model-writtenrun 2865 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Nov - Dec 2019 · 3 entries
- Zhenghua Lyu opened the thread with a TPC-DS query whose enormous estimated cost overwhelmed the fixed disable_cost penalty, allowing a disabled nested loop to win #1.
- Andres Freund explained why DBL_MAX would not preserve meaningful comparisons between disabled plans, Robert Haas proposed counting violations separately from ordinary costs, and Tom Lane discussed the need for a count and its performance implications #4 #8 #11 #12.
- Jim Finnerty tried encoding disabled status in floating-point bits, but his inline experiment had regression failures and Tom Lane objected to making optimizer costs depend on floating-point representation #13 #17.
Aug 2023 · 1 entry
- After a long gap, Jian Guo supplied a boolean Path.is_disabled implementation; Robert Haas found incorrect comparisons between disabled paths and later demonstrated that disabled status must accumulate as a count through the plan tree #19 #20 #28.
Mar - Oct 2024 · 19 entries
- Robert Haas separately explored suppressing disabled index paths, and David Rowley's tests showed that disabling index-only scans could also eliminate useful ordinary index scans, prompting a revision #36 #38 #53.
- Attempts to avoid generating disabled paths more generally ran into mutually parameterized scans that left no valid join plan, while discussion also exposed a related CURRENT OF qualification-selection problem #41 #47 #49.
- Haas assembled forced-TID cleanup, revised index controls, and bitmap/gather-merge suppression experiments, then replaced the latter experiments with accumulated disabled-node costs and a separate EXPLAIN patch #53 #54.
- The counting design deliberately retained the oversized-hash-bucket disable_cost penalty because Haas considered planner memory estimates different from explicit user disabling, while Andres Freund benchmarked the additional path-comparison work #54 #55.
- Heikki Linnakangas approved the TID cleanup but questioned index-setting compatibility and cumulative EXPLAIN output; Tom Lane and David Rowley also raised compatibility concerns about changing the index controls #60 #66 #72.
- The TID component of #59 landed unchanged as e4326fbc60, replacing penalty arithmetic with generation of only the executor-supported CURRENT OF path.
- Haas then posted the remaining counting and EXPLAIN components, dropping the controversial index-control changes and explaining test-plan changes previously obscured by enormous, fuzzily equal costs #73.
- David Rowley's detailed review uncovered missing count propagation and redundant join counting, which Haas fixed; their subsequent discussion clarified why join prechecks need disabled-node counts during initial costing #74 #75 #78 #79.
- The two components of #75 landed unchanged in separate commits, e222534679 for planner costs and c01743aa48 for EXPLAIN, with Haas explicitly leaving display preferences open for further feedback #82.
- Jonathan Katz reported a pgvector HNSW failure after disabled-node priority overtook DBL_MAX costs; discussion proposed positive infinity as an unsupported-index-path signal, and Katz checked its behavior on older PostgreSQL versions #85 #86 #97 #100.
- Alexander Lakhin's UBSan report identified reversed cost_sort arguments, leading Richard Guo to commit the inline correction as 87b6c3c0b7; regression tests had missed it because they suppressed the affected displayed costs #101 #103 #105.
- Laurenz Albe objected to repetitive cumulative EXPLAIN counts, and David Rowley supplied a delta showing Disabled booleans only on the responsible nodes while repairing missing make_sort propagation #107 #108.
- Rowley's revised implementation handled SubqueryScan and CustomScan children explicitly and demonstrated why runtime-pruned Append children made the old cumulative output misleading; Haas remained concerned about maintaining the relationship between internal counts and displayed markers #122 #124 #130.
- Alena Rybakina proposed documentation for cumulative counts, while Laurenz Albe independently drafted an inline explanation of unavoidable disabled scans using the proposed boolean output #126 #128.
- Tom Lane proposed removing the remaining oversized-hash-bucket penalty, but Haas objected that treating memory estimates as disabled nodes could interfere with user and extension control; the alternatives remained unresolved in this thread #135 #140 #147 #150.
- Shayon Mukherjee's testing, Tom Lane's endorsement, and Alena Rybakina's inspection supported the boolean display, and Haas acknowledged sufficient consensus to proceed #132 #133 #138 #142.
- Rowley polished and tested the EXPLAIN revision in #152, incorporated Rybakina's missing ICU expectations from #153, and landed the combined change as 161320b4b9, replacing the cumulative presentation without reverting the underlying counting design.
- Rowley turned Albe's inline documentation into an attached patch using a regression-database example; their subsequent wording revisions clarified that many flags discourage rather than prohibit nodes and that fallback is intentional #156 #157 #158 #159 #160.
- The final documentation in #160 landed unchanged as 84b8fccbe5, crediting Laurenz Albe and David Rowley; the earlier cumulative-count documentation was superseded by the boolean explanation #162.
Jul 2025 · 1 entry
- At the thread's close, Michael Paquier reported that PostgreSQL 18's disabled-node priority simplified pg_hint_plan and improved plan interpretation, and Haas welcomed the result; the main implementation, display revision, and documentation had landed, while the residual hash-join penalty proposal had no recorded resolution here #163 #164 #150.