model-writtenrun 8815 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Ildar Musin reported a reproducible duplicate-LSN assertion during logical slot creation and proposed assigning tuple CID mappings to the subtransaction in #1, crediting Murat Kabilov and Ildus Kurbangaliev with discovering and reproducing the failure.
- Thomas Munro found a tuple-CID lookup crash with the initial patch in #2, and Masahiko Sawada explained in #4 that CID mappings must remain attached to the top-level transaction, replacing that approach with decoding ASSIGNMENT records before the snapshot is complete.
- Alvaro Herrera attempted to limit early assignment processing to the BUILDING state in #5, although the attached condition still returned for all states below FULL_SNAPSHOT.
- Andres Freund explained in #6 why NEW_CID records matter for recognizing catalog-changing transactions, proposed assigning children directly while processing those records to avoid tracking unrelated transactions, and identified a separate risk of incorrect snapshots during fast-forwarding.
- Petr Jelinek proposed removing DecodeHeap2Op()'s early fast-forward check in #7, and Simon Riggs supplied that separate fix with explanatory comments in #8.
- Alvaro Herrera implemented the child-assignment alternative in #9, but encountered an error and an assertion that required weakening two sanity checks; Andres Freund replied in #10 that assignment should occur regardless of snapshot state and that the proposed conditional was unreachable.
- After initially reporting successful tests in #11, Alvaro Herrera reproduced the original failure and reported that the alternative still did not fix it in #13.
- Alvaro Herrera posted both competing implementations in #16, confirming that Sawada's early-ASSIGNMENT fix resolved the assertion while the revised NEW_CID child-assignment implementation still failed.
- The ASSIGNMENT attachment from #16 landed unchanged as bac2fae05c, leaving out the unsuccessful alternative; Alvaro Herrera announced the commit and backports in #17, and Andres Freund objected to the review timing in #18.
- Months later, Alvaro Herrera revived the outstanding fast-forwarding fix in #19, extending Simon Riggs's patch to execute standalone invalidations as well as process NEW_CID records.
- At the thread's end, the assertion fix had landed, while the fast-forwarding work remained under review: Amit Kapila asked in #20 about HEAP_INPLACE handling, commit-time invalidations, unnecessary CID mappings, and a reproducing test case.