model-writtenrun 7399 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
May - Jun 2021 ยท 17 entries
- Andres Freund proposed moving attcompression into existing padding to avoid increasing each pg_attribute entry from 112 to 116 bytes, and Tom Lane suggested grouping attbyval, attalign, attstorage, and attcompression together #1 #3.
- Michael Paquier supplied the initial reordering, and Tom expanded it to keep documentation and attribute-manipulating code consistent, discovering that equalTupleDescs() omitted attcompression and questioning relation-kind restrictions #6 #8.
- Andres investigated materialized-view compression and reported unnecessary recompression during concurrent refresh, then separately found that VACUUM FULL retained detoasted values until transaction end #9 #10.
- Tom committed the expanded reordering with additional initialization fixes, comments, and a catalog-version bump as f5024d8d7b, then proposed making zero attcompression mean the current default rather than resolving that default when a table is created #11.
- Michael patched the memory leak, and Dilip Kumar's review of redundant value copying prompted a simplification that reused the existing values array; Michael committed the revised cleanup as fb0f5f0172 #12 #16 #18 #19.
- The compression-default discussion settled on deferring to the current setting, with Justin Pryzby drawing a parallel to default tablespaces and Michael accepting the benefit of changing installation-wide behavior without altering every column #21 #22.
- Tom's new series removed relation-kind restrictions, redefined zero attcompression, and removed pg_dump's handling of the dump-time default; an add-on introduced COMPRESSION DEFAULT and fixed grammar that had accepted an empty ALTER COLUMN SET command #23 #25.
- Tom consolidated the series and removed the trivial default-compression accessor; Alvaro Herrera tested partitioned-table behavior, while Michael questioned test and documentation coverage that Tom identified as already present #27 #28 #31 #32.
- Andres then measured overhead from checking compression during VACUUM FULL even without recompressing data, and experiments showed that avoiding array initialization alone did not remove the cost of inspecting each attribute #33 #38 #41.
- Tom proposed dropping forced recompression, while Robert Haas argued for retaining some way to convert existing data and Dilip highlighted removing dependencies on compression libraries; the discussion considered optional recompression but faced the PostgreSQL 14 feature freeze #42 #45 #47 #49.
- The default-semantics work landed separately as e6241d8e03, based on #27 with a broader toastable-type check, additional output-order cleanup, and a catalog-version bump.
- Tom clarified that only forced recompression during VACUUM FULL was under consideration for removal, and that his goal was to avoid designing new user-visible behavior late in the release cycle #53 #54.
- Michael's wider-table benchmarks confirmed overhead without actual recompression, leading him to post a patch removing the checks and updating documentation and regression expectations #58 #61 #62.
- Alvaro benchmarked an alternative that inlined toast_get_compression_id(), initially recovering most of the performance in his test, then added casts to fix C++ checks; Ranier Vilela revised the helper to use a const argument and direct returns #63 #65 #66.
- Michael's 200-column benchmark still showed substantial overhead with Alvaro's revision, while Ranier proposed an additional loop and cleanup optimization; Alvaro concluded that investigating further would require time he could not provide #67 #68 #69.
- The thread settled on removing forced recompression for this release, leaving the proposed optimizations and any future explicit recompression facility uncommitted #70 #71 #72.
- Michael committed #62 with only a documentation wording difference as dbab0c07e5 and confirmed completion in the last message; the catalog layout and default-semantics changes remained, while VACUUM FULL and CLUSTER no longer forced conversion between compression methods #73.