Rongsong Shen
d26ff8870d
The buglog was one 27K-line JSON file whose fields covered only symptom,
root cause and fix, as free text: 2% of entries named the code version a
bug was found on, 12% the build profile, 36% a way to reproduce, and 58% of
the file was OpenWolf hook auto-detections ("Missing guard clause",
"Function not marked async" on .md files).
- .claude/skills/buglog/bug-record.md: the record standard. Eight
sections (Summary as the title, Description, Environment, Reproduction;
Root cause, Fix, Verification, References once fixed), each with the
content it must contain -- e.g. Environment needs the OS, the libc the
failing binary loads (from its ELF interpreter: glibc 2.42 in the nix
shell vs 2.40 on the host), toolchain, build profile and commit.
- .wolf/bugs/<id>.md: 500 records migrated from buglog.json; every field
verified to round-trip. They carry legacy: true and an explicit "not
recorded" where the old log had no such field; component derived from
file paths, severity from keywords (else unknown), fix_commits from
hashes in status_note/fix that git resolves (91 of 463 fixed).
- .wolf/buglog-archive.json: the 710 hook auto-detections, verbatim.
- .wolf/buglog.json is now an index generated from the records, keeping
the fields OpenWolf's hooks and `openwolf bug search` read.
- scripts/buglog.py rewritten: new (template + Environment collected from
the machine and --binary), check, status (refuses a status the record
doesn't support; fixed needs Root cause, Fix, Verification and
--commit hashes git resolves), note, recur, set, show, list, triage,
env, index. 29 unittest cases in scripts/tests/test_buglog.py.
- scripts/git-hooks/pre-commit runs `check --staged`: incomplete staged
records, a legacy record hand-changed to fixed without the fix-time
sections, a new record claiming legacy: true, or a stale index refuse
the commit.
- scripts/gitea-sync.py reads the records: issue body = the record's
sections plus linked fix commits (origin remote, GITEA_COMMIT_REPO to
override), and a body hash so an issue's text is refreshed when the
record changes.
- bug-1866 rewritten to the full standard as the worked example.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
38 lines
5.5 KiB
Markdown
38 lines
5.5 KiB
Markdown
---
|
|
id: "bug-1737"
|
|
status: "fixed"
|
|
status_updated: "2026-09-24"
|
|
status_note: ""
|
|
component: "lib"
|
|
severity: "unknown"
|
|
timestamp: "2026-08-30T22:05:00Z"
|
|
last_seen: "2026-09-19T09:00:00Z"
|
|
occurrences: 1
|
|
files: ["src/controller/sched.c (sched_job_exit_fun", "renamed sched_job_finish_fun)", "src/scheduler/src/handler.c (sched_handle_notify's NotifyData_data_schedExitJob case)", "src/schema/controller.capnp (schedExitJob field)", "src/include/typemap.h (xce_notify_data_t's sched_exit_job union member)", "src/lib/helper.c (xce_notify_data_ptr)"]
|
|
tags: ["scheduler", "controller", "wal", "notify", "capnp", "job-state", "exit-code", "done-vs-exit", "confirmed-live"]
|
|
related: ["bug-1731", "bug-1734", "bug-1738"]
|
|
legacy: true
|
|
---
|
|
|
|
# bug-1737: User-reported (live 3-node HA cluster + worker, 50-element array job `sleep 100 && hostname`…
|
|
|
|
## Description
|
|
User-reported (live 3-node HA cluster + worker, 50-element array job `sleep 100 && hostname`, xce_scheduler -i -s <replay> vs controller joblist): job 0-8-1788143749-0[38] shows EXIT in the scheduler's replay (`info jobs`) but DONE in the controller's joblist -- confirmed a REAL state divergence (JOB_STAT_EXIT=0x04 and JOB_STAT_DONE=0x08 are genuinely different bits, src/include/xce/types.h), not cosmetic.
