model-writtenrun 4747 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Michael Paquier proposed generating query-jumbling code from node definitions so utility statements with equivalent parsed trees would share query IDs despite textual differences, while retaining existing normalization behavior #1.
- Peter Eisentraut requested explanations for ignored fields and discussed location detection; Michael introduced explicit location tracking and additional documentation, and Peter supported ignoring fields by default through annotations while deferring broader corrections to existing jumbling semantics #3 #5 #6 #7.
- At Peter's request, Michael separated comment formatting and file relocation from the generator feature, adopted query_jumble naming, and reported benchmarks that raised the possibility of retaining string hashing for utility statements #6 #10.
- After resolving a misunderstanding about location handling, the series returned to ignoring locations unless explicitly marked and added an optional utility_query_id setting to choose string hashing or parsed-tree jumbling #11 #12 #13.
- Peter approved the preparatory patches with fewer location-comment changes, and Michael committed the relocation as 8eba3e3f02 and the reduced comment cleanup as 5d29d525ff from #13, then resent the remaining two patches #14 #15.
- Peter supplied a small fix-up and requested clearer Param handling, a custom jumbling attribute for RangeTblEntry, and better GUC documentation; Michael incorporated those changes and subsequently corrected the GUC variable used by the optional branch #16 #17 #18.
- Peter endorsed the updated implementation, requested centralized guidance on ignored fields, and argued for testing whether the string-hashing alternative could be removed #19 #20.
- Michael measured nanosecond-scale overhead for simple utility statements and comparable timings for ordinary queries, and configured a buildfarm member to exercise jumbling through pg_stat_statements; Peter favored keeping only one method #21 #23 #24 #26.
- Michael committed the generator portion of #18 as 3db72ebcbe, omitting the optional GUC, centralizing documentation, correcting field selection, and distinguishing zero-hash utility query IDs, while leaving reconsideration of the GUC open for beta #28.
- Michael subsequently found that A_Const values were not correctly incorporated into query IDs, causing utility statements with different values to be grouped together, and posted a custom type-sensitive implementation with regression tests #29.
- The A_Const fix from #29 landed as 9ba37b2cb6 with only formatting and test-comment corrections; Michael confirmed completion and left broader utility normalization for separate work #30.
- Tom Lane identified unused generated code for planner nodes, prompting Michael to propose no_query_jumble exclusions and an accompanying experimental patch to force jumbling during regression tests #31 #32.
- Andres Freund suggested exercising pg_stat_statements in existing pg_upgrade or streaming-regression tests, and Michael discussed coverage, memory sizing, and possible verification checks #33 #34.
- Tom recommended inheriting no_query_jumble from parent node types and retaining support for parsed-query nodes; Michael revised the exclusions accordingly, and Tom approved the revision while leaving questions about other ignored fields for separate analysis #36 #37 #38.
- Michael committed the exclusion portion of #37 as 2a507f6fd8, adding ForeignKeyCacheInfo to the excluded nodes, while deferring the test-forcing changes and broader coverage work to another thread #39.
- During PostgreSQL 16 beta, Michael reopened the question of retaining string-based utility hashing; Andrei Lepikhov supported a single tree-based method and separately suggested alternative query-ID algorithms for extensions #40 #41 #42 #43.
- Michael ultimately removed the optional utility_query_id GUC from the open-item list, leaving the generated tree-based implementation and its follow-up fixes landed, with broader testing and alternative algorithms outside this completed thread #44.