model-writtenrun 7234 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Apr - Jul 2023 ยท 13 entries
- Evgeny Morozov reported recurring PostgreSQL 15.2 crashes and zeroed pg_class_oid_index pages after replacing faulty disks and restoring databases, initially leaving hardware corruption as a plausible explanation #1 #2.
- Checksum discussion established that catalog indexes are protected like other relations, while all-zero pages might avoid checksum warnings, so enabled PostgreSQL and ZFS checksums did not resolve the diagnosis #6 #10 #12.
- A recurrence on newly created test databases without reported disk errors exposed an important pattern: the client had timed out and retried DROP DATABASE on both affected databases #14.
- Thomas Munro examined the relation map and WAL evidence, while Evgeny identified hundreds of zero-filled files confined to the two affected databases, shifting attention toward PostgreSQL's WAL_LOG database-copy path #24 #25 #27 #28.
- Michael Paquier reproduced catalog failures with concurrent database creation, table changes, checkpoints, and crashes, but Thomas distinguished the previously fixed source-database buffer problem from the reported zeroed target databases #35 #36 #37.
- Alvaro Herrera proposed restoring FILE_COPY as the default, but Tom Lane considered that premature without stronger evidence that the copy strategy caused the problem #39 #41.
- Andres Freund identified interrupted DROP DATABASE as the likely cause because DropDatabaseBuffers() discards dirty pages before file removal finishes, and Evgeny confirmed that client timeouts triggered cancellation #42 #43.
- Thomas reproduced the exact failure by injecting cancellation after the storage-manager barrier, and Michael independently reproduced zeroed catalog files on the development branch #47 #48.
- Andres considered committing catalog deletion before irreversible operations and splitting buffer invalidation, then favored retaining a database entry marked invalid so interrupted drops could be retried without orphaning their files #44 #49.
- Andres posted a prototype using datconnlimit = -2 as a backpatchable invalid marker, blocking connections and ALTER while permitting DROP, and identified unfinished tests, diagnostics, documentation, and a planned move of further development to another thread #50.
- For an interim workaround, Andres advised avoiding DROP DATABASE cancellation and optionally renaming databases before dropping them; Evgeny's questions about the long-running drops and differing PANIC behavior remained unanswered in this thread #58 #59.
- Development continued in thread #47599, where the later submission #538445 supplied the interrupted-DROP patch committed as c66a7d75e6, superseding this thread's prototype #50 with WAL flushing, invalid-template rejection, vacuum and client-tool handling, documentation, and broader tests; the proposed new catalog column and CREATE DATABASE lifecycle changes remained outside that commit #49.
- The thread ended with Andres confirming that the fixes had been pushed to all branches, completing the interrupted-DROP repair developed further in thread #47599 #61 c66a7d75e6.