model-writtenrun 7031 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Feb 2024 · 1 entry
- Ashutosh Bapat proposed a standalone regression-database dump/restore test in #1 to cover normal dumps that pg_upgrade's binary-upgrade tests missed, while acknowledging that comparing dumps cannot detect objects omitted by both dumps.
Apr - Jul 2024 · 3 entries
- Michael Paquier and Tom Lane objected to another full regression run, while Daniel Gustafsson and Peter Eisentraut favored reusing pg_upgrade's existing database; Ashutosh adopted that placement in #12 following #2, #5, #8, and #9.
- Michael requested a separate normalization module instead of hardcoded expected diff output in #13, leading Ashutosh to introduce AdjustDump.pm, run the roundtrip before upgrade, and add all-format testing in #16.
- Runtime concerns initially produced a plain-format default with other formats behind PG_TEST_EXTRA in #20, followed by a schema-name adjustment for changed regression objects in #21.
Oct 2024 - Apr 2025 · 18 entries
- Michael's timing measurements and Tom's objection to default runtime overhead in #22 and #23 led to the fully opt-in test in #26, which also added module installation, normalization assertions, and a reusable dump-comparison helper.
- The new test exposed inherited NOT NULL constraint differences in #26; the correction was developed in thread #50568 from #578069 and landed as fd41ba93e4, after which Ashutosh reported that this thread's test passed again in #27.
- Daniel reviewed diagnostics, format selection, and cross-version behavior in #28 and #30; Ashutosh incorporated the feedback through #32, including checks for different server versions and custom installations.
- After the January rebase in #33, Michael split the helper from the roundtrip test and generalized it into compare_files() in #35, with an optional line comparator usable by pg_combinebackup.
- The helper portion landed with small cleanup as 169208092f, unifying comparison diagnostics across pg_upgrade, pg_combinebackup, and recovery tests while leaving three specialized diff users untouched.
- Ashutosh continued with the uncommitted roundtrip portion in #40, adding generated-column normalization and filtering problematic inherited-table COPY sections so comparisons could include data rather than only schemas.
- Statistics mismatches prompted temporary --no-statistics support in #43, and a subsequent materialized-view problem kept that workaround necessary in #44, which also disabled destination autovacuum.
- Alvaro Herrera found a money-format mismatch in #45 and traced it to locale settings in #49; Ashutosh first proposed explicit lc_monetary initialization in #51, then replaced it with --create during dump and restore in #52 to preserve database settings generally.
- Vignesh suggested parallel execution in #53, and Alvaro argued for one format enabled by default in #54; timing experiments led to directory-format dump and restore with two workers in #63.
- Alvaro's full-suite measurements in #65 supported default enablement, while Michael remained concerned about the slowest test and schedule maintenance in #70.
- Attempts to shorten the regression schedule in #76 and #77 met Tom's coverage objection and proposal for dedicated object-creation tests in #78; Daniel recommended retaining the existing schedule for this release in #82.
- Ashutosh removed statistics exclusion after the materialized-view fix in #81, but restored it in #84 after another mismatch appeared.
- Alvaro combined the final series into 172259afb5, enabling the directory-format roundtrip by default while excluding statistics, replacing the test subroutine with a SKIP block, and simplifying assertions and documentation; he reported successful CI in #85.
- The remaining statistics mismatch proved to be a test-ordering problem rather than another pg_dump defect: citing Jeff Davis's diagnosis, Ashutosh submitted #92 to run the roundtrip after source autovacuum shutdown and preserve separate old/new initialization parameters.
- Alvaro split #92 into the parameter cleanup abe56227b2 and the test relocation plus restored statistics comparison 8806e4e8de, confirming both in #93; this later implementation superseded the statistics-only proposal #588164 in thread #51340.
- Andres Freund reported Valgrind timeouts and suggested reducing statement logging in #95, prompting Alvaro's three-cluster logging patch in #96.
- Ashutosh proposed reusing the upgraded cluster for restoration and bundled an adapted logging patch in #97, reporting additional correctness checks in #98; the cluster-reuse optimization did not land among the supplied commits.
- The original three-cluster logging change landed with revised comments as 64fba9c617, while Andres's measurements in #99 and #100 showed that overall test-cycle overhead remained substantial.
Aug 2025 · 3 entries
- In August, Alvaro reconsidered default enablement and submitted the opt-in guard in #101; Daniel and Tom endorsed it in #102 and #103.
- The guard landed as d185161e47 with the option renamed regress_dump_restore and documentation added, making the roundtrip test opt-in on the development and PostgreSQL 18 branches.
- By the thread's end, Michael had enabled regress_dump_restore on three buildfarm animals in #105, and Ashutosh acknowledged the added documentation in #106; the test and stability fixes were committed, while cluster reuse and broader runtime redesign remained uncommitted.