA layer exists to
protect anything.
Task 4.2's RED rotation-guard test now passes; full regression sweep
(test_persist_checkpoint, test_persist_hooks, test_commit_barrier_reply,
test_commit_durability, test_e2e_hooks, test_hooks, test_config_repl,
test_ha_liveness) all clean. Both controller_unit_test and production
xce_controller build clean; nm confirms only the real production helper
(snapshot_lsn_in_flight) appears, no test-only symbols leaked in.
Executed directly (no subagent dispatch) per the hybrid execution-mode
decision.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JDYFbrgyDHbvNYcxApKhrX
A checkpoint file whose position matches an in-flight peer transfer
(task 4.1's snapshot_inflight_lsn[]) must survive rotate_snapshots even
when snap_keep_n would otherwise delete it. Genuinely fails today (file
removed, access() returns -1) since rotate_snapshots has no awareness of
in-flight transfers yet.
Executed directly (no subagent dispatch) per the hybrid execution-mode
decision -- small, single-file, isolated test.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JDYFbrgyDHbvNYcxApKhrX
Split remaining tasks by risk after a mid-session spend-limit hit: small/
isolated tasks (4.2, 4.3, 7.1, 8.1) execute directly with self-review;
load_snapshot restore (5.1/5.2, the exact function that caused task 2.2's
rejection) and the production wiring (6.1) keep the full implementer +
independent-reviewer subagent cycle, since that's where adversarial review
already proved its worth and where the remaining risk is concentrated.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JDYFbrgyDHbvNYcxApKhrX
Adds snapshot_inflight_lsn[MAX_HA_CANDIDATES] to ha_state_t, alongside the
existing snapshot_inflight[] flag it mirrors exactly: set to the
checkpoint's LSN at ha_snapshot_send_dispatch before a transfer's first
byte goes out, cleared to 0 at every completion/failure path
(snapshot_lane_fn's normal completion, and all 3 early-return failure
sites: malloc failure, slowlane dedup-refuse, slowlane submit failure).
This is the foundation task 4.3 will consult from rotate_snapshots to
avoid deleting a checkpoint file a peer is still mid-transfer reading
from -- the one genuinely new retention guard this feature needs (the
existing write_snapshot-before-trim_wal_segments sequencing already
satisfies the WAL-trim side of Requirement 3).
Proven via a new gated unit test (test_ha_liveness.c) that observes the
tracked LSN mid-flight while the send body is parked at a test gate, not
just before/after -- confirms the lifecycle ordering, not just the
endpoints.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JDYFbrgyDHbvNYcxApKhrX
write_snapshot replaces the queue_count=0 placeholder with a real
collect_queue_frames helper: enumerates every queue in the live table
under nested table+per-entity RTTI_LOCK_READ (no retention filter --
table presence already implies config-provisioned existence, unlike
jobs), writes each queue's (name, status) as a plain fixed record (no
capnp -- matches restore_queue_state's exact contract). Any failure
anywhere still discards the whole checkpoint, including any
already-collected job frames, no partial file at the final path.
Entirely contained to persist.c -- load_snapshot and persist.h remain
byte-identical to before this task, honoring the boundary lesson task 2.2
established (queue restoration stays task 5.2's scope).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JDYFbrgyDHbvNYcxApKhrX
Queue's (name, status) round-trip through write_snapshot's raw on-disk
bytes -- a plain fixed record (no capnp; queue definitions are
config-owned, only runtime admin status is WAL/snapshot-tracked, per
design.md's corrected queue model). Fails today (Expected 4 Was 0) since
write_snapshot still hardcodes queue_count=0 from task 2.2's placeholder.
