model-writtenrun 9085 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Jul - Aug 2019 · 5 entries
- Justin Pryzby opened the thread seeking a parallel-worker stress test after database restores failed with unexplained postmaster exits, while Tom Lane recognized similar buildfarm failures and initially suspected OOM kills #1 #2.
- Thomas Munro helped recover the error details from a core dump, revealing an ENOSPC checkpoint failure, but distinguished that failure from the unexplained postmaster deaths; Justin also confirmed that deliberate OOM failures produced kernel logs #9 #10 #12 #15.
- Tom's buildfarm survey narrowed the recurring failures mainly to particular machines, including two sharing a Linux PPC64 host, shifting attention toward a platform-specific problem #16.
- Thomas proposed collecting the postmaster's exit status, and Tom suggested an intermediate process launched by pg_ctl; they concluded that subreaper machinery was unnecessary #17 #18 #19.
- Heikki Linnakangas found neither OOM evidence nor a core file on chipmunk, leaving the cause unresolved and strengthening the need for better crash diagnostics #22 #24 #26.
Oct 2019 · 9 entries
- After a two-month gap, Tom submitted temporary pg_ctl instrumentation that retained a shell to print the postmaster's exit status, explicitly accepting degraded PID-dependent behavior; #27 landed unchanged as 6a5084eed4, with removal intended after diagnosis.
- The instrumentation captured exit status 139 on wobbegong, proving SIGSEGV rather than an unexplained normal exit; Mark Wong supplied access and Andrew Dunstan adjusted core collection #29 #30 #32.
- Tom obtained a backtrace showing hundreds of nested reaper and SIGUSR1 handler frames, explaining how rapid parallel-worker creation and termination could exhaust the postmaster's stack #33.
- Tom then observed crashes near 1 MB despite an 8 MB stack limit and submitted a separate fix using sigaction to block signals throughout handler execution; his tests stopped crashing and used much less stack, while Windows support and helper cleanup remained unfinished #35.
- Andres Freund questioned memory pressure and interruptibility inside handlers, and Tom supplied memory and signal diagnostics while explaining that the patch changed only masking during handler entry and exit #40 #41.
- Thomas pointed to PowerPC stack-expansion checks, and Tom connected signal delivery near the stack boundary to infinite_recurse failures, reproduced the kernel problem with a standalone program, and filed a kernel bug #37 #44 #45 #51.
- The signal fix from #35 landed as 9abb2bfc04 after Tom introduced pqsignal_pm, moved its implementation into backend code, and preserved Windows' manual masking; the initial landing targeted HEAD while backpatching was deferred for aging.
- The separate infinite_recurse failures remained, and Tom proposed isolating that test from concurrent tests to avoid signal-triggered kernel crashes, without submitting that workaround here #52.
- Justin closed the original customer incident by identifying an Ansible timeout that sent SIGKILL to the postmaster, establishing that his reported failure had a different cause from the PPC64 buildfarm crashes #53.
Dec 2019 · 1 entry
- Thomas reported a proposed kernel fix in December and further corrections the following July: kernel-generated signal frames needed stack expansion, and the assumed signal-frame size was also too small #54 #55.
Jul - Aug 2020 · 2 entries
- By August, the kernel fix had reached upstream and then the relevant stable branches, leaving distribution uptake as the remaining deployment concern #57 #58 #59.
- In October, the separate thread thread #43077 supplied #460080 for ae0f7b11f1, which split out infinite_recurse and skipped its execution on Linux PPC64 while older affected kernels remained in use.
Oct 2020 · 1 entry
- The thread ended with Thomas confirming that a Debian buster kernel update included the stack-expansion fix; both attached PostgreSQL changes had landed, and the remaining discussion concerned availability of the kernel repair #60.