model-writtenrun 4510 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Jun - Jul 2021 · 3 entries
- Paul A Jungwirth opens the thread with application-time PERIODs, temporal primary/unique keys, FOR PORTION OF updates/deletes, and temporal foreign keys, crediting Vik Fearing's foundational PERIOD work and supporting native range columns throughout #1.
- The initial design implements temporal uniqueness through GiST exclusion semantics and deliberately treats null period bounds as unbounded; discussion with Corey Huinker keeps system versioning separate and preserves the pg_period catalog approach #1 #9 #10 #12.
- Early revisions correct bounds handling, catalog documentation, build problems, and an omitted trigger source file, while partition support and concurrency questions remain open #2 #3 #8 #14.
Nov 2021 · 1 entry
- Peter Eisentraut challenges hardcoded operator names and the trigger implementation's primary-key prerequisite, and requests a separate BRIN rename; the unchanged rename hunks from #14 land as ee41960738 #15 #21.
May 2023 · 1 entry
- After a long gap, Paul replaces FOR PORTION OF triggers with executor logic so the feature no longer requires a temporal primary key, and Peter supplies build fixups #24 #25.
Jul 2023 · 2 entries
- Following PGCon discussions, the series puts range-based functionality first, separates cascading referential actions, and moves named PERIODs last using generated range columns; btree_gist need not first move into core #27 #29.
- Paul incorporates Peter's fixups and adds partitioned-table support, while PERIOD dependency handling remains unfinished #30 #32.
Aug 2023 - Apr 2024 · 12 entries
- Jian He's crash reports and NO ACTION, RESTRICT, and deferrability tests drive successive foreign-key fixes, followed by a btree_gist operator-lookup correction #39 #41 #46 #44 #47 #49.
- Discussion explores multiple WITHOUT OVERLAPS columns, but Paul finds multidimensional containment too complex and favors extensible types; Peter accepts retaining a single temporal column #34 #35 #51 #62.
- Jian supplies executor and column-lookup fixups, Peter cleans up names and parsing, and Paul folds selected changes into the full series; subsequent discussion drops UNBOUNDED syntax and confirms that implicit leftover inserts need no separate INSERT privilege #56 #57 #59 #62 #65 #67 #71 #73.
- The mismatch between GiST and btree_gist strategy numbers leads Paul to propose an optional translation function, accepted by Peter and implemented alongside aggregate and without_portion extensibility for multiranges and custom types #65 #72 #74 #77.
- Partition attribute mapping and review fixes prepare the first landing: Peter commits the strategy-function and primary-key components of #88 as 6db4598fcb and 46a0cd4cef, with light documentation and error-message edits #78 #80 #90.
- Paul resends the remaining work with generated-column dependency and transition-table improvements, and the separate comment correction in that submission lands unchanged as 86232a49a4 #91.
- Peter recommends a shared conperiod catalog flag and simpler foreign-key infrastructure; Paul incorporates Jian's test cleanup and changes leftover functions to SETOF, while the rename from #100 lands with small documentation additions as 030e10ff1a #94 #95 #97.
- Peter supplies an unreserved-PERIOD fixup and asks to hardcode range_agg, leading to shared RI routines and a simpler core foreign-key patch, with generalized support temporarily separated #101 #105.
- Test readability improvements from #108 land separately as a88c800deb for daterange/YMD literals and 5577a71fb0 for half-open notation, the latter extracted from the larger FK patch.
- Peter accepts inferred temporal REFERENCES as an extension, and the series drops generic aggregate support in favor of built-in ranges and multiranges; deferred-check and collation review culminates in the first two components of #120 landing together as 34768ee361 #110 #111 #117 #119 #121.
- Paul then discovers that empty ranges violate assumptions about temporal uniqueness, exposing ON CONFLICT and planner problems, and starts a separate repair series with an eventual proposal to generate a not-empty CHECK constraint #123 #128.
- The replica-identity regression test from #128 lands as 482e108cd3, and the revised ON CONFLICT component from #132 lands as 144c2ce0cc after removal of an unnecessary IndexInfo field.
May 2024 - Mar 2025 · 21 entries
- Jian adds internal dependencies for the implicit CHECK constraint and Paul integrates them, but empty-value semantics and constraint lifecycle behavior remain unsettled; Matthias also identifies broader GiST uniqueness and index-administration inconsistencies #130 #140 #133.
- Peter judges the unresolved problems too risky for PostgreSQL 17, Paul agrees to revert, and the ten rollback attachments in #146 are squashed essentially unchanged into 8aee330af5 #142 #143.
