model-writtenrun 1333 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Dimitrios Apostolou reported that PostgreSQL's use of posix_fallocate disabled Btrfs compression during database restores, verified that compiling without it restored compression, and requested a runtime workaround, preferably per tablespace #1 #3.
- Thomas Munro proposed a configurable allocation threshold and a tablespace override, with the buffer manager consulting catalogs before taking the extension lock and passing the result into smgr #4.
- Tomas Vondra challenged the threshold's counterintuitive zero value, Tom Lane rejected the tablespace implementation because of layering risks, and Jakub Wartak preferred a simpler switch while identifying similar motivation from XFS allocation failures #6 #7 #8.
- Thomas argued that copy-on-write filesystems could not provide the intended guarantee against later out-of-space errors anyway, making ftruncate worth exploring; he proposed a method selector covering allocation, truncation, and zero writes, with a separate threshold #5 #10.
- Thomas posted that redesigned selector, and Dimitrios tested its ftruncate path on PostgreSQL 17, observed the calls with strace, and confirmed that Btrfs compression worked during a large restore #13 #14.
- Thomas supported backpatching the workaround but favored leaving threshold tuning out of stable branches, while Bruce Momjian warned that existing postgresql.conf files would not automatically acquire a newly added setting #16 #11.
- Magnus Hagander reopened the case for tablespace-specific behavior, prompting discussion of replicated catalog settings versus server-local filesystem choices; Thomas ultimately chose to proceed with global settings and defer a general solution for tablespace configuration #18 #19 #20 #21.
- After the delay, Thomas reported successful light ZFS testing, and Jakub pointed out that the fix should reach PostgreSQL 16 and proposed making the settings reload-controlled rather than user-settable; Bruce reiterated the need for release-note guidance #23 #24 #25.
- Thomas split the revised work into a minimal posix_fallocate/write_zeros selector for backpatching, a configurable threshold, and ftruncate/chsize options for development only, adding documentation and adopting PGC_SIGHUP #26.
- Jakub reviewed and lightly tested the revised series successfully, supported landing the minimal selector, questioned the practical value of threshold tuning, and noted that he had not tested the third patch #27.
- Thomas landed only the minimal selector as 33e3de6d77, taking the first patch from #26 with base adaptations and editorial changes and backpatching through PostgreSQL 16; he confirmed that the remaining patches would return for development later #29 #31.
- The stable-branch GUC changes triggered crake's ABI checker; Andrew Dunstan discussed suppression support, while Tom Lane demonstrated that outside code could not take sizeof the incompletely declared array and recommended updating .abi-compliance-history before the release freeze #32 #35 #36 #37.
- Thomas posted the ABI-history update, Tom approved it and cautioned that broad type suppression needed research, and Thomas committed it with a small explanatory wording change as 3c3b34bbee #38 #39.
- The thread ended with the minimal workaround and ABI-history correction landed and Thomas awaiting crake's result; configurable thresholds and ftruncate/chsize remained deferred, while a longer-term ABI suppression solution was still unresolved #31 #39 #40.