model-writtenrun 2484 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Aug - Oct 2024 · 12 entries
- Bertrand Drouvot proposed pg_logicalsnapinspect in #1, providing SQL functions for snapshot metadata and contents to aid debugging and education, with isolation tests and documentation.
- Amit Kapila questioned exposing snapshot structures in #2; after discussion of accessor functions and Bharath Rupireddy's support for built-in functions in #6, Bertrand supplied a core implementation in #8.
- Amit argued that low-level debugging functions fit a contrib extension in #9, and Bertrand proposed exposing the infrequently changed serialized structures through an internal header in #12; Amit endorsed that arrangement and a broader module name in #13.
- The revised extension in #14 returned to the original implementation as pg_logicalinspect, leaving inspection of other logical-decoding files as a possible future extension rather than part of this submission.
- Shveta Malik's reviews in #15, #17, and #23 improved directory macros, header naming, tuple-column indexing, and reader naming; David G. Johnston explained why NULL input should not cause function-resolution failure in #19.
- Peter Smith supplied detailed cleanup suggestions in #25 and #27, while Shveta and Amit advocated readable snapshot states; #30 incorporated textual states, clearer tests, and permission documentation.
- Shveta's testing found negative checksum values in #32, leading to an int8 checksum result in #34, and Peter Eisentraut's error-handling review in #38 led to the simpler open-failure path in #39.
- Masahiko Sawada's review in #41 prompted the substantial revision in #43: snapshot restoration became shared core code, SQL functions accepted filenames directly, and a switch replaced the state lookup table, eliminating duplicated reading code and extra CRC exports.
- Further review in #46 identified incorrect xid-array construction and unnecessary memory-context and file-opening duplication; #53 corrected those issues and introduced a separate XIDOID array-helper patch.
- Peter Smith questioned unused deconstruction support in #55, and #56 retained only construction support, improved initialization and validation order, and named the shared reader SnapBuildRestoreSnapshot.
- Masahiko repackaged the work in #59, separating structure relocation from the extension, folding array-helper support into the extension patch, and supplying small count-type and documentation changes; Bertrand accepted this split and incorporated Peter's final comments in #62.
- After Peter accepted the final differences in #63, Masahiko committed the two parts of #62: e2fd615ecc relocated the structures with a redundant-typedef correction, and 7cdfeee320 added the extension unchanged from its attached component.
Mar 2025 · 5 entries
- Months later, Andres Freund reported an isolation-test failure with an unexpected third snapshot in #65; Masahiko attributed it to an additional RUNNING_XACTS record and suggested checking for suitable snapshots instead of exact counts in #66.
- Bertrand supplied that fix in #67 and simplified and documented the test in #69; Tom Lane reported failures without Valgrind in #70, and Bertrand reproduced the problem against the original extension commit in #72.
- Masahiko found that a checkpoint could still remove a required snapshot in #73, while Amit identified concurrent autoanalyze as another source of varying counts in #74; #75 kept a transaction open through inspection and checked minimum counts in a snapshot containing both ongoing and committed catalog changes.
- Amit accepted the autoanalyze handling in #76, Masahiko supplied cosmetic refinements in #77, and Bertrand approved them in #78; a49927f04c landed that final attachment unchanged.
- The thread ended with Euler Taveira's whitespace report in #80 and Masahiko's confirmation in #81 that David had already fixed it; the submitted extension and test-stabilization work had landed, with no attached patch left awaiting review.