Does not call persist_load_snapshot, following task 2.2's established
raw-byte-parsing pattern.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JDYFbrgyDHbvNYcxApKhrX
write_snapshot now Cap'n-Proto-encodes every retained job under nested
table+per-entity RTTI_LOCK_READ (table lock alone does not exclude a
concurrent per-entity writer -- design.md's central correctness finding),
appends length-prefixed job frames after a new format_version marker, and
fixes a previously-unchecked fdatasync return. Any failure anywhere in the
path (encode, write, sync) discards the whole checkpoint -- no partial
file at the final path, proven by a real mid-encode-loop fault-injection
test (forces one specific job's encode to fail among N, asserts nothing
survives).
Also fixed two pre-existing bugs in task 1.1's test fixture: terminal-state
seeded jobs need a non-zero finish_time or job_should_retain's retention
window silently excludes them from any checkpoint; the fixture needs
xstring_table_init or every seeded job's submit/start_info silently
decodes as NULL.
Round 1 of this task also touched load_snapshot/persist.h to make the
round-trip tests pass end-to-end, but review caught a live hazard: it
advanced current_lsn/checkpoint_lsn BEFORE the entity-decode loop ran, so
a decode failure still left the position advanced -- reachable today
through the unmodified ha_handle_snapshot_chunk production path, not
hypothetical. Remediated by fully reverting load_snapshot/persist.h to
byte-identical pre-task-2.2 state and rewriting the round-trip tests to
parse write_snapshot's raw on-disk bytes directly (read_Job/decode_Job)
instead of round-tripping through persist_load_snapshot -- proves the
write side without depending on load_snapshot, which stays task 5.2's
scope entirely, not started here.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JDYFbrgyDHbvNYcxApKhrX
Three tests against the current sentinel-only write_snapshot/load_snapshot:
job content round-trip and concurrent-mutation-not-torn genuinely FAIL
(RED) since no entity content is written today; write-failure-discard
already passes (write_snapshot's open()-failure path already discards
correctly) -- flagged transparently as a GREEN-from-day-one regression
guard rather than faked RED, per real fault injection (directory removal
-> genuine ENOENT).
Carries forward two notes for task 2.2: fdatasync's return value is
currently unchecked in write_snapshot (persist.c:1505), and a mid-encode-
loop fault-injection test should be added once the per-job encode loop
exists.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JDYFbrgyDHbvNYcxApKhrX
Adds the ha-entity-snapshots Kiro spec (requirements, design, research,
tasks) for task 5.10a / raft-ha-core-safety: making the HA checkpoint an
actually-restorable copy of committed job/queue state instead of a
position-only sentinel, closing the residual chaotic-soak divergence gap.
Task 1.1 (foundation): reusable test fixture snapent_make/snapent_teardown
in a new test_persist_snapshot_entities.c, seeding N jobs (varying state)
and M queues (varying status) into a real persist_t+cluster_t for later
write/load/rotation-guard tests to build on. Build clean; new suite 1/1;
regression spot-checks (test_persist_hooks 19/19, test_commit_barrier_reply
22/22) clean.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JDYFbrgyDHbvNYcxApKhrX
A dead primary's control fd RDHUPs faster than an election settles. The 5.6
guard only drops an RDHUP while the role is instantaneously CTRL_ELECTING/
prevote; an RDHUP landing in the gap after a lost election (role briefly back
to CTRL_STANDBY) re-triggers a fresh ha_begin_failover -> PreVote -> real
election (term++). One bump per RDHUP -> unbounded escalation under a sustained
cadence (observed round 4: 48 begin_failovers, terms 6->24, round overran the
harness window).
Fix: ha_state_t.last_failover_ms (atomic) stamped at ha_begin_failover;
ha_on_primary_ctrl_fd_rdhup drops an RDHUP within 2*election_timeout_ms of the
last trigger (closes the dead fd, returns) so the election's own jittered retry
(5.7) runs without the external RDHUP cadence adding term bumps. First trigger
never debounced; genuine loss still fails over, at most once per window.
