model-writtenrun 9580 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Aug - Oct 2017 · 2 entries
- Peter Eisentraut proposed SQL-standard generated columns in #1, initially implementing computation on read while planning stored columns, using pg_attrdef for expressions and excluding generated columns from ordinary COPY column lists.
- Jaime Casanova's testing exposed missing values through views, inheritance crashes, and ALTER TABLE and trigger problems in #3 and #4; Peter addressed the view, inheritance, dependency, and DDL problems in #20, while explaining that computing virtual values for every trigger remained difficult.
Dec 2017 - Mar 2018 · 3 entries
- After proposals to permit timestamp-like expressions, Peter retained the SQL-standard immutability requirement in #22, explaining that stored versus virtual computation should be an optimization rather than a semantic difference.
- Michael Paquier reviewed #26 and identified unsupported-table crashes, unfinished STORED parsing, and missing tests in #27; Peter's #28 removed the unfinished syntax, added checks and tests, and improved virtual-column handling in procedural-language triggers.
- Virtual-column row types and trigger visibility remained disputed in #32, and Michael proposed shipping stored columns first in #36; Peter closed the work for that commitfest in #37 because he had not completed an update.
Oct - Nov 2018 · 4 entries
- Peter revived the work months later in #38 with both stored and virtual implementations, executor-side stored-expression evaluation, tuple-descriptor caching, and renewed discussion of the catalog representation.
- Sergei Kornilov found that adding a stored column failed to populate existing rows in #42, while Erik Rijkers demonstrated logical-replication initialization failures involving COPY in #48.
- Simon Riggs initially advocated replicating stored values in #54, but accepted Peter's explanation in #55 that subscribers must recompute their own generation expressions, making transmission of generated values unnecessary.
- Alvaro Herrera called for adapting stored-expression execution to the slot API and expanding tests in #60; Peter explained in #62 that generated partition keys remained unsupported because routing precedes stored-value computation and virtual keys had dependency-deletion problems.
Jan - Apr 2019 · 13 entries
- The updated implementation in #63 fixed ADD COLUMN population and added logical replication, postgres_fdw integration, more precise trigger behavior, caching improvements, and fuller documentation.
- Pavel Stehule exposed incomplete volatility checking in #64, and Michael found NULL-expression crashes and growing bulk-insert memory usage in #67 and #74; Peter addressed these issues and expanded tests in #77.
- Peter judged stored columns ready for PostgreSQL 12 in #77, but deferred virtual columns because expression expansion still missed cases and eliminating their physical NULL placeholders required unresolved planner and storage work.
- Peter retained the separate attgenerated catalog field in #78, concluding that the discussion had not established a compelling reason to combine it with default and identity metadata.
- Following Michael's request in #67 and follow-up review in #79, Peter extracted the unchanged walreceiver server-version API from #77 into 69039fda83 and announced the separate landing in #84.
- Peter submitted the stored-only remainder in #85, removing virtual-expression rewriting and concentrating runtime computation in the executor.
- Pavel reported a dropped-generated-column failure in #86, and Michael reproduced it while requesting uniqueness and prohibited-expression tests in #87; Peter fixed the stale attgenerated flag and added that coverage in #88.
- Pavel's dblink failure attachment in #89 proved to be checkout contamination, which he retracted in #90; his subsequent review in #91 confirmed successful check-world, documentation and behavior checks, and better performance than his trigger-based comparison.
- Peter committed #88 as fc22b6623b, with light rebase and catalog-version adjustments, and announced completion in #92.
- After landing, Justin Pryzby reported a documentation remnant and unintended acceptance of GENERATED BY DEFAULT in #94; Peter reported both fixed in #97.
- Jeff Janes reported that the new catalog field broke upgrades between development snapshots in #96; Peter explained in #99 that such compatibility was not maintained, and Michael withdrew his contrary concern in #101.
- Erik's file_fdw report in #95 developed into a reproducible case in #104, with Amit Langote connecting the failure to explicit COPY column lists in #105; Peter announced a fix in #106.
- Amit questioned the remaining NULL values returned by file_fdw in #107, and Peter explained foreign-table storage semantics in #108; Amit requested documentation in #109, Peter confirmed its addition in #110, and the thread ended with acknowledgment in #111, leaving the stored feature and prerequisite API landed while virtual columns remained deferred.