model-writtenrun 9547 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Dec 2018 - Apr 2019 ยท 17 entries
- Alvaro Herrera proposed monitoring ordinary, concurrent, and partitioned CREATE INDEX operations, separating generic wait and partition metrics from AM-specific build phases, then posted an initial generic skeleton in #1 and #2.
- Pavan Deolasee identified missing validation reporting and stale wait counters in #3, while Rahila Syed proposed clearer phase information and blocker PIDs in #4; Herrera retained AM-specific subphases because heap scanning alone would conceal lengthy btree sorting and loading work #5.
- The expanded patches added an AM callback for phase names, btree scan and tuple counters, blocker PIDs, and validation metrics, while leaving sorting itself without numerical progress measurements #7 #8.
- Amit Langote's callback-naming review and Tatsuro Yamada's view-definition corrections were incorporated alongside block-based validation reporting, unsupported-AM handling, and monitoring documentation #11 #12 #13.
- Herrera supplied hash-index reporting as a delta, then bundled it with the main patch while fixing stale block counters identified by Syed #14 #16 #17.
- Robert Haas argued that numbered phase labels were misleading when phases had unequal durations or could be skipped; Herrera removed them while rebasing over CLUSTER monitoring and table-AM changes and extracting a scan-progress helper #21 #27 #31.
- After coordination with Andres Freund about table-AM integration, the next revision moved scan reporting into the heap implementation, absorbed the helper patch, and extended basic build reporting to all index AMs #34 #35 #36.
- Per-AM instrumentation for concurrent validation index scans remained incomplete and was left for subsequent work because Herrera considered it a relatively small portion of the operation #36.
- Simon Riggs continued advocating phase numbers, but Haas recommended treating numbering as a separate proposal, potentially using separate view columns, and the feature proceeded without numbered labels #38 #40 #41.
- Herrera discovered that nested index builds overwrote CLUSTER progress counters; after discussion with Freund, he assigned non-overlapping parameter slots and enlarged the progress parameter array #42 #43 #45 #46.
- A proposal to expose arrays of blocker PIDs or object OIDs was deferred after Haas raised fixed-memory, overhead, and complexity concerns #46 #47 #49 #50.
- Syed's testing and review exposed regression-output problems and fragile parallel block accounting, leading to centralized total-block reporting and a final scan-completion update; further partition testing persuaded Herrera that the reporting was sufficient #30 #51 #52 #54.
- Herrera committed both final CREATE INDEX patches from #51 as ab0dfc961b, with light naming, documentation, and API-call cleanup announced in #55.
- Peter Eisentraut then proposed reusing the infrastructure for REINDEX and REINDEX CONCURRENTLY, initially leaving later concurrent wait phases unfinished #56.
- Herrera recommended reporting those waits and documenting the additional phases; Eisentraut's revision added them, exposed the index OID, and clarified locker-versus-snapshot terminology, while deferring index-count and command-identification refinements #57 #58.
- Eisentraut committed the revised REINDEX patch as 03f9e5cba0, additionally enabling older-snapshot wait reporting; the thread ends with his commit announcement and both features landed #60.
- The later CLUSTER/VACUUM FULL repair, 6212276e43, guards REINDEX progress initialization and termination and originated in thread #37325; its link to this thread recalls the earlier nested-progress problem in #42.