Validation: unit-proven — test_rdhup_within_cooldown_is_debounced (recent
last_failover -> stays CTRL_STANDBY) vs test_rdhup_single_node_reaches_primary
(last_failover=0 -> CTRL_PRIMARY) is a differential proof. Soak: 10/10 elect,
terms 4->14 bounded, begin_failover 14/14/12 (was 48), committed_job_loss=0, no
regression. The stochastic storm did not recur this run (debounce fired 0x), so
soak-reproduction of the storm is still pending; the mechanism rests on the
unit test. Election suites green (rdhup 8/0, prevote 12/0, core_safety 24/0,
election 28/0, failover_prevote 6/0).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JDYFbrgyDHbvNYcxApKhrX
The round-4 term ladder is driven by repeated RDHUP->begin_failover->
PreVote-won->real-election(term++) cycles from the dead primary's socket, not
a behind-node vote loop. 5.6's instantaneous 'suppress while electing' guard is
insufficient under a sustained RDHUP cadence. Fix direction: per-failover
cooldown/debounce + hold the failover latch across the full prevote->election->
settle window. Needs a deterministic dead-primary-cadence repro.
Their WAL_MODE_WAIT hook (wal_hook_append_classified) does local-append only
and carries no durability verdict through the hook return, so there is nothing
for a later hook to clobber. JOB was unique in running the full commit barrier
inside the hook. Their commit barrier is the separate ha-commit-barrier-rollout
handler barrier — verify independently, do not mechanically extend the JOB fix.
freshness-deny escalation loop under a just-restarted-survivor overlap: a
behind-log candidate is freshness-denied every term (18->24 observed), times
out, re-campaigns at term+1 -> ~18 escalations before convergence. NOT RNG
(seed is /dev/urandom); 5.7 jitter bounds retry rate, not the escalation.
Fix direction: a repeatedly-freshness-denied candidate must recognize it is
behind and defer/catch-up instead of laddering the cluster term. Needs a
deterministic 2-survivor repro before a fix (5.5/5.6/5.7 family).
raft-apply-decoupling item #1 (down payment). xce_run_hooks returns only the
LAST-registered hook's rc, so a later JOB AFTER_EVENT subscriber (sched/
apiserver) returning 0 clobbered persist_job_hook_fun's commit -ETIMEDOUT.
handle_jobsubmit then replied 'jobsubmit ok' for an entry that never reached
majority, and a new-term leader legally discarded it (committed-job-loss,
Commit Durability violation).
- persist.c/.h: new persist_submit_wal_and_commit(p, jlist) — the explicit raft
step (WAL append + majority-commit barrier) returning the commit verdict
directly. persist_job_hook_fun's WAL_MODE_WAIT branch is now a no-op.
- job.c handle_jobsubmit: call the explicit step, read commit_rc DIRECTLY,
reject on <0 before firing the now notification-only AFTER hook
(canonical commit -> notify -> reply).
- tests: +test_job_hook_new_is_notify_only regression pin; 4 hook tests
relocated to drive the explicit step. persist_hooks 19/0, e2e_hooks 3/0,
commit_barrier_reply 22/0, commit_durability 38/0.
Proven: 10-round chaotic soak committed_job_loss 5 -> 0 (same election code).
Residual soak FAILs (divergence catch-up lag + a round-4 election-escalation
tail) are NOT job loss — tracked as task 5.10.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JDYFbrgyDHbvNYcxApKhrX
10-round soak waldump+log bisect: the chaotic committed_job_loss is NOT the
catch-up/entity-snapshot prefix gap. All 3 WALs gapless/CRC-intact/full
payloads; lost job only on ctrl-1@LSN30 term9 vs term-10 NOOP on the majority.
The durable-majority gate worked (commit_submit -> -ETIMEDOUT), but
xce_run_hooks (hook.c:134-157) returns only the LAST-registered hook's rc, so a
later JOB AFTER_EVENT hook (sched/apiserver) returning 0 clobbered persist's
-ETIMEDOUT -> handle_jobsubmit replied jobsubmit ok for an uncommitted entry ->
term-10 leader legally overwrote it. Fix per raft-apply-decoupling item #1:
move the WAL append + commit barrier out of the hook into an explicit
handle_jobsubmit-driven raft step whose rc is read directly (down payment on
the full apply-after-commit inversion). Entity-snapshot gap is a separate,
still-real defect (not this soak's mechanism).
