model-writtenrun 503 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- wenhui qiu proposed making wal_compression=on select lz4 when available, otherwise pglz, to accompany the TOAST compression change; Christoph Berg supported targeting PostgreSQL 19 and Japin Li said the patch looked good #1 #2 #3.
- Michael Paquier suggested preferring zstd over lz4, explaining that TOAST lacked zstd support for separate compatibility reasons, and wenhui updated the patch to select zstd, then lz4, then pglz #4 #5.
- Fujii Masao opposed making the meaning of on depend on build options, prompting Japin Li to withdraw the CommitFest entry, but Michael, Jelte Fennema-Nio, and Christoph argued that users choosing on should receive an efficient algorithm; Jelte clarified that the actual default would remain off #7 #8 #9 #10 #11.
- Tomas Vondra revived the proposal by arguing that on should authorize PostgreSQL to choose an algorithm, while Andres Freund described pglz performance problems and the lower compatibility risk of changing WAL compression compared with TOAST #13 #14.
- Christoph supplied an editorial revision of wenhui's zstd-first patch and renewed the request to include it in PostgreSQL 19 #18.
- Tomas initially preferred matching TOAST's lz4 choice, and Andrey Borodin reported a hint-bit workload where lz4 was faster than both zstd and uncompressed WAL, leading Christoph to submit an lz4-first revision; Tomas reported no objections from the release management team to including the change in PostgreSQL 19 #20 #22 #23 #24.
- Michael argued that zstd's lower I/O cost and existing cloud-provider usage favored it, while Hannu Krosing proposed comparing compression levels; Tomas kept compression-level tuning outside the PostgreSQL 19 change and verified that WAL already used zstd's default level of three #27 #29 #30 #32.
- Tomas benchmarked OLTP and TPC-H workloads on four systems and reversed his preference to zstd because its better compression generally came with comparable performance; his attached wal_compression_level patch was explicitly testing infrastructure, not intended for commit, and the existing library compression levels were retained #34.
- Hannu questioned slower zstd recovery on Azure, and Tomas reproduced the difference and attributed it to decompression CPU costs on fast local storage; Andrey accepted zstd as the better general choice while suggesting context reuse as future work #35 #39 #40.
- Tomas tested decompression-context reuse from another thread and found that it nearly eliminated the Azure recovery gap, but deferred that optimization beyond PostgreSQL 19; Christoph and Michael endorsed retaining zstd as the preferred algorithm #41 #42 #43 #44.
- Tomas submitted the final revision restoring zstd-first selection and renaming the macro to WAL_COMPRESSION_ON to avoid implying that compression became enabled by default; Christoph requested consistent algorithm ordering in the documentation, and Daniel Gustafsson approved the patch with that correction #47 #48 #49.
- Tomas committed #47 with documentation ordering and commit-message edits as 931c9701f2, then confirmed the PostgreSQL 19 backpatch; the thread ended with wenhui's thanks, leaving the requested alias change landed and the benchmark-only GUC excluded #50 #51.