model-writtenrun 8606 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Andres Freund opened the thread because txid_current() returned bigint without a convenient conversion to xid, complicating system-column regression tests and calls to age(xid) #1.
- Thomas Munro proposed a dedicated unsigned 64-bit transaction-ID type and a new function family, while Tom Lane and Andres Freund supported keeping that type outside the numeric hierarchy #2 #5 #6 #7.
- Thomas posted a two-patch proof of concept providing xid8, an explicit cast to xid, and parallel xid8 transaction/snapshot functions that shared C implementations with the older bigint-based functions #9.
- Takao Fujii found an incorrect snapshot upper-bound comparison and requested missing FullTransactionId comparison helpers; Thomas incorporated those fixes and Tom Lane's advice on catalog OID allocation while retaining the SQL name xid8 #12 #11 #13.
- Yoshikazu Imai questioned the changed wraparound handling, and Thomas explained the practical lifetime of 64-bit IDs while distinguishing this groundwork from eliminating wraparound in the existing heap implementation #15 #16.
- After a rebase, Mark Dilger reported successful installcheck testing, and Thomas later corrected the function-name typo Takao had identified, updated error messages, and reported check-world and 32-bit testing #18 #19 #14 #20.
- Fujii Masao's review led to ordering operators, btree and merge-join support, expanded documentation, and separate tests for the new and compatibility APIs; Thomas left broader input-validation improvements outside this project because of portability complications #21 #22.
- Thomas clarified the snapshot-to-full-ID conversion assumptions, added btequalimage support, and subsequently renamed txid.c and internal identifiers to reflect the new API while preserving shared implementations #23 #24.
- Alvaro Herrera proposed an epoch/value display format, but Andres Freund argued for retaining plain 64-bit values and keeping migration of existing system views separate; Mark Dilger accepted the existing input/output representation while highlighting transition concerns #27 #28 #32 #33.
- Mark's review prompted a third patch migrating remaining test callers to the new functions, and he reported that the expanded series passed check-world on macOS while suggesting clearer snapshot documentation #25 #35 #36.
- Thomas consolidated the series back to two patches and chose core-style names such as pg_current_xact_id(), pg_current_snapshot(), and pg_snapshot; Mark examined the resulting xid/xid8 mismatches, and Thomas defended explicit narrowing casts while discussing future interface migration #37 #38 #39.
- Thomas committed both components of #37 as aeec457de8 and 4c04be9b05, adding only catalog-version bumps to the posted changes, and announced that the naming could still be reconsidered before release #40.
- After landing, Mark raised concern about two differently typed functions named TransactionIdInRecentPast, and Thomas explained that the duplicate naming predated this work; no corrective patch followed in this thread #41 #42.
- Robert Haas and Alvaro Herrera continued to argue that an epoch/value representation could help troubleshooting, while Andres maintained that another representation would increase confusion, leaving that design debate unresolved #43 #44 #45 #46 #47.
- The main xid8 series remained committed, but Thomas ended with an explicitly experimental xid-to-xid4 rename that lacked a suitable backward-compatible alias mechanism and documentation, leaving that separate proposal unfinished #48.