model-writtenrun 5194 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Feb - Mar 2021 · 5 entries
- Andrey Borodin proposes selectable PGLZ, LZ4, and zlib compression for WAL full-page images after measuring a 21% pgbench improvement with LZ4, while treating compression of all WAL as separate future work #1.
- Justin Pryzby requests build integration and considers a shared compression API, while Michael Paquier objects to enlarging the WAL image header and recommends rejecting unsupported algorithms when configuring the setting #2 #3.
- Justin takes over revision work, keeps the small compression calls local instead of requiring a generic API, adds conditional GUC choices and build support, and expands the series with Zstandard and CI-only defaults #6 #7.
- The March revisions add documentation, compression-level experiments, and correct decompression lengths around page holes; Justin then replaces the extra header byte with existing flag bits and allocates buffers large enough for every supported compressor #9 #10 #11.
- After a March rebase and a gap until May, Michael renews review for PostgreSQL 15, asks for controlled performance measurements, and advocates one existing wal_compression setting plus explicit compression-method flags #12 #13.
May - Jun 2021 · 8 entries
- Justin updates build support and documentation, then converts wal_compression into an enum, reduces its flag requirements, and exposes the method in pg_waldump; Dilip Kumar requests clearer flag names and recommends LZ4_compress_default() #14 #16 #17.
- Justin argues that CPU-bound and I/O-bound workloads need different choices, supplies index-build measurements, and adds both configurable compression levels and compression details in recovery context #21.
- Discussion involving Andres Freund and Heikki Linnakangas examines configuration privileges, after which Justin clarifies that the actual patch retains PGC_SUSET; Michael and Justin also settle on preserving PGLZ as the meaning of existing true/on aliases #24 #28 #27 #29 #31.
- Michael substantially rewrites the series for benchmarking, using individual method flags and a separate level setting, and finds LZ4 close to uncompressed CPU cost while Zstandard offers stronger compression; he recommends dropping zlib and adjustable levels from the initial feature #33.
- Andrey favors retaining Zstandard, while Heikki raises compression of multiple pages together; the discussion leaves grouped-page and whole-WAL compression as separate work and favors a minimal LZ4 implementation first #34 #35 #37 #38 #43.
- Michael splits the implementation into a near-committable LZ4 patch and a Zstandard extension needing further work, and Andrey's review leads to clearer unsupported-build errors and an explicit unknown-method label in pg_waldump #39 #42 #45 #46.
- The imported recovery-test and XID-persistence fixes are ultimately excluded because the failure predates alternate compression methods and depends on WAL-size fluctuations, so participants redirect that issue to its own discussion #51 #52 #54.
- Michael commits the LZ4 portion of #42 as 4035cd5d4e, adding exhaustive enum switches, LZ4_compress_default(), and corrected error returns after testing all three compression settings; he resends only the Zstandard remainder, still lacking complete MSVC support #55.
Sep 2022 · 6 entries
- In September 2022, Justin reports that Zstandard support has also landed and revives the recovery-context hunk omitted from his earlier series because replay on a build lacking the required compressor produces an unhelpful error #57.
- Justin reproduces the missing-Zstandard recovery failure and proposes an inline change that exposes RestoreBlockImage()'s existing error buffer through ereport() #58.
- A separate question about combining compression flags establishes that stacking algorithms is unsupported, with Justin recommending compression-level tuning instead #59 #60.
- Michael replaces the context-only proposal with a broader attached fix that generates messages for invalid block states and propagates RestoreBlockImage() errors through both backend callers, retaining ERRCODE_INTERNAL_ERROR #62.
- Justin questions that error code for unsupported build options, but Michael declines additional return states or error-code plumbing because the detailed message identifies the problem and preserves the frontend-compatible interface #63 #64.
- Michael applies #62 unchanged apart from rebase and whitespace as df4a056619 for development and PostgreSQL 15, closing the diagnostic issue; the last message leaves open whether older releases should receive the same improvement #65.