5.8 endpoint refcount fix validated: 182 unit tests/0 fail, 6-round ASAN chaotic
soak ASan-clean, all rounds elect. bug-1298 marked fixed.
The soak now reaches rounds 4-6 for the first time (every prior run crashed at
round 4) and FAILs the verdict on committed_job_loss=4/divergence=7. Evidence shows
this is the pre-existing bug-795/828/851 failover catch-up gap, NOT a regression:
one job missing from round 4 onward across 3 primaries; rounds 1-3 clean; ZERO
WAL/commit/catch-up code changed this session; ASAN clean. Routed to task 5.9 /
ha-commit-durability.
The io_worker freed an endpoint (endpoint_killer, cluster.c) on connection
teardown while the vtask dispatcher concurrently read ep->busy
(endpoint_is_task_blocked, netio.c) on a queued vtask — a heap-use-after-free
that ABORTed a controller mid-chaotic-soak (round 4), dropping the cluster
below quorum so no primary could be elected.
Replace the accumulation of ad-hoc ep-lifetime patches (bug-1282 lockless busy,
bug-1197 dispatch_inflight drain + deferred teardown vtask) with one principled
invariant: an endpoint lives until its reference count reaches 0.
- endpoint_t.refcnt (_Atomic) + heap_alloc flag; endpoint_get starts a heap ep
at refcnt=1 (the owner ref).
- endpoint_ref/endpoint_unref: the last unref fini's + frees a heap ep.
Embedded eps (client ctx->ep, heap_alloc=0) are never freed here — their
container owns them — so a vtask-balanced ref/unref pair is a safe no-op.
- vtask_enq takes a per-vtask ref on the bound ep BEFORE publishing the task to
the dispatcher; run_task / taskq_item_killer release it. So the dispatcher can
read ep->busy on any queued vtask without the ep being freed underneath it.
- every HEAP-ep teardown site (endpoint_killer, handle_endpoint_error, the
vote/prevote connect-fail paths, the vote-ep loser/teardown paths, the standby
reconnect/eviction paths) now drops its owner ref via endpoint_unref instead of
raw endpoint_fini+free. Several of those also leaked the ep struct before
(endpoint_fini-only); unref fixes the leak too.
Validation: controller_unit_test 182 tests / 0 failures (test_ha_vep_teardown
updated — its OFF-path test asserted the old struct leak, now fixed). 6-round
ASAN chaotic soak: ZERO AddressSanitizer reports, ctrl-3 no longer crashes at
round 4, all 6 rounds elect at bounded terms. (The pre-existing iomgr_fini ->
SSL_CTX_free abort under openssl 3.6 in the aggregate test binary is unrelated.)
Note: with the crash gone, the soak now reaches rounds 4-6 for the first time and
surfaces a SEPARATE, pre-existing committed-job-loss/divergence (the bug-795/828
failover catch-up gap; ha-commit-durability domain) — not a refcount regression
(ASAN clean; rounds 1-3 are loss-free).
5.7 election-timeout resonance FIXED + soak-confirmed: 6-round chaotic soak now
elects with bounded terms (rounds 1-3 at terms 4/5/6, not 25-30+), zero job loss.
The election-liveness layer is done.
Round-4 stall re-roots to a NON-election crash: ctrl-3 ASAN heap-use-after-free
in endpoint teardown — io_worker (endpoint_killer, cluster.c:813) frees an endpoint
while the vtask dispatcher reads ep->busy (endpoint_is_task_blocked, netio.c:1008).
ctrl-3 dies -> cluster drops to 1/3 -> no quorum -> no primary. Same async-io
dispatch surface as bug-1282/1197; UPSTREAM-owned (raft-drive-async / libxce).
