model-writtenrun 5038 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Mar 2021 · 1 entry
- Thomas Munro opened #1 with a proposal to retain tombstones only at low WAL levels, aiming to avoid expensive checkpoints during DROP TABLESPACE.
Jun 2021 · 1 entry
- Heikki Linnakangas suggested tracking dropped identifiers in shared memory in #2, while Robert Haas questioned retaining a rarely exercised tombstone mechanism in #3 and Andres Freund proposed borrowing fork-number space for 56-bit identifiers in #4.
Sep - Oct 2021 · 1 entry
- Thomas identified stale bgwriter descriptors in #6, supplied fault-injection patches demonstrating lost updates in #8, and withdrew the original conditional-tombstone proposal in #10.
Jan - Mar 2022 · 5 entries
- Robert revived the non-recycling identifier approach in #11, explaining its usefulness for WAL-logged database moves, and Dilip Kumar took over implementation in #17.
- Dilip's first implementation in #18 introduced a separate 56-bit allocator and immediate unlinking; #19 completed more interfaces, and #25 split the work into independently reviewable patches.
- Robert argued that fork packing belonged only in BufferTag rather than general RelFileNode structures in #26 and #28, leading to the representation redesign in #30.
- Dilip identified allocation durability and upgrade-counter hazards in #30, then added WAL flushing and counter advancement in #31 using solutions discussed with Robert.
- Robert's extensive review in #35 requested clearer terminology, proper extension versioning, and non-wrapping allocation; after a gap, Amul Sul reviewed bounds and batching in #36, and Thomas reviewed recovery safety in #38.
May - Oct 2022 · 24 entries
- Dilip isolated the terminology refactor in #40, preserving SQL-visible names, added the RelFileNumber type in #42, and reunited the rename with the substantive series in #50.
- Simon Riggs proposed a mapped database/tablespace identifier in #46, but discussion of complexity and identifier exhaustion led him to accept proceeding with the existing 56-bit approach in #55.
- Robert's review in #57 led to corrected packing, output conversion, and allocation batching in #59; Andres then raised fsync costs and requested safety assertions in #63, which Dilip incorporated in #66.
- Robert committed the rename component of #69 as b0a55e4329, with whitespace, comment, and invalid-value cleanup described in #70.
- The remaining series evolved through #72 and #74 to use explicit two-word packing, integrated assertions, and deferred allocator WAL flushing, while immediate tombstone removal was held back.
- Andres rejected relying on larger atomic sector writes for widened relmaps in #76 and proposed temporary-file replacement in #78; Robert supplied #79 and revised it to use durable_rename() in #81.
- Robert declined variable-length relmaps in #81 because the additional complexity and bug risk offered little benefit over increasing a fixed mapping limit.
- Following Robert's suggestion in #75, Dilip supplied the macro-to-inline preparation in #82 and incorporated it together with Robert's relmap patch into #89.
- Testing identifiers above 32 bits exposed filename, extraction, and upgrade problems in #90; #91 addressed these and preserved initial system-catalog filenames to prevent upgrade collisions.
- Thomas reported multi-platform failures in #92, leading Dilip to fix an uninitialized locator used by existence assertions in #100.
- Robert committed the relmap component of #100 as d8cd0c6c95, adopting Andres's RelationMapReplace terminology; Dilip had also reviewed Robert's original patch in #88.
- Vignesh's naming review in #93 shaped the inline-function component of #110, which Robert committed as 3ac88fddd9 after correcting BufferTagsEqual in #112.
- Dilip split missed type conversions from the remaining work in #113, and Robert reported partially committing those corrections in #114; subsequent review addressed translated formatting, extension interfaces, and documentation.
- Alvaro Herrera, Tom Lane, and Thomas examined locator return-by-value and generated calling behavior in #123, #127, and #130, resolving concerns about the new accessors.
- Robert committed the unchanged BufferTag preparation component of #136 as 82ac34db20, announcing in #152 that it could proceed independently of unresolved upgrade design.
- Robert proposed rewriting new-cluster catalogs to fully decouple filenames from OIDs in #129, and Amit Kapila proposed rewriting only conflicts in #142, prompting Dilip's experiments reported in #153.
- The upgrade experiments encountered binary-upgrade restrictions and database recreation problems; Dilip concluded in #162 that reproducing dump and ACL logic was error-prone, and Robert chose to retain initial system-table filename equality with OIDs in #163.
- After resubmitting the remaining widening patch in #159, Dilip identified a risk of allocating unflushed identifiers in #164 and separated next, logged, and flushed counters in #165.
- Further review by Amul and Robert in #166 and #167 led to unsigned identifiers, additional assertions and comments, and a separate tablespace-key simplification patch in #169.
- Robert supplied declaration relocation in #171, which Dilip reviewed and incorporated into #172; it landed as 2f47715cc8, and the widening component landed as 05d4cbf9b6 with light edits.
- In #173, Robert reported an additional headerscheck fix and left tablespace-key simplification uncommitted because its is_shared argument complicated callers without a clear benefit.
- Thomas proposed further tombstone cleanup in #174 and questioned segment-deletion sequencing in #176, while Dilip explicitly confirmed in #175 that widening had landed but tombstone removal remained pending.
- Maxim Orlov reported a mismatch between 19 control-data attributes and an 18-slot tuple descriptor in #177, and Robert acknowledged the mistake in #178.
- Tom requested missing control-data test coverage in #179; Michael Paquier supplied invocation tests in #180 and expanded them into sanity checks in #182 following Tom's review in #181.
Jan 2023 · 1 entry
- After the final rebase request in #183, Dilip withdrew the ongoing work because of extra WAL overhead in #184; this thread records no subsequent landing of the remaining tombstone, tablespace-key, or control-data test patches and no explicit revert announcement.