- After debate about mathematically correct empty-range behavior, Paul chooses direct execution-time rejection for both temporal primary and unique keys, while Robert Haas advises avoiding another catalog flag #148 #151 #152 #154.
- Paul revives the full series with executor checks and catalog bookkeeping; Jian offers a nonempty-overlap operator alternative, but Paul's experiment shows that rejected inserts can leave index entries that make later checks fail #156 #159 #161 #162.
- The final empty-value design moves validation into check_exclusion_or_unique_constraint, eliminating the special operator and catalog/relcache machinery; further tests and review prepare a second landing #164 #165 #168.
- Peter approves and reintroduces the strategy, temporal-key, empty-check, and foreign-key work, announcing the second landing in September; Paul's benchmarks support retaining range_agg for the common valid-reference case #170 #172 #173.
- The GiST strategy-function component of #168 lands as 7406ab623f, unchanged apart from a catalog-version bump.
- The temporal primary-key and empty-value rejection components of #168 are combined into fc0438b4e8, with formatting cleanup, preserving the revised validation reviewed by Jian in #166, #167, and #169.
- The temporal foreign-key component of #168 lands with matching changed lines as 89f908a6d0, restoring range and multirange containment checks while leaving cascading actions and named PERIOD declarations for later.
- The remaining FOR PORTION OF work changes leftover insertion to cached SPI statements to obtain correct statement triggers, transition tables, and partition routing, while grammar placement is corrected and replication failures become a new focus #174.
- Jian's documentation findings lead to fixes in #177, which land as bf62105950, d56af4c882, and f683ba0867; the combined ALTER TABLE patch is split and lightly corrected #169 #180.
- Paul adds PL/pgSQL access to temporal trigger information, but Sam Gabrielsson reports that valid shrinking referenced ranges can fail NO ACTION checks; Paul responds with revised coverage queries and a proposed RESTRICT fix #177 #178 #187.
- The publisher-side replica-identity portion of #187 lands separately as 79b575d3bc, leaving subscriber lookup and replication tests for subsequent work #188.
- Peter reworks subscriber index selection around Relation metadata and per-opclass equality lookup, first extracting the simplification from #191 as 7727049e8f.
- Vignesh C reviews Peter's expanded replication series and Paul tests and revises it; the five components of #195 land as 321c287351, a2a475b011, 13544e790e, 74edabce7a, and 939b0908c8, with only light cleanup #192 #194 #196.
- Paul resumes the feature remainder with PERIOD adaptations for cataloged NOT NULL constraints, then removes the redundant intersect support function and resolves NOT ENFORCED conflicts #197 #199.
- Peter extracts test whitespace from #199 into 888d4523f0 and lands its NO ACTION coverage fix with light edits as 1772d554b0 #200.
- Tom Lane reports intermittent RESTRICT failures after that landing, and investigation suggests inappropriate reuse of cached plans; Paul's correction in #210 lands as 9926f854d0, omitting the diagnostic reproducer #201 #207 #209.
- The cache-key change eliminates observed failures, but the mechanism that could have confused constraints remains unexplained, so Peter calls the result provisional and Paul leaves the underlying cache mystery unresolved #213 #214.
- Peter proposes RESTRICT semantics that ignore the temporal column, while Paul argues that temporal matching must depend on the affected time span; lacking agreement, Paul removes temporal RESTRICT support in #218, which lands with a documentation addition as b83e8a2ca2 #215.
- Peter closes the PostgreSQL 18 scope because subsequent features still need structural and SPI review, and Paul rebases the uncommitted remainder while planning a new commitfest entry; their disagreement over RESTRICT semantics persists #220 #222 #223 #224.
May - Jun 2025 · 3 entries
- Robert Haas asks for clearer concepts, documentation, and execution behavior; Paul agrees to add an introductory temporal section, improve leftover terminology, and remove the unnecessary GiST support-function dependency, with trigger and permission details still requiring work #225 #226.
- Peter separately aligns the landed GiST support-function names with compare-type translation, Paul reviews the patch, and #227 lands unchanged apart from its catalog-version bump as 32edf732e8 #228 #229.
- At the thread's end, temporal keys and their fixes have landed and temporal RESTRICT is disabled, while FOR PORTION OF, cascading temporal actions, PL/pgSQL temporal trigger variables, and named PERIOD declarations remain uncommitted; Tom accepts the final operator-family terminology while leaving future generalization open #224 #226 #230.