model-writtenrun 9579 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Amit Kapila opened the thread by identifying that multiple transaction-ID wraparounds between checkpoints could make txid_current() reuse an epoch-qualified value, and Alexander Korotkov confirmed the code lacked protection against that case #1 #2.
- Amit proposed explicit shared-memory epoch tracking, while Andres Freund preferred a single 64-bit next-XID value and suggested using existing CLOG WAL records for persistence; Tom Lane supported treating this as a separable step toward wider transaction IDs #3 #4 #6 #7.
- After Amit declined to implement the proposal, Thomas Munro supplied an experimental patch in July 2018 that widened the generator and checkpoint value and inferred recovery epochs from existing 32-bit WAL transaction IDs, without adding the proposed CLOG logging #10 #11.
- Andres recommended a struct wrapper to prevent accidental conversions, which Thomas adopted as FullTransactionId; subsequent review produced inline increment logic and explicit conversions, while briefly changing diagnostic output to full-width numbers #12 #16 #17 #18.
- Craig Ringer advocated extending full-width IDs to SQL-facing interfaces, but accepted Andres's argument that a new SQL type and its operators should be separate work #19 #21 #22.
- The changed control-data output broke pg_upgrade parsing, prompting Thomas to add a compatibility parser after check-world exposed the failure #23.
- Tom tested struct passing across several architectures and concluded that the wrapper was practical, while rejecting his earlier assert-build-only wrapper idea because it would create incompatible extension ABIs #24 #26.
- Thomas then narrowed the patch to the type, generator and conversions, dropping the control-data output changes and optional replication cleanup; he later rebased this reduced patch as development continued #27 #28.
- Review resumed in February 2019 with naming objections from Michael Paquier and Tom, while Andres argued that the new values should be understood as actual 64-bit transaction IDs and questioned the remaining epoch-inference code #31 #32 #33 #34.
- Thomas answered by rewriting the recovery advancement helper, documenting why its inferred epoch is bounded, removing unnecessary locking options, and reporting no measurable generator slowdown; widening WAL and two-phase state remained outside this patch #35.
- Thomas expanded the submission into two patches by adding FullTransactionId to the transaction stack and exposing full-width accessors for table access methods; Heikki's interest in a full-width RecentGlobalXmin led to further discussion of shared-memory size and locking costs #36 #37 #38 #39.
- The final submission replaced the transaction-state serialization layout with SerializedTransactionState and refined the accessor interfaces, while leaving wider user-facing output and explanatory documentation for follow-up work #40.
- Heikki tested the final series and found that its startup-process assertion rejected single-user recovery; Thomas restored the missing exception and committed the infrastructure component as 2fc7af5e96 from #40, with the fix confirmed in #42 following #41.
- Thomas also committed the transaction-stack component of #40 as ad308058cc, changing only an equivalent validity assertion before announcing both patches pushed #42.
- Jeff Janes subsequently noticed that GetNewTransactionId could return a stale full_xid after releasing and reacquiring its lock, but immediately confirmed that the bug had already been fixed through another thread #43 #44.
- Alexander Korotkov later proposed extending 64-bit IDs across non-heap code, SLRUs, multixacts and table-access interfaces; Robert Haas raised snapshot cache costs, and Andres suggested retaining compact transaction arrays while widening snapshot bounds #46 #47 #48.
- Alvaro Herrera opposed widening multixacts because of storage and upgrade costs, Alexander explored alternatives, and Robert emphasized that replacement locking schemes still need to handle large sets of shared lockers efficiently #49 #50 #51 #52.
- At the thread's end both submitted components had landed, while broader adoption remained design work: Thomas supported non-recycling SLRU addressing and described an undo-backed multixact-like experiment without submitting another patchset here #42 #53 #54.