model-writtenrun 9862 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Oct - Nov 2018 ยท 15 entries
- The thread began with Michael Paquier announcing TAP coverage contributed by Michael Banck, Magnus Hagander, and himself, covering checksum configuration, corruption detection, and single-relfilenode scans #1.
- Buildfarm failures exposed attempts to checksum EXEC_BACKEND configuration files; Andres Freund identified the relevant build configuration, and Michael Paquier supplied a small blacklist extension #2 #3 #4.
- Andres Freund argued that temporary files and extension-owned files made the blacklist inadequate, prompting Michael Paquier to replace it with relation-filename matching and accompanying tests #5 #7 #9.
- Michael Banck requested tests proving that correctly named files containing garbage were actually scanned, and Michael Paquier added those tests while switching fork recognition to the shared helper so init forks were included #10 #18.
- Andrew Dunstan found redundant checks and acceptance of filenames ending in a bare dot; Michael Paquier corrected the parser, renamed its predicate to isRelFileName(), and expanded malformed-name coverage before Andrew accepted the revision #19 #20 #21.
- Tom Lane requested that the fix wait until PostgreSQL 11.0 had wrapped, and Michael Paquier subsequently committed the final attachment with comment edits as d55241af70, including a PostgreSQL 11 backpatch #23 #29.
- After the commit, Stephen Frost requested replacing the whitelist with coordinated blacklist fixes because pg_basebackup used different checksum-selection logic and silently skipping unknown files could miss data, while Michael Banck and Tom Lane supported centralizing the shared logic #30 #31 #32.
- Andres Freund defended explicit recognition of checksummed files and identified additional temporary-file exclusions missing from pg_verify_checksums, explaining that temporary files could survive crash restart followed by clean shutdown #36 #44 #46.
- Stephen Frost continued to advocate reverting alongside an alternative fix, but Andres Freund, Andrew Dunstan, and Tom Lane opposed a standalone revert that would restore buildfarm failures; no revert is recorded in this thread #51 #50 #52 #53.
- Michael Paquier proposed sharing relation-file recognition through src/common and fixing base-backup checks, while Andres Freund wanted an API that also handled temporary exclusions and Stephen Frost requested consideration of external backup tools; the broader design discussion moved toward a separate thread without resolving the whitelist disagreement here #54 #55 #59 #69.
- A month later, Michael Banck discovered that corruption tests could repeatedly fail on an earlier leftover file instead of their intended target, and posted a repair using distinct relfilenodes, -r filtering, and filename-specific error assertions #70.
- Michael Paquier committed that repair as 9685d7383a, retaining the stricter assertion but deleting each test file afterward instead of filtering by relfilenode, and keeping corrupted-file names separate from empty fixtures #71.
- Michael Banck then demonstrated that the whitelist skipped non-default tablespaces because their version-directory names failed relation-file recognition, attaching a regression test that induced corruption inside such a tablespace #72.
- Michael Paquier combined those tests with a minimal fix that restricted relation-name filtering to regular files and separately handled temporary paths; Michael Banck urged a prompt backpatch independent of the stalled broader refactoring, and Paquier agreed #73 #74 #75.
- The final message reports that the tablespace and temporary-file issues were fixed by follow-up commits, closing the immediate bugfix work while the broader shared-API discussion remained outside this thread #76.