model-writtenrun 5808 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Noah Misch opened the thread with a patch exposing the underlying XLogReadRecord() error when reading two-phase state, and Michael Paquier approved it while moving broader diagnostic improvements into another thread #1 #2 #4.
- The diagnostic patch landed unchanged as 3354746910, and Noah reported that the observed failures were concentrated on SPARC64 Debian machines #5.
- The enhanced diagnostics exposed zero page addresses, zero record lengths, and checksum failures; Tom Lane questioned synchronization, while Noah traced the locking and WAL-availability checks without finding a gap #6 #7.
- Andrey Borodin suggested more crash-recovery testing, but Tom Lane considered that premature until the existing failure was understood and subsequently collected the related buildfarm failures #8 #9 #12.
- After obtaining access to an affected machine, Noah isolated a kernel I/O problem in which concurrent reads and writes could return zeros, and ruled out the earlier build-option theory #13.
- Prompted by related discussion from Thomas Munro, Noah narrowed the zero-reading bug to SPARC64 with ext4, reproduced it on newer kernels, and proposed avoiding repeated writes to existing WAL bytes #14 #15.
- Thomas identified the streaming regression test as another likely victim, and Noah raised selective test skipping as a possible stopgap #16 #17.
- Noah posted the write-once WAL workaround after an affected machine survived 8.5 hours of testing, intending to restrict it to SPARC Linux before committing #18.
- Thomas Munro identified O_DIRECT alignment failures, and Andres Freund raised read-modify-write costs and future direct-I/O constraints, leading Noah to abandon the runtime workaround in favor of restoring useful buildfarm coverage #19 #20 #21.
- The discussion considered buildfarm-side expected-failure handling and an environment variable, but Andres advocated detecting affected configurations inside the tests and marking known failures as TODO #22 #24 #25 #27.
- Noah posted automatic SPARC Linux/ext4 detection with test skips; Michael Paquier favored continuing to run tests under TODO, Tom Lane required platform-specific conditions, and Andres supplied the conditional Perl mechanism #28 #29 #30 #32.
- Noah settled on TODO for the amcheck test and skip_all for the streaming test because the latter could die outright, and ce6d79368e landed that lightly revised form of #28 #33.
- A subsequent snapper failure revealed that older Test::More versions could miss the localized TODO variable; Noah proposed calling todo_start directly, and that inline fix landed as adbd00f7a5 #34.
- At the thread's last message, Noah had reported the kernel bug to Debian and upstream, and another tester had reproduced it on Linux 5.17.0-rc5; the PostgreSQL diagnostic and test mitigations had landed, but the underlying kernel problem remained unresolved #35.
- After the visible discussion ended, ad76c9708b extended the same mitigation to contrib/bloom/t/001_wal.pl following snapper and kittiwake failures, using skip_all because the test could die.
- A further kittiwake failure led to a9f8ca6005, which likewise skipped src/test/recovery/t/032_relfilenode_reuse.pl on configurations detected by has_wal_read_bug.