model-writtenrun 5992 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Georgios Kokolatos proposed LZ4-compressed WAL streaming for pg_receivewal in #1, extending the directory writer while leaving tar-based archives for possible future work.
- Michael Paquier, Dilip Kumar, and Magnus Hagander reviewed the initial interface in #2, #3, and #4; the resulting #10 used --compression-method, added documentation and restart-position handling, and made external lz4 integrity checks conditional on command availability.
- The discussion settled on default-only LZ4 compression: dictionary support would complicate decompression, and acceleration controls offered insufficient benefit for their maintenance cost, as discussed in #13, #15, #16, and #17.
- After a two-month gap, #18 incorporated Michael Paquier's review in #11, replacing filename macros with is_xlogfilename(), requiring an explicit compression method, rejecting --compress=0, and initially choosing gzip level 5; Jian Guo's descriptor-leak report in #19 was addressed in #20.
- Michael Paquier's testing and review in #21 prompted LZ4 flushing, cleanup, and test fixes in #22, which also corrected diagnostic filenames to include compression and partial suffixes.
- The diagnostic cleanup from #22 landed separately as cddcf7842c, with minor edits, and Michael Paquier confirmed the extraction in #23.
- Michael Paquier supplied a revised patch in #25, correcting the compression callback's confusion between method and level, while identifying that declaring a full segment's contentSize caused errors when closing incomplete LZ4 streams.
- Following another gap, Georgios Kokolatos returned with #26, using the library's default gzip level and attempting to rewrite incomplete-segment headers; Michael Paquier explained in #28 that partial WAL must remain recoverable and archives compressed by external lz4 commands must also work.
- The replacement approach in #32 measured segment size by decompressing into a disposable buffer, abandoning contentSize-dependent validation and header rewriting, and split the series into an options refactor and LZ4 support at Michael Paquier's request in #31.
- Magnus Hagander and Robert Haas argued in #33 and #34 that the user-facing format should remain gzip rather than zlib, and #37 restored gzip terminology throughout the two patches.
- The options refactor from #37 landed as d62bcc8b07 with revised option validation, proper zlib header inclusion, and clearer compression-level field names; #38 reported the landing and supplied the remaining LZ4 patch.
- Manual testing exposed a decompression input-pointer error, fixed by Michael Paquier in #39; Georgios Kokolatos then corrected file-versus-global compression selection, heap-allocated the buffers, and refined the loop in #41.
- LZ4 support from #41 landed with light edits as babbbb595d; Michael Paquier reported testing all zlib/LZ4 build combinations in #42 and declined an additional oversized-segment test because of its extra I/O and command overhead.
- Jeevan Ladhe questioned unnecessary finalization after a header-write failure in #44, and Robert Haas supported removing it in #48; Michael Paquier agreed in #49, leading to the discussion-derived cleanup f79962d826.
- A February follow-up clarified that the directory-writer work had already landed and tar-method LZ4 support remained outside this thread's scope, as Michael Paquier explained in #52 and Robert Haas acknowledged in #53.
- Justin Pryzby reported the incorrect lz4frame.h inclusion in #54, producing f512efb2d5 to use angle brackets and place the library header before PostgreSQL headers; Michael Paquier confirmed the correction in the final message #55, leaving no submitted patch pending.