model-writtenrun 4957 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Thomas Munro reported that mamba's replication test failed to start a WAL receiver in #1; Tom Lane identified children surviving SIGQUIT and proposed sending SIGABRT instead of SIGKILL to obtain diagnostic core dumps in #2.
- Thomas Munro considered and then withdrew a runtime symbol-resolution hypothesis in #3 and #4, while Tom Lane proposed configurable SIGABRT handling and replacing the existing -T option's SIGSTOP behavior in #5.
- Andres Freund recommended a GUC after encountering excessive core dumps with a compile-time approach in #6; Tom Lane supplied #7, adding send_abort_for_crash and send_abort_for_kill, changing -T, removing the defunct -n option, and asking whether grandchildren should also receive SIGABRT.
- Tom Lane committed #7 unchanged as 51b5834cd5 and enabled send_abort_for_kill on mamba; in #8, he concluded that signaling grandchildren was necessary for prompt shutdown.
- The new core dumps revealed WAL receiver and autovacuum processes stuck in recursive runtime-linker symbol resolution in #9; Thomas Munro suggested checking signal masks and breaking on the resolver in #10, correcting the proposed inspection function to sigprocmask in #11.
- Andres Freund proposed that eager binding had missed shared-library dependencies and suggested LD_BIND_NOW in #12; subsequent discussion examined incomplete stack frames, unchecked signal-handler installation failures, and OpenSSL signal-mask handling without establishing those as the cause #13 #14 #15 #16 #17.
- Tom Lane traced the deadlock to a lazily resolved call from libpthread's select implementation into libc during the postmaster's first wait, confirming Andres Freund's dependency hypothesis; he enabled LD_BIND_NOW=1 on mamba and verified early resolution, while noting that confidence in the workaround would require weeks of observation #18.
- The diagnostic patch was landed, but the underlying platform issue remained open: Andres Freund discussed static linking, explicitly resolving library symbols, and retaining eager-binding flags in #19, while Tom Lane favored trying a dummy select call before unblocking signals in the thread's final message #20.