model-writtenrun 9658 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Michael Paquier reported that backend errors in get_controlfile() could leak file descriptors and proposed replacing open() with BasicOpenFile(); he also considered a read wait event unnecessary for these SQL-facing calls #1.
- Fabien COELHO raised control-file locking and concurrent maintenance concerns, while Andres Freund argued that page-sized writes should be atomic and flock() would be disproportionate; Fabien clarified that his concern was coordination across whole maