model-writtenrun 7219 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Thomas Munro opened the thread with a proposal to keep Windows headers and WIN32 definitions out of Cygwin translation units, accompanied by experimental CI support intended eventually to run on demand #1.
- Tom Lane and Robert Haas questioned continued Cygwin support because unreliable signal masking caused recurring crashes, while Thomas preferred an upstream fix and later floated—but did not implement—signal-handler guards; no decision to remove support followed #2 #4 #5 #10 #15 #17.
- Justin Pryzby took over development of the CI patch, addressing slow configuration, getopt failures and TAP portability, then removed unnecessary workarounds and added stackdump decoding; testing exposed intermittent fsync, shared-memory and semaphore failures despite occasional successful runs #6 #11.
- After the summer investigation, Justin converted the CI work to Meson and Ninja and separated out optional-PROVE and Cygwin build prerequisites, initially disabling precompiled headers because of GCC crashes #18.
- The optional-PROVE change first attached here in #18 landed unchanged as ab72a31f6c, but its source attribution belongs to the later matching submission #519359 in thread #46498.
- Justin incorporated Andres's feedback into Cygwin regression host identification and revised the CI configuration in #19; Andres then committed its unchanged Meson-support component as 79f7c482f6, leaving CI and test-infrastructure work outstanding #20.
- Andres recommended sharing Windows VM images, shortening repeated shell commands through an environment variable, reconsidering stackdump handling and separating non-CI fixes, while Thomas connected the signal failures to separate postmaster WaitEventSet work #20 #21.
- Testing the CI patch with the postmaster changes nearly passed the full TAP suite, but semctl() still failed with EAGAIN; Thomas proposed bypassing cygserver's SysV semaphores and committed unnamed POSIX semaphore selection directly from that investigation as f2857af485, then reported two successful runs #24 #25.
- Thomas subsequently withdrew the CI entry because successful testing depended on data_sync_retry masking snapshot-file failures attributed to non-atomic rename behavior, and advised Andrew Dunstan to wait for the postmaster changes before re-enabling lorikeet's TAP tests #26 #27 #28.
- Justin reopened the work with the remaining CI patch, using a Windows VM and replacing the broad workaround with ERROR-level handling in SnapBuildSerialize(), then enabled automatic CI execution #29 #30.
- Andres challenged the fsync change's correctness and rejected weakening synchronization merely to pass Cygwin tests; Justin explored alternative synchronization sequences and reported further snapshot failures, leaving the proposed workaround unresolved #31 #32 #33 #34 #35 #36.
- Thomas later found the same snapshot panic on native Windows, and by the final message suggested investigating checkpoint-driven snapshot synchronization, reconsidering full TAP coverage on lorikeet and separating wal_sync_method concerns; the two Meson changes and semaphore fix had landed, with optional-PROVE attributed to thread #46498, while the CI patch and original header cleanup had no demonstrated landing in this thread #37 #38.