Logged bug-1298; tracked as task 5.8.
The per-campaign election watchdog was armed with a FIXED election_timeout_ms
(raft_start_election -> on_schedule_timeout -> raft_cb_schedule_timeout). Under
cold-connection churn a just-restarted survivor opens a fresh TCP+SSL endpoint
per vote round, so its single needed grant can land a hair AFTER a constant
deadline on EVERY campaign -- a resonance: the campaign times out, the term
re-escalates, and the now-stale grant is dropped (DropStaleResponse). Forever.
Terms climb unbounded; the freshest survivor never tallies quorum. This is the
final round-4 chaotic-soak election-liveness stall.
Fix (Ongaro thesis 3.4): re-randomize the watchdog per campaign to the canonical
[T, 2T] interval in the HA adapter, so some campaign's deadline falls AFTER its
grant RTT and the grant lands in-window. Done in raft_cb_schedule_timeout so the
pure (sim-tested) raft library stays deterministic. Independent of
election_jitter_max_ms (which tunes only the sub-second startup/stagger jitter)
so it spreads even when jmax is small/0; never below the configured floor so
genuine-loss failover is not sped up past election_timeout_ms.
Preserves Election Safety (no change to term/vote/DropStaleResponse logic).
new ha_jittered_election_delay() helper + 2 unit tests (RED verified: stubbing
jitter to 0 fails test_5_7_campaign_timeout_is_randomized with the resonance
message; PIN guards floor + modulo-by-zero). Suite: 24 Tests / 0 Failures.
5.6's RDHUP-no-preempt guard (81cb09a) is unit-verified and slows the round-4 term
churn, but the soak still stalls — a 5th layer (task 5.7). The freshest survivor
ctrl-2 (15@t5) wins PreVote, enters the real election, and ctrl-3 GRANTS it
(vote_granted) — but ctrl-2 never tallies a majority: every grant arrives as
vote_reply_stale_drop (reply term < ctrl-2's current term). ctrl-2's term
escalates via repeated election timeouts faster than ctrl-3 (cold peer
connection, fresh SSL connect per vote, plus dead-ctrl-1 connect timeouts) can
reply, so each grant is one term stale -> dropped -> timeout -> term++.
Self-perpetuating; DropStaleResponse itself is correct Raft.
Adds task 5.7 (don't re-escalate the term while a vote RTT is outstanding /
warm-reuse the peer vote connection / wider jitter). 5 layers of the round-4
stall now peeled, 4 fixed (deadlock, segfault, advertised-term, RDHUP-preempt);
the cold-connection term-lag remains. Consensus SAFETY intact throughout.
Co-Authored-By: Claude <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019dbTHR9CsEYK9RsqmQpQpC
After 5.5 let the freshest survivor win PreVote and ENTER the real election,
round 4 still stalled: the dead primary's control fd keeps RDHUP'ing on a
deterministic ~15s cadence, and ha_on_primary_ctrl_fd_rdhup's 5.2 suppress only
catches CTRL_STANDBY — so an ELECTING node fell through to ha_begin_failover and
restarted a fresh PreVote, preempting its own in-progress real election. Every
cycle bumped the term (21->25->28->29...) and kept two survivors dueling in
lockstep, defeating the jittered election-timeout tiebreak.
Guard the RDHUP entry: if role==CTRL_ELECTING or a PreVote round is in progress,
close the dead control fd and return WITHOUT restarting failover — let the
in-progress round finish or time out (its election timeout is jittered, which
desynchronizes the duel so one survivor wins). A node not currently electing
still takes the normal 5.2-suppress / genuine-loss failover path.
RED+PIN in test_ha_core_safety.c (22 Tests / 0 Failures): CORE RED — CTRL_ELECTING
+ RDHUP must not start a new PreVote (Expected 0 Was 1 on baseline); PIN — genuine
loss (STANDBY, not electing) still fails over. Preserves Election Safety, 5.2
suppression, and genuine-loss liveness. Re-running 7.2.
