60 lines
6.7 KiB
Markdown
60 lines
6.7 KiB
Markdown
---
|
|
id: "bug-1913"
|
|
status: "fixed"
|
|
status_updated: "2026-10-03"
|
|
status_note: ""
|
|
component: "controller"
|
|
severity: "wrong-result"
|
|
timestamp: "2026-10-03T14:18:58Z"
|
|
last_seen: "2026-10-03T14:18:58Z"
|
|
occurrences: 1
|
|
files: ["src/controller/ha.c", "src/controller/repl.c", "src/controller/config_repl.c"]
|
|
tags: ["ha", "raft", "replication", "latency"]
|
|
related: ["bug-1485"]
|
|
fix_commits: ["2aefa939"]
|
|
---
|
|
|
|
# bug-1913: after a late-joiner election the primary cannot commit client writes for ~2 heartbeat cycles
|
|
|
|
## Description
|
|
**Observed:**
|
|
```
|
|
repl_broadcast: no standbys registered ... not replicated
|
|
...
|
|
WAL NACK from standby lsn=2 ... marking catch-up
|
|
repl_wait_majority_ack: timeout ... OP_upgradeConfig rejected (status=-110)
|
|
```
|
|
In some elections the vote connections are not converted into standby channels (`raft_cb_become_leader`), so both standbys register later through `on_standby_connected` (ha.c ~9643-9660). `ha_arm_own_term_gate` (ha.c ~5178) has already appended the own-term WAL_NOOP (lsn 1) while no standby was registered (repl.c ~306-311), and `repl_add_standby` does not resend it. Catch-up then runs from the 3 s heartbeat, one action per standby per cycle (config push first, WAL stream on the next cycle), so `committed_lsn` stays 0 for ~6 s. Any client write in that window (e.g. `OP_upgradeConfig`) is NACKed by both standbys and fails with -ETIMEDOUT after the 500 ms majority wait.
|
|
|
|
**Expected:** a newly elected primary commits its own-term no-op, and so accepts client writes, promptly after its standbys register, not two heartbeat cycles later.
|
|
**Impact:** for ~6 s after such an election every client write is rejected with -110 (fail-loud, no data loss, no false success). Clients retry, but tests and operators see spurious failures right after failover (the root cause of bug-1485's versionTag=0 flake). A fix must respect the async-I/O rules (no blocking on the delaymgr/heartbeat path; streaming offloaded) in `.kiro/steering/ha-async-io.md`.
|
|
**Found by:** test (root cause of bug-1485)
|
|
|
|
## Environment
|
|
- OS: no /etc/os-release; Docker container; NixOS kernel; kernel Linux 6.18.33 x86_64
|
|
- libc: glibc 2.42-47 (interpreter /nix/store/wb6rhpznjfczwlwx23zmdrrw74bayxw4-glibc-2.42-47/lib/ld-linux-x86-64.so.2)
|
|
- Toolchain: gcc (GCC) 15.2.0, meson 1.9.1
|
|
- Build: asan (xce.build)
|
|
- Code: scheduler-framework @ 91782e40-dirty
|
|
- Deployment: local 3-node HA (xceadmin live e2e fixture), CPU-loaded sandbox (40 busy loops on 32 cores) makes late-joiner elections frequent
|
|
|
|
## Reproduction
|
|
1. Start a 3-node HA cluster (e.g. the `live_ha_cluster` fixture in `src/xceadmin/tests/test_e2e_live_ha_push.py`) under CPU load.
|
|
2. In the primary log, find an election with no "promoted" standby channels and a later "repl_add_standby"/late registration; push a config (`xceadmin config push ... --controller <primary>`) within ~6 s.
|
|
3. The push fails with `controller returned status -110`; the log shows the NACK and majority-ack timeout above.
|
|
- Frequency: 7/28 runs of the HA push test under 40-spinner CPU load, 0/6 idle (2026-10-03); all 7 failures were late-joiner elections.
