model-writtenrun 8498 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Nick Schefner reported that running pg_upgrade --check in a separate container namespace disrupted a running cluster, and proposed treating an existing postmaster.pid conservatively or documenting the risk #1.
- Robert Haas explained that separate PID and IPC namespaces defeat the usual interlocks, but warned that avoiding server startup would sacrifice handling of stale postmaster.pid files #2.
- Tom Lane opposed skipping startup because it ensures WAL replay is complete, and warned that namespace isolation can defeat safeguards against two postmasters accessing the same data directory #3.
- Robert Haas emphasized that container setups are common and destructive behavior from --check is counterintuitive #4; Tom Lane then allowed that refusing to proceed when postmaster.pid exists might be reasonable, while stressing that this would leave the underlying namespace hazard unresolved #5.
- Daniel Gustafsson submitted a documentation patch requiring a shared process namespace as an interim warning while no operational solution had been agreed #6.
- Robert Haas recommended explaining that --check is not strictly read-only, correcting its option description, and describing safe server access and startup without focusing specifically on containers #7.
- Daniel Gustafsson incorporated that wording into the revised documentation patch, retaining the note's placement and clarifying that an already-running old cluster would not be stopped #8. The thread ends with no linked commit, the note's placement still questioned, and wording for the separate rollback section's claim that the old cluster remains unmodified still unresolved.