model-writtenrun 5941 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Andres Freund opened the thread by proposing per-process personality(ADDR_NO_RANDOMIZE) for Linux EXEC_BACKEND builds, avoiding both intermittent test failures and system-wide ASLR changes #1; Andrew Dunstan and Tom Lane supported an unobtrusive solution #2 #6.
- Robert Haas reported especially frequent failures on macOS #4, while Andres identified that disabling ASLR only inside the postmaster could leave its existing layout incompatible with its children and suggested doing it in pg_ctl instead #7.
- Thomas Munro posted an experimental macOS patch using posix_spawn() with an undocumented ASLR-disable flag in both pg_ctl and the postmaster #8; he then identified pg_regress's separate temporary-cluster launch path as another place needing the change #11.
- Andrew Dunstan asked whether initdb also needed adjustment #12, and Tom Lane explained that its standalone backend does not have the cluster's shared-memory attachment problem #14.
- Thomas abandoned the macOS spawning approach after finding that dyld remained randomized, and reported consistent Intel and ARM test success using PG_SHMEM_ADDR=0x80000000000 instead #15.
- After Robert requested an automatic workaround and Tom suggested incorporating it into startup #16 #17, Thomas posted a default-address implementation in shared-memory creation, avoiding duplicated pg_ctl and pg_regress changes #18.
- Tom requested a more extensible placement of the platform conditional #19, and Thomas incorporated that and wording changes into 88cbbbfa3e, announcing its backpatch to all live branches while noting that future macOS layouts might require another address #20.
- FreeBSD enabling ASLR by default prompted Thomas to return with a separate Linux/FreeBSD patch that disables ASLR through pg_ctl and pg_regress; direct server launches remain outside its coverage #22.
- Nathan Bossart reported successful Linux testing #23, after which Thomas committed the patch with comment, header, and MSVC configuration cleanup as f3e78069db, initially on master only #24.
- A year later, Michael Paquier reported shared-memory attachment failures on buildfarm member gokiburi's older branches and proposed backpatching the Linux/FreeBSD change #25; Tom and Thomas agreed, and Michael reported successful testing before delaying the change until the upcoming minor releases were tagged #26 #28 #29 #30.
- Michael applied the existing implementation to stable branches through PostgreSQL 12, represented by 0801345758; he left PostgreSQL 11 out because its missing procctl detection required additional backport work and it was nearing end of support, choosing to reduce gokiburi's coverage instead #31.
- The thread ended with both platform solutions landed and the stable-branch scope settled; Thomas flagged that backpatching did not require resetting authorship, and Michael acknowledged that he had accidentally done so for Thomas's patch #33 #34.