|
|
- Evidence: not kept (scratchpad loop logs from the bug-1485 investigation).
|
|
|
|
## Root cause
|
|
Nothing asked a late-registered standby for its state, so its catch-up waited for the heartbeat clock. `ha_arm_own_term_gate` (ha.c ~5180) appends the own-term no-op while no standby is registered; `repl_broadcast` drops it (repl.c:306-311) and `repl_add_standby` never resends it. `on_standby_connected` (ha.c ~9511-9690) sends only a ctrlhello (no reply) and sets `last_push_time = now`. Catch-up is driven solely by a heartbeat reply's state report → `repl_ack_decode_and_dispatch` (peer-control vtask) → `ha_dispatch_repl_action_for_standby` (ha.c:4455), one action per report (config push if the tag is behind, else WAL stream). A standby rejoining with tag 0 therefore got the config push on the next 3 s tick and the WAL stream carrying the no-op on the tick after: ~6 s with committed_lsn below the no-op and every client write NACKed. The config-before-WAL order is a pacing rule of the ha-repl-hardening design ("no further decisions until the next state report confirms"), not a correctness requirement. Why the vote connections were not promoted in the first place is tracked separately (bug-1915).
|
|
|
|
## Fix
|
|
Make the state report arrive sooner; the decision loop and its one-action-per-report pacing are unchanged. New `ha_send_catchup_probe()` (ha.c) sends the ordinary `xce_ha_heartbeat_req` (same fields, read the same way as `primary_heartbeat_fn`), only while `CTRL_PRIMARY`, encode + enqueue only (same non-blocking path as the ctrlhello and the delaymgr tick); a send failure is a WARN and the periodic tick stays the fallback. Triggers: (1) once at the end of `on_standby_connected` (iomgr connect callback, after `msg_handler` is set), which also arms `ha->catchup_kick_armed[peer]` (ha.h); (2) the accepted config-push-ACK branch of `repl_ack_decode_and_dispatch` (peer-control vtask) after a 1→0 CAS on the arm, so the WAL stream follows the config push at ~RTT while keeping config-before-WAL. The arm is cleared on the standby's disconnect (`standby_ep_err_handler_primary`) and on every `raft_cb_become_leader` / `raft_cb_become_follower`, so it never outlives its registration or leadership; at most two probes per connect. No new executor, opcode or protocol change; heartbeat interval unchanged; WAL streaming still runs only on the vtask.
|
|
|
|
## Verification
|
|
- Test: `src/controller/testsuite/test_late_joiner_catchup_kick.c` — `test_late_joiner_registration_probes_standby_state`, `test_config_ack_after_registration_reprobes_once`, `test_rejected_config_ack_does_not_probe` RED before the fix (Expected 1 Was 0), GREEN after; `test_step_down_clears_catchup_arm` RED with the become_follower clear removed (Expected 0 Was 1); guards for no-registration and non-primary.
|
|
- Suites: controller 5/5 meson suites, main 1285/0 (asan); `xce-build.sh all --incremental` green (C tests 127/0, scheduler 16/0, xceadmin 373), 2026-10-03.
|
|
- Live: `/xce-test-ha` sequential 20 rounds PASS (63 jobs, zero loss, all 3 led, SIMVERIFY 21/0); chaotic 6 rounds PASS (zero committed divergence/loss); sequential 8 rounds under 40 busy loops PASS (with the bug-1916 fix; earlier loaded runs aborted on bug-1916). In those soaks catch-up started 0-6 ms after each late standby registration and a restarted standby was fully caught up 46 ms after registering (was up to ~6 s). Proxy (wiped standby) before/after: WAL stream at +5.8 s → +18-60 ms. Live HA push tests 6/6 under load; promoted-vote elections commit the no-op in median 6 ms (max 16 ms). Observation, not a regression: after catch-up the decision loop re-streams already-acked ranges on later stale state reports (harmless; standby discards duplicates).
|
|
|
|
## References
|
|
- Related: none
|