|
|
|
|
## Environment
|
|
_Not recorded (migrated from buglog.json)._
|
|
|
|
## Reproduction
|
|
_Not recorded (migrated from buglog.json)._
|
|
|
|
## Root cause
|
|
CONFIRMED via direct code reading (both static and, indirectly, live -- the fixed behavior was verified against a real captured replay; the pre-fix defect itself was not separately re-reproduced live in this session since fixing it was done in the same pass). HOOK_DETAIL_JOB_EXIT fires for EVERY job finish, both success and failure -- job.c's record_job_finish_fact()/apply_job_finish_transition() decide DONE (exit_code==0) vs EXIT (exit_code!=0) BEFORE firing this exact hook detail either way (mirrors persist.c's own WAL_JOB_FINISH -> HOOK_DETAIL_JOB_EXIT mapping, `case HOOK_DETAIL_JOB_EXIT: return WAL_JOB_FINISH;`). sched.c's sched_job_exit_fun() (the HOOK_DETAIL_JOB_EXIT subscriber that notifies the scheduler) receives xce_jobfinish_t* info (which HAS an exit_code field) but never read it -- it unconditionally sent a bare jobid via NotifyData_data_schedExitJob (controller.capnp's schedExitJob was a plain UInt64, no exit_code at all on the wire). scheduler/src/handler.c's sched_handle_notify() then unconditionally mapped ANY schedExitJob to JOB_STAT_EXIT, with no way to ever land on JOB_STAT_DONE via this live path. Under STABLE connectivity this was normally self-correcting: the SAME WAL apply also fires HOOK_DETAIL_JOB_UPDATE moments later (persist.c:4854-4855, immediately after the HOOK_DETAIL_JOB_EXIT fire, same lock scope), which carries the job's REAL, already-final job->job.state via schedUpdateJob -> XCE_CORE_EVENT_JOB_UPDATE's `job->state = updated->state;` (core_events.c), overwriting the wrongly-forced EXIT with the correct DONE within the same notify batch. But this is TWO SEPARATE notify() calls, not atomic -- if the scheduler's connection blipped between them (the same reconnect-flakiness bug-1731/bug-1734 already documented live under this exact 50-element-array load), the correcting schedUpdateJob could be dropped (c->sched.ep NULL at fire time) while the earlier, wrong schedExitJob(EXIT) had already landed -- leaving the scheduler permanently stuck at EXIT for a job the controller correctly recorded as DONE. This exactly explains why only ONE array element (out of 50) showed the divergence: only that job's particular pair of notifies happened to straddle a reconnect gap.
|
|
|
|
## Fix
|
|
FIXED (2026-08-30). Reused the EXISTING JobFinishInfo wire shape (jobid + code) apiserver_job_exit_fun() (apiserver.c) already sends for this identical hook, rather than inventing a new capnp struct: controller.capnp's schedExitJob field changed from `UInt64` to `JobFinishInfo`; typemap.h's xce_notify_data_t.sched_exit_job union member changed from `uint64_t` to `xce_jobfinish_t*`; sched.c's sched_job_exit_fun() (renamed sched_job_finish_fun() for accuracy -- it covers both DONE and EXIT outcomes, not just EXIT) now sends `.sched_exit_job = info` (the whole struct, mirroring apiserver_job_exit_fun's `.finishinfo = info` pattern) instead of `info->jobid`; scheduler/src/handler.c's sched_handle_notify() now reads finish->exit_code and mirrors job.c's own decision exactly: exit_code==0 -> JOB_STAT_DONE, else -> JOB_STAT_EXIT, instead of unconditionally assuming EXIT. helper.c's xce_notify_data_ptr() updated to match the new pointer type (dropped a now-redundant void* cast). TDD: added test_handle_notify_job_exit_maps_exit_code_to_done_or_exit (test_handlers.c) -- exit_code==0 -> DONE, exit_code!=0 -> EXIT, NULL payload -> -EINVAL (no crash). Full lib+controller+scheduler+worker build and test suite (both a dedicated full build via xce-build.sh and an incremental rebuild+retest after further edits): 61/61 Ok for controller+scheduler (66/67 overall including lib, the 1 failure a pre-existing unrelated ASAN-leak fixture bug already logged separately, not caused by this diff), xceadmin 288/288. Live-verified: deployed a fresh 3-node HA cluster + worker with the fixed binaries, submitted a 51-job workload (50-element array + 1 standalone job, all exiting 0), let everything run to completion -- the scheduler's replay showed DONE for every single job, matching the controller's joblist exactly, with zero EXIT/DONE divergence.
|
|
|
|
## Verification
|
|
_Not recorded (migrated from buglog.json)._
|
|
|
|
## References
|
|
- Related: bug-1731, bug-1734, bug-1738
|