Co-Authored-By: Claude <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019dbTHR9CsEYK9RsqmQpQpC
5.5's real root cause was a stale advertised lastLogTerm (c->persist.term vs
persist_entry_term), not the last_activity hypothesis — fixed in b8453b2 and
VERIFIED: with it, the freshest survivor now wins PreVote and enters the real
election (it could not before). But round 4 still does not complete: a deeper
layer (task 5.6) — under the dead-primary RDHUP cadence (~15s) the two survivors
duel in lockstep and the term escalates unbounded (21->25->28->29...) without
reaching majority. Adds task 5.6 (randomized failover backoff to desync the duel
/ don't re-trigger failover mid-real-election). Consensus SAFETY intact; this is
election LIVENESS under sustained churn. 4 layers of the round-4 stall now peeled
(deadlock, segfault, advertised-term fixed; dueling-candidate remains).
Co-Authored-By: Claude <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019dbTHR9CsEYK9RsqmQpQpC
The election cross-deny that stalled round 4 (after bug-1282+bug-1286 fixed): the
freshest survivor (e.g. ctrl-2 tip 15 @ term 5) could not win because it
advertised lastLogTerm = c->persist.term (the persist layer's RUNNING term,
stale/0 after a restart) instead of the term STAMPED ON its last WAL entry. So a
voter (ctrl-3, last entry term 4) saw the candidate's lastLogTerm <= its own and
raft_log_ok rejected the genuinely-fresher candidate -> every round timed out ->
permanent stall. This was a documented stopgap (ha.c "task 1.3" NOTE) and is the
bug-851 term-identity family.
Fix both advertise sites (real-vote raft_cb_send_vote_request + PreVote
raft_cb_send_pre_vote_request): lastLogTerm = persist_entry_term(&c->persist,
ha->raft.log_len) when log_len>0 — the SAME accessor voters use to compute their
own last-log term (on_get_entry_term -> persist_entry_term), so logOk is now
symmetric and the freshest survivor wins. Strengthens Election Safety/Leader
Completeness (only genuinely-fresher candidates are granted); the index source
(ha->raft.log_len, seeded from the recovered WAL tip) is unchanged.
Root-caused from a frozen round-4 soak: ctrl-2 (15@t5) vs ctrl-3 (11@t4); ctrl-3
denied ctrl-2 on logOk (mislabeled "eligibility"). Re-running 7.2 to confirm.
Co-Authored-By: Claude <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019dbTHR9CsEYK9RsqmQpQpC
6-round chaotic soak with both fixes: rounds 1-3 elect; round 4 has NO segfault,
NO fdatasync, NO iomgr deadlock — both fixes held. The round-4 stall is now the
pure election-liveness cross-deny that the crash/deadlock masked: the freshest
survivor (floor 15) is the legitimate winner, but the other survivor (floor 12)
denies it on eligibility because it still believes its dead primary is alive
(in_catch_up + heartbeat-not-yet-stale). Suspect 5.2's last_activity re-arm
keeping a node falsely-fresh after primary death.
Adds task 5.5 (release the heartbeat-fresh/in_catch_up deny on genuine primary
death within the failover window) as the remaining 7.2 blocker. bug-1282 and
bug-1286 are fixed and verified; consensus safety unaffected.
Co-Authored-By: Claude <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019dbTHR9CsEYK9RsqmQpQpC
Root-caused via the non-ASAN debug build + gdb: on a chaotic restart an inbound
endpoint dispatches a request (e.g. walCatchupEntry op 37) while ep->owner
(cluster_t) is NULL — not yet wired at accept, or cleared during teardown. The
handler did `c = ep->owner` then `ha_dispatch_gate(&c->ha, ...)`; with c==NULL,
&c->ha = NULL+offsetof(cluster,ha) = 0x3ff0, a garbage pointer ha_dispatch_gate
dereferenced (atomic_load &ha->role) -> SIGSEGV, dropping the node below quorum
and stalling the election.
