model-writtenrun 7395 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Aug 2018 - Mar 2019 · 11 entries
- Edmund Horner opened the thread with CTID ordering support and a separate range-scan executor, aiming to make bounded CTID queries efficient without constructing arrays of individual tuple identifiers #1.
- David Rowley proposed extending the existing TidScan to combine individual TIDs and ranges, and Edmund implemented that approach with range sorting, merging, and new selectivity estimates #2 #4 #8.
- David's testing exposed incorrect costing and the failure to use all available range bounds; Edmund responded by resolving arbitrary AND-lists in the executor and splitting selectivity, execution, backward scanning, and ordering into four patches #10 #12.
- Further review found offset-overflow errors and obsolete tuple-storage comments, producing corrected range handling and an independent comment-cleanup patch for sequential and index scans #13 #15 #17.
- David supplied a selectivity correction, Tomas Vondra reviewed range-array allocation, and Edmund incorporated their feedback alongside a separate adjustment for the expected half-full final page #18 #19 #21 #23.
- Tom Lane questioned the value of explicit descending CTID ordering and directed pathkey generation toward existing planner helpers; Edmund also explained why nullness estimates affect range selectivity, leading to agreement that system-column nullness should be handled separately #25 #32 #34 #35 #36 #37.
- Changes to TidPath's RestrictInfo handling exposed the complexity of mixing OR and AND semantics, so Tom recommended separate range path and executor types restricted to top-level inequalities; Edmund accepted the loss of multiple-range support and posted the redesign without ordering pathkeys #38 #39 #40 #43.
- David tested the selectivity estimator against actual row counts, and Edmund refreshed the series for the new table AM API while identifying remaining heap-specific dependencies #44 #47 #49.
- With table AM questions unresolved and the release deadline approaching, Tom committed the two selectivity patches from #49 as f7111f72d2, making light editorial changes and leaving executor work for later #54.
- Andres Freund advised using optional table AM support and avoiding direct heap access; after further discussion, Edmund posted the remaining range executor with an optional block-limit callback and removed backward cursor scanning to simplify the implementation #52 #53 #57 #61 #63.
- By the end, selectivity estimation and range execution had landed and the node-output omission was reported fixed, while CTID ordering and MIN/MAX optimizations remained deferred extensions and the earlier mixed-range design had been abandoned for simplicity #40 #63 #93 #97.
Jul - Sep 2019 · 2 entries
- David corrected the planner's callback-support flag and simplified the patch, while Edmund favored retaining enable_tidscan as the shared control rather than adding another setting #64 #65.
- Andres then objected that block-based limits assumed too much about storage layouts and called for an API expressed in TID ranges; Edmund described an attempted implementation but eventually requested help with the heap internals, and the patch was returned with feedback #66 #68 #70 #77 #78.
Jan - Feb 2021 · 6 entries
- After more than a year, David Fetter revived the last posted version with a rebase, motivated by possible parallel dumps of large tables, and Edmund checked the updated expression-fixing call #79 #80.
- David Rowley took over implementation, moved range enforcement into the access method through an optional ItemPointer-based callback, restored backward scanning, added rescan tests, and represented AM capabilities with flags #81.
- The revived series included a prerequisite heap backward-scan fix discussed separately; once that fix landed, David reposted only the range-scan remainder #81 #82.
- Zhihong Yu identified small code and comment defects, and Andres questioned the lifetime of bounds passed repeatedly to the fetch callback; David incorporated the corrections and split range setup from tuple retrieval, storing bounds in the scan descriptor #83 #84 #85 #87.
- David aligned API names with existing conventions, then addressed Andres's rescan concerns by making range changes restart the scan; he rejected moving range-expression construction into planning because the extra node and EXPLAIN machinery outweighed the benefit #88 #90 #92.
- David committed the final posted range-scan patch as bb437f995d, renaming the restart wrapper to table_rescan_tidrange, simplifying empty-table tests, adjusting comments, and adding cursor-plan lookup support #92 #93.
Jun 2021 · 1 entry
- Peter Eisentraut later found that TidRangePath lacked node-output support and supplied a focused fix; Edmund and David acknowledged the omission, and Peter reported pushing it in the thread's final message #94 #95 #96 #97.