model-writtenrun 6409 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Mar 2021 · 1 entry
- Ashutosh Bapat reported WAL-sender memory exhaustion caused by unconfirmed speculative inserts retaining toast hashes, supplying reproduction steps and diagnostic instrumentation in #1.
May - Jun 2021 · 15 entries
- After a two-month gap, Amit Kapila proposed cleaning toast hashes when releasing transactions, and Dilip Kumar implemented cleanup in both transaction-return and streaming-truncation paths in #3, #8, and #9.
- Tomas Vondra confirmed that transaction-return cleanup solved his customer case, questioned streaming cleanup, and identified redundant null checks; Dilip explained the streaming constraints and submitted revised branch-specific patches in #10, #11, and #18.
- Further analysis established that incomplete speculative inserts were not streamed, but Amit identified a subsequent-delete hazard and Dilip reproduced both an assertion failure and failures when inserting into another relation in #20, #23, #25, and #27.
- Dilip consequently changed the design to queue speculative-abort records and clear toast state when processing them in #28; Amit deferred distinguishing toast-table abort records through new WAL information because the fix needed backpatching in #29.
- Dilip presented explicit abort handling alongside an alternative that tolerated leftover toast hashes in #31, and Amit selected explicit handling while retaining transaction-return cleanup for error paths in #32 and #33.
- The revised implementation gained an isolation test in #35, but Amit found incorrect lock identifiers and a scheduling race in #36, prompting the controller-session rewrite and separate test patch in #38.
- Amit and Dilip kept the test in test_decoding because moving it into the general isolation suite would require changing that suite's WAL configuration, while Alvaro Herrera requested a nonconflicting helper name; those revisions and back-branch patches appeared in #41, #43, and #44.
- Amit verified that the test failed without the fix and passed with it across branches, corrected a missing PostgreSQL 13 change-size case, and supplied assertion and comment refinements that Dilip propagated in #45 and #46.
- Before committing, Amit flagged intermittent failures in the test on which this one was based, and Dilip agreed to defer the test and posted code-only patches in #48, #49, and #50.
- Amit committed the code-only speculative-abort fix from #50 as 4daa140a2f, with backpatches through PostgreSQL 9.6, and announced the push in #51.
- Tom Lane then reported a buildfarm Valgrind failure in pgoutput.c in #52; Amit traced it to a pre-existing PostgreSQL 13 bug that allowed invalidation callbacks to inspect an uninitialized tuple-conversion map, rather than to the speculative-abort fix, in #53.
- Amit Langote supplied early RelationSyncEntry initialization in #56, and Dilip's review led to revised comments plus a HEAD comment-only companion in #57 and #58.
- Amit retained master's original comment wording and committed only the PostgreSQL 13 initialization fix from #58 as 357cb8f07f, leaving the HEAD comment edit out, as explained in #59 and confirmed in #60.
- Tomas Vondra subsequently noticed that inserting SPEC_ABORT before TRUNCATE changed an externally visible enum value and proposed appending it instead in #61; Tom Lane and Michael Paquier supported repairing stable branches while retaining HEAD's existing ordering in #64 and #65.
- The discussion-derived ABI repair landed as 56e366f675, and Amit's final message confirmed backpatching through PostgreSQL 11 because older branches already placed SPEC_ABORT last in #66; the fixes were complete, while the speculative-abort isolation test remained deferred pending reliable scheduling.