Guard c==NULL after decode (so the decoded request is freed, no leak): log a
loud WARN and drop the request — there is no cluster context to dispatch through
and no c->myhost to address a reply. The controller stays alive (never
self-terminate on an operational condition); the peer retries once the endpoint
is owned. Backtrace: ha_dispatch_gate(ha=0x3ff0) <- myreq_handler cluster.c <-
vtask_entry <- run_task.
Second of the two round-4 blockers (bug-1282 deadlock already fixed). Re-running
the 7.2 chaotic soak with both fixes.
Co-Authored-By: Claude <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019dbTHR9CsEYK9RsqmQpQpC
/xce-build now takes a build PROFILE: asan (default, debug+ASAN -> xce.build,
backward-compatible), debug (non-ASAN -O0 -g3 -fno-inline -fno-omit-frame-pointer
-> build.dbg, for gdb/cores), release (-O3 -> build.release). Each profile uses
its own default build dir so variants don't clobber each other; arg2 may still be
a legacy build_dir. asan runs the unit suites; debug/release are build-only (the
suites are ASAN-oriented and the meson test harness is overlay-fs fragile, bug-924)
and print how to run a suite directly.
Why: bug-1282/bug-1286 are heisenbugs ASAN perturbs away (no SIGSEGV) and that
gdb-on-release turns into a hang/timeout. The non-ASAN -O0 -g3 build reproduces
them AND yields a symbolized backtrace under gdb.
It immediately paid off: the debug profile ROOT-CAUSED bug-1286 (the round-4
segfault) — myreq_handler (cluster.c:251) does `c = ep->owner` then
`ha_dispatch_gate(&c->ha,…)`; on chaotic restart a walCatchupEntry (op 37) vtask
runs with ep->owner==NULL, so &c->ha = 0x3ff0 -> deref crash at ha.c:3194. Fix is
a NULL-guard in myreq_handler (+ owner-lifecycle). 7.2 now has one remaining
blocker (bug-1286, root-caused); bug-1282 deadlock is fixed.
Co-Authored-By: Claude <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019dbTHR9CsEYK9RsqmQpQpC
bug-1282 (ABBA iomgr deadlock) is FIXED and verified (commit 25e741d): zero
__lll_lock_wait threads post-fix, unit suite green. But a plain release soak WITH
the fix still segfaults a restarted standby at round 4 — and deadlocks hang, they
do not SIGSEGV, so that crash is a SEPARATE bug (bug-1283), most likely an
endpoint-lifetime UAF under chaotic restart (bug-1197 family).
bug-1283 is a heisenbug (invisible under ASAN and gdb; no usable core in the
sandbox) and likely UPSTREAM raft-drive-async / core libxce endpoint quiescence.
7.2 stays blocked on it. Consensus correctness (Phase-2/5.2/5.3) is unaffected.
Co-Authored-By: Claude <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019dbTHR9CsEYK9RsqmQpQpC
Root-caused via gdb (all-thread backtrace on two independently-wedged controllers
during chaotic failover): an ABBA lock-order deadlock in the async dispatch path.
io_worker (netio.c): holds ep->lock, blocks on q->lock (vtask_enq -> dataq_enq)
vdispatcher: dataq_deq holds q->lock, blocks on ep->lock via the
vtask_eligible -> endpoint_is_task_blocked match callback
endpoint_is_task_blocked took ep->lock while the caller (dataq_deq) held the dataq
q->lock, inverting io_worker's ep->lock -> q->lock order. Under churn the two
orders collide and the iomgr wedges (no more accepts / vote replies).
Fix (mirrors the dispatch_inflight precedent, bug-1197): promote endpoint_t.busy
from a 1-bit bitfield to _Atomic int and make endpoint_block_tasks /
endpoint_unblock_tasks / endpoint_is_task_blocked lockless (release/acquire). The
ep->lock that wrapped each access protected nothing else (busy only gates the
vtask queue), so removing it is safe; the inherent set-then-check TOCTOU is
unchanged (the lock never prevented it — the decision is advisory).
Verified: post-fix gdb soak shows ZERO threads in __lll_lock_wait and
endpoint_is_task_blocked no longer appears in any lock path (deadlock gone);
controller_unit_test all green (no regression). NOTE: a SEPARATE release-only
restart segfault remains and still blocks the raft-ha-core-safety 7.2 gate
(tracked separately) — deadlocks hang, they do not SIGSEGV, so it is a distinct
memory bug.
Co-Authored-By: Claude <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019dbTHR9CsEYK9RsqmQpQpC
Using gdb (SIGQUIT all-thread backtrace on the wedged nodes), nailed the round-4
blocker to a deterministic ABBA lock-order deadlock in src/lib async dispatch —
identical on both wedged survivors:
- io_worker (netio.c:672): holds ep->lock, blocks on q->lock (vtask_enq->dataq_enq)
- vdispatcher dataq_deq (dataq.c:65): holds q->lock, blocks on ep->lock via the
vtask_eligible (vtaskmgr.c:255) -> endpoint_is_task_blocked (netio.c:1001) cb
io_worker wedges -> node stops accepting/answering PreVotes -> two restarted
survivors can't elect -> round-4 stall. The release SIGSEGV is the same ep/q race
corrupting under churn.
Defect: dataq_deq runs its match callback while holding q->lock, and the callback
takes ep->lock — inverting io_worker's ep->lock->q->lock order. Minimal fix: make
ep->busy atomic + lockless in endpoint_is_task_blocked (or io_worker drops ep->lock
before dataq_enq, or one global lock order).
Ownership: src/lib async-I/O dispatch (vtaskmgr/dataq/netio) = raft-drive-async /
core libxce, a Non-Goal of this spec. Per upstream-ownership protocol, routed to
raft-drive-async (bug-1282); 7.2 stays blocked until the dispatch lock-order is
repaired, then re-run the gate. The gate worked as designed — 7.2 must be GREEN
before raft-drive-async flips dispatch flags ON, and it caught a real dispatch
deadlock. Task 5.4 (root-cause + assign ownership) is DONE. Consensus correctness
(Phase-2/5.2/5.3) is unaffected.
Co-Authored-By: Claude <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019dbTHR9CsEYK9RsqmQpQpC
Used the now-available gdb (wrapped the release controller under gdb -batch via
XCE_HA_BIN) to expose the round-4 blocker's second face. Findings:
- The release-build SIGSEGV and a gdb-run accept-loop HANG are the SAME restart-
time I/O-subsystem corruption: under gdb the memory bug is perturbed away from
a crash and instead wedges the inbound-accept path (ctrl-3 stops accepting at
03:42:03 while still firing its own election timer), so two restarted survivors
can never exchange PreVotes -> permanent round-N stall.
- This is a restart-STABILITY bug, NOT a consensus-logic gap: PreVote/eligibility/
freshness (incl. 5.3) are correct; the restarted node is simply I/O-dead. It
predates this session (baseline 21e26a1 failed round 3-4).
- Heisenbug: invisible under ASAN and under gdb; no usable core (read-only
core_pattern -> non-running systemd-coredump). Needs a non-ASAN DEBUG (-O0 -g)
build under gdb, or a host with a live coredump pipeline, to backtrace.
Updates research.md, refines task 5.4 (the unified restart-stability bug +
the debug-build path), updates bug-1281. Consensus correctness (Phase-2/5.2/5.3)
is unaffected and validated for every round that completes (2-round soak PASSES;
6-round rounds 1-3 elect; zero job loss; committed prefix byte-identical).
Co-Authored-By: Claude <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019dbTHR9CsEYK9RsqmQpQpC
