model-writtenrun 2872 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Thomas Krennwallner proposed collecting every invalid database into a report so users would not repeatedly encounter one connection failure per pg_upgrade attempt #1.
- Nathan Bossart suggested skipping invalid databases, while Tom Lane raised possible loss of recoverable data and argued that consistency with other checks meant reporting and failing #2 #3 #5.
- Thomas explained that customers had encountered invalid databases left behind by apparently interrupted cleanup of Prisma Migrate shadow databases; Daniel Gustafsson supported reporting them all while warning that pg_upgrade could not distinguish disposable databases from valuable ones #6 #8.
- Bruce Momjian favored failing on unusual conditions unless cleanup was explicitly enabled, Nathan proposed an opt-in flag, and Daniel requested a report after successful skipping #9 #10 #11.
- Thomas implemented --skip-invalid-databases with a completion report while retaining failure by default; skipped databases would remain in the old cluster rather than being explicitly dropped #12.
- Bruce questioned demand for this specialized skip option, and Daniel supplied a smaller replacement that moved connection checks earlier and reported invalid databases through the existing check; broader skipping or repair behavior was deferred for separate design discussion #13 #14.
- Bruce agreed with the narrower approach #15, and Daniel committed #14 with only a comment clarification as f638aafd1e.
- Daniel closed the thread by confirming that reporting had landed and directing any further discussion of skipping or repairing problems to a new thread #16.