model-writtenrun 4507 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Apr 2024 - Feb 2025 · 31 entries
- Peter Eisentraut revived virtual generated columns in #1, proposing computation on read and separate, aligned stored/virtual regression files, with preparatory test and code cleanups.
- Corey Huinker endorsed the test split and questioned immutability and inheritance restrictions in #2; Peter kept the established inheritance rules while leaving possible relaxations for later in #4.
- The updated implementation in #3 replaced unfinished cases with explicit unsupported-feature errors, including indexes, not-null constraints, foreign keys, domains, and several ALTER TABLE operations.
- Tomasz Rybak reviewed the preparatory patches and successfully built and tested every cumulative stage in #5; Jian then reported RETURNING and MERGE failures and a pg_attrdef documentation omission in #6.
- The documentation report produced da486d3601 directly from discussion, while the two cleanup components of #3 landed as da2aeba8f5 and e26d313bad, as acknowledged in #7 and #8.
- Jian's subquery, statistics, and ANALYZE reports led to #14, and his further testing in #16 produced collation, recursive domain-type, TOAST, and statistics corrections in #17.
- Peter extracted the small ATExecSetNotNull() cleanup from #17 into 7ff9afbbd1, then committed its test-renaming and schema-isolation components as b9ed496925 and 894be11adf, leaving the feature remainder in #22.
- Dean Rasheed identified missing whole-row expansion in #15 and explained in #19 how ReplaceVarsFromTargetList() could replace the custom walker, reduce catalog lookups, and support future RETURNING OLD/NEW behavior.
- Nazir Bilal Yavuz independently submitted a transaction-ID exhaustion test-message correction in #25; none of the supplied candidate commits represents that unrelated patch.
- Peter implemented Dean's rewriting design as a second patch in #27, then consolidated it in #34 after resolving the interaction between RLS processing and foreign tables.
- The November update #35 incorporated additional tests and avoided unnecessary table rewrites, but accidentally restored an older rewrite layout; Dean supplied an RLS correction in #38, which Peter incorporated and acknowledged in #44.
- Amit Kapila's logical-replication investigation in #40 and Vignesh's row-filter testing in #41 exposed the need to distinguish evaluating virtual expressions in filters from actually publishing virtual-column values.
- Peter explained in #55 that publishing virtual columns required broader restrictions and validation than he intended to implement, while #59 added SET EXPRESSION, trigger-column support, stronger not-null checks, and publication-filter validation.
- Jian's CREATE DOMAIN diagnostic report in #42 became the independent patch #54, which landed unchanged as 84a67725cd and was confirmed in #62.
- Jian supplied trigger and executor improvements during December, including replacing trigger-written virtual values with NULL in #65; Peter incorporated that behavior and related tests into #69, retaining a shared helper to avoid duplication.
- Peter confirmed in #68 that expression overflow may occur on read because virtual columns behave like views, and in #71 and #77 retained the domain and immutability restrictions rather than adopting broader relaxations.
- Dean reviewed column-level access control in #81 and found a whole-row CHECK crash caused by an incomplete dummy RangeTblEntry in #82; #85 fixed the crash and added a separate logical-replication restriction patch.
- Peter separately committed 44b61efb79, removing obsolete stored-column suffixes from two regression-test table names to support the aligned test-file design in #1; this cleanup was not posted as an attached patch.
- Peter extracted the catalog-check relocation from #85 into 86749ea3b7, implementing Jian's suggestion from #63.
- The stored-column regression additions accumulated in #85 were extracted and extended as 41084409f6, crediting Jian and Dean as co-authors and including RETURNING OLD/NEW tests associated with #87.
- Dean supplied the RETURNING OLD/NEW integration in #87, and Vignesh moved the publication-option redesign to another thread in #90; Peter integrated the resulting interfaces and remaining feature work in #92.
- Dean identified an incorrectly scoped executor-validity flag in #93, and Shlok Kyal found a missing replica-identity restriction in #94; Peter supplied both corrections alongside the feature in #95.
- After Dean's final review in #97, all three components of #95 landed as 83ea6c5402, with additional publication regression tests, a catalog-version bump, and whitespace changes.
- Peter confirmed the feature landing in #99, retaining SQL-standard default display syntax without an explicit VIRTUAL keyword; indexes, foreign keys, not-null constraints, domain types, DROP EXPRESSION, and publishing virtual values remained outside that landing.
- Alexander Lakhin's SQLsmith report in #101 exposed missing nulling-relation metadata, which Zhang Mingli analyzed in #102; Richard Guo then demonstrated incorrect outer-join results and join removal in #103, #106, and #107.
- Jian's patches #104 and #108 propagated nulling bits but did not solve general outer-join semantics; Peter suggested temporarily rejecting affected expressions in #109, while Dean instead moved expansion into the planner in #110.
- Jian developed a parallel planner-based repair in #111, but Richard's alternative #113 reused the existing subquery-pullup replacement machinery, following his grouping-set analysis in #112.
- Dean refactored the shared replacement logic in #114 and added tests in #115, discovering a pre-existing grouping-set defect involving simple Vars; Jian independently reproduced that defect in #116 and proposed a separate adjustment in #117.
- Richard split the planner repair and callback refactor into separate patches in #118, fixing nesting-depth propagation, dropped-column handling, and expansion ordering; Dean refined the ReturningExpr handling in #119.
- Richard's nested-EXISTS example required expansion after SubLink pullup in #120, and Jian's review in #122 led to final comment, assertion, and regression-test improvements in #123.
- Following Dean's approval in #124, the two components of #123 landed unchanged as 1e4351af32 and 363a6e8c6f, with Richard confirming the push in #125.
May - Jun 2025 · 3 entries
- Alexander found another unexpanded reference in ALTER COLUMN TYPE USING in #126; Richard initially suggested expanding during table rewriting in #127, while Jian argued for earlier processing in #128.
- Richard's patch #129 expanded USING expressions before expression planning and rewrite decisions, received Jian's approval in #130, and landed unchanged as 5069fef1cf.
- The thread ends with Richard's push confirmation in #131: the main feature, preparatory test cleanups, planner repair, callback refactor, and ALTER COLUMN TYPE fix had landed, while the explicitly deferred capabilities and separate simple-Var grouping-set issue were not resolved by these candidate commits.