62 lines
4.9 KiB
Markdown
62 lines
4.9 KiB
Markdown
---
|
|
id: "bug-1914"
|
|
status: "fixed"
|
|
status_updated: "2026-10-04"
|
|
status_note: ""
|
|
component: "controller"
|
|
severity: "wrong-result"
|
|
timestamp: "2026-10-03T15:07:26Z"
|
|
last_seen: "2026-10-03T15:07:26Z"
|
|
occurrences: 1
|
|
files: ["src/controller/config_repl.c", "src/controller/config_repl.h"]
|
|
tags: ["ha", "config-push", "timeout", "load"]
|
|
related: ["bug-1485", "bug-1913"]
|
|
fix_commits: ["06e99b90"]
|
|
gitea_issue: 341
|
|
gitea_synced_status: "fixed"
|
|
gitea_synced_hash: "71c6272fe0e6"
|
|
---
|
|
|
|
# bug-1914: config push fails -110 under heavy CPU load: commit lands just after the 500 ms barrier
|
|
|
|
## Description
|
|
**Observed:**
|
|
```
|
|
Error: controller returned status -110
|
|
```
|
|
`xceadmin config push` to a healthy, normally elected HA primary (standbys promoted from the vote connections, own-term no-op already committed) was rejected with -ETIMEDOUT: the config marker's majority commit landed ~75 ms after the `repl_wait_majority_ack(..., CONF_PUSH_TIMEOUT_MS)` barrier (`config_repl.c` ~1658; `CONF_PUSH_TIMEOUT_MS` = 500, `config_repl.h:33`) expired. Seen only with the CPU heavily oversubscribed (40 busy loops on 32 cores).
|
|
|
|
**Expected:** a push to a healthy primary with a reachable majority succeeds under load, or the barrier is sized/structured so normal scheduling delay cannot turn a soon-to-commit marker into a rejection. (A commit that lands after the client was told -110 also leaves the outcome ambiguous to the client: rejected per the reply, yet committed later.)
|
|
**Impact:** spurious push failures under heavy load (fail-loud, no data loss). If the marker then commits after the -110 reply, the client believes the push failed while the cluster activated it -- a retry is idempotent per tag, but the reply is misleading. Check before changing: whether the late commit is actually applied after the -110 reply, and the async-I/O rules for the commit-wait barrier (`COMMIT_WAIT_RESERVED`, `.kiro/steering/ha-async-io.md`).
|
|
**Found by:** test (bug-1485 investigation, 2026-10-03)
|
|
|
|
## 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 @ 4cd5fb7d-dirty
|
|
- Deployment: local 3-node HA (xceadmin `live_ha_cluster` fixture), CPU oversubscribed with 40 busy loops on 32 cores
|
|
|
|
## Reproduction
|
|
1. Start 40 CPU busy loops on a 32-core host.
|
|
2. Loop `src/xceadmin/tests/test_e2e_live_ha_push.py` (with the bug-1485 readiness wait in place).
|
|
3. Rarely, the push fails with `controller returned status -110` on a primary whose own-term no-op had already committed; the primary log shows the marker commit ~75 ms after the 500 ms barrier expired.
|
|
- Frequency: 1/70 HA push loops under heavy load, 0 idle (2026-10-03)
|
|
- Evidence: not kept (scratchpad loop logs from the bug-1485 investigation)
|
|
|
|
## Root cause
|
|
`config_commit_change` appends the `WAL_CONFIG_MARKER` first and then waits for the out-of-band byte-delivery quorum (`repl_wait_majority_ack(..., CONF_PUSH_TIMEOUT_MS)`, configAck branch). `CONF_PUSH_TIMEOUT_MS` was 500 ms, which is below ordinary scheduling delay on an oversubscribed host: with 40 busy loops on 32 cores a standby's ACK arrived ~575 ms in, so the push was rejected with -ETIMEDOUT although the majority was healthy. The marker, already in the log, still committed and activated, so the client was told the push failed for a change the cluster applied (the reply text even said "upgrade not committed"). The configAck wait was also a raw condition wait on a vtask outside the commit-wait admission reservation that ha-async-io.md requires for bounded blocking there.
|
|
|
|
## Fix
|
|
- `CONF_PUSH_TIMEOUT_MS` 500 -> 2000 ms (src/controller/config_repl.h).
|
|
- The configAck wait in `repl_wait_majority_ack` now holds a commit-wait admission slot for its duration (new `commit_blocking_admit`/`commit_blocking_release` in commit.c, the same budget `commit_submit` WAIT uses); a full budget rejects fast with -XCE_ERR_NOT_READY, so the longer bound can never take workers reserved for liveness traffic.
|
|
- The timeout reply no longer claims "upgrade not committed": it says the change was not confirmed on a majority, may still activate if it was already ordered in the log, and should be re-pushed to confirm.
|
|
|
|
## Verification
|
|
- Test: `test_configack_late_ack_under_load_meets_barrier` (src/controller/testsuite/test_commit_barrier_rollout.c): a config ACK arriving 700 ms in must meet the production bound, and the wait must hold one admission slot (released afterwards). RED before the fix (-110 at 500 ms), GREEN after.
|
|
- Suites: `xce-build.sh all` asan, 127/127 meson suites OK, Go tests ok (2026-10-04).
|
|
- Live: `test_live_config_push_against_primary_reaches_success_path` looped 80 times under 40 busy loops on 32 cores, 2026-10-04: 80/80 passed, no -110 (before the fix: 1/70 failed with -110 under the same load).
|
|
|
|
## References
|
|
- Related: none
|