model-writtenrun 9610 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Peter Eisentraut proposed SQL-standard COMMIT AND CHAIN and ROLLBACK AND CHAIN to preserve isolation, read/write, and deferrability settings across transactions in procedures, supplying six preparatory cleanups plus SQL and PL/pgSQL implementations #1.
- Andres Freund argued that the submission was too late for PostgreSQL 11, while Simon Riggs considered the standard syntax a modest extension; Andres emphasized the transaction-control changes and new interface that needed review #2 #3 #4.
- Alvaro Herrera reviewed the preparatory cleanups and suggested comment corrections, after which Peter reported committing those cleanups with edits #5 #6.
- Heikki Linnakangas found that converting transaction_isolation to an enum stopped accepting the quoted string 'default'; Peter explained that this undocumented form was unnecessary because RESET and SET TO DEFAULT remained available #7 #8.
- Heikki distinguished the proposed SQL-standard behavior from the broader notion of chained transactions retaining locks and cursors, but the implementation remained focused on preserving transaction characteristics #9.
- After the patches became stale, Peter returned with the rebased two-patch chaining remainder #11, and separately landed the original enum-conversion component from #1 as f8c10f616f, documenting the quoted-default compatibility change #12.
- Fabien COELHO tested the chaining patches and requested psql completion and broader transaction-characteristic checks; Peter incorporated those changes, followed by SQL-conformance documentation and clearer, shorter tests across #14 and #19 #13 #15 #17.
- Fabien objected to breaking SPI_commit and SPI_rollback signatures and making their names conceal transaction restarting; Peter initially defended the change for other procedural languages and explained the SPI memory-protection flag, while Fabien proposed separate _and_chain entry points #16 #20 #21 #22.
- Andres also questioned the global storage for saved transaction characteristics; Peter retained it because a passed structure would expose transaction internals or require allocation surviving transaction completion #23 #25.
- Peter merged SQL and PL/pgSQL support into #24, accepted Fabien's separate SPI entry points, and explained that the SAVEPOINT regression case exercised distinct transaction-state branches confirmed by coverage analysis.
- Fabien successfully built and tested the merged patch and requested documentation indexing #26; Peter added those documentation details and landed #24 as 280a408b48, retaining the existing state-assignment order and internal function names #27.
- The thread ended with SQL and PL/pgSQL chaining committed and no attached patches outstanding; Fabien's remaining comment concerned internal helper naming, while chaining support for other procedural languages remained outside the submitted implementation #1 #28.