3.9 KiB
| id | status | status_updated | status_note | component | severity | timestamp | last_seen | occurrences | files | tags | related | fix_commits | legacy | gitea_issue | gitea_synced_status | gitea_synced_hash | ||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| bug-1450 | fixed | 2026-09-27 | 7aac018c: demotion-convergence offload ON by default; inline fallback when slow lane not started |
controller | unknown | 2026-07-18 | 2026-09-19T09:00:00Z | 1 |
|
|
|
|
true | 25 | fixed | 69e69f68fbf7 |
bug-1450: xce-audit-invariants controller pass (dim: async I/O thread-role) found ha_demotion_convergence_fn…
Description
xce-audit-invariants controller pass (dim: async I/O thread-role) found ha_demotion_convergence_fn scheduled directly as a delaymgr task body with no vtask/slow-lane hop: it truncates the WAL (persist_truncate_wal_above, file I/O under p->lock) then calls ha_self_restart (tlog_exit synchronous log flush + execv) -- all inline on the delaymgr thread, the SAME single thread that drives heartbeats and election timeouts cluster-wide for the whole cluster.
Environment
Not recorded (migrated from buglog.json).
Reproduction
Not recorded (migrated from buglog.json).
Root cause
ha_demotion_convergence_fn (ha_convergence.c) ran its full body -- truncate + self-restart -- synchronously on whatever thread invoked it (the delaymgr worker in production). A prior fix attempt (2026-07-10, see bug-1429) tried moving this to a VTASK_LIVE_BLOCK vtask via c->ha.pool and deadlocked controller_unit_test, because c->ha.pool is never initialized/started by this file's lightweight unit-test fixtures (dc_setup), so a vtask enqueued VTASK_LIVE_BLOCK blocked forever waiting for a worker that never existed; that attempt was reverted.
Fix
Split ha_demotion_convergence_fn into a dispatcher (fast admin work: single-flight CAS on new ha_state_t field demotion_convergence_inflight, then slowlane_submit) and a new demotion_lane_fn slow-lane worker body running the actual truncate+self-restart (extracted into ha_demotion_convergence_do_run) -- the SAME established slow-lane pattern already used for snapshot send / config-push / peer connect in ha.c (ha_snapshot_send_dispatch, ha_config_push_offloadable, ha_connect_to_peer), not the vtask pool that deadlocked before. New feature flag g_drive_async_demotion_offload (ha_set_demotion_offload_for_test / ha_demotion_offload_enabled, ha.h/ha_convergence.c) gates the offload. IMPORTANT: defaulting this flag to ON (1) was tried first and reproduced a NEW real hang -- raft_cb_become_follower (which schedules this task) is exercised by a very broad set of pre-existing controller_unit_test fixtures across many files that never call slowlane_init, so the offloaded work was silently queued on an uninitialized c->slowlane (queued, never run -- no live workers) and downstream code waiting on the now-never-happening completion hung. Reverted the default to OFF (0), matching the ORIGINAL rollout discipline of the three reference offloads (all shipped OFF-by-default, flipped ON only in a later dedicated compatibility-audit pass -- raft-drive-async task 3.4). The mechanism is correct and covered by dedicated tests (test_demotion_offload_task_returns_promptly / test_demotion_inline_task_blocks_RED in test_demotion_convergence.c) that explicitly flip the flag ON against a fixture with a real slowlane_init'd lane. Flipping the cluster-wide default to ON is a follow-up compatibility audit across the full test suite, explicitly out of scope for this fix. Verified: standalone controller_unit_test binary run directly (bypassing meson) reports 976 Tests 0 Failures 0 Ignored, no hang/timeout.
Verification
Not recorded (migrated from buglog.json).
References
- Related: bug-1429, bug-1451
- Gitea: #25