model-writtenrun 8363 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Jun 2020 ยท 15 entries
- Kyle Kingsbury reported Jepsen/Elle finding two committed transactions that failed to observe each other's writes under SERIALIZABLE on a healthy PostgreSQL 12.3 instance #1.
- Peter Geoghegan clarified PostgreSQL's isolation-level mapping, while Kyle reproduced the anomaly without ON CONFLICT and questioned why the manual did not explicitly identify REPEATABLE READ as snapshot isolation #4 #10.
- Daniel Verite and Kyle distinguished the SQL standard's anomaly definitions from later research definitions, establishing the documentation problem alongside the implementation bug #16 #17.
- Kyle ruled out autovacuum timing, provided a simpler local testing workflow, and reproduced the problem on PostgreSQL 9.5 through 12; Peter subsequently reproduced it on PostgreSQL 13 after resolving test-harness setup issues #18 #20 #26.
- After SQL-log analysis exposed ordering ambiguities, Peter captured an execution with rr chaos mode, and Kyle mapped the reported dependency cycle onto its exact SQL operations #42 #43.
- Peter traced the missed conflict to a third transaction updating an invisible, concurrently inserted tuple and posted a tentative fix selecting the original inserter's XID; his Jepsen runs then stopped reporting anomalies #45.
- Peter refined the fix, Andres Freund requested an isolation test, and Thomas Munro identified an assertion failure that required rejecting XIDs older than TransactionXmin; the resulting version also established that updater abort was not necessary to trigger the bug #47 #51 #54.
- Thomas tested the final posted bug fix, confirmed that it detected the anomaly and passed check-world, and proposed small test cleanups #56.
- Thomas opened two independent follow-ups: manual references explaining SI and SSI plus catalog visibility, and corrected comments describing the conflict-out API contract #57 #58.
- Peter committed the bug fix from #54 as 5940ffb221, retaining its executable logic while refining comments and tests, including DELETE coverage, and announced backpatches to all supported branches #60.
- Kyle requested fuller isolation-level comparisons, while Andres cautioned against dense academic terminology and asked that catalog wording distinguish internal access from explicit queries; Thomas incorporated concise clarifications and retained external references for the detailed theory #59 #62 #64 #65.
- Following Peter's approval #61, Thomas committed the independent comment cleanup from #58 unchanged as 7aa4fb5925, with backpatching to PostgreSQL 13.
- Thomas's final documentation revision explicitly compared SI and SSI with locking implementations, while leaving a possible SNAPSHOT ISOLATION alias and broader user guidance for future work #70.
- Peter requested a clearer statement that other systems' separate isolation levels also have different behavior, and Thomas incorporated it and standardized the bibliography before committing #70 as 1575fbf1ac #72 #73.
- All three patch strands had landed by the thread's end; the remaining discussion corrected a comparison with MySQL's REPEATABLE READ behavior, which Thomas acknowledged without proposing another patch #74 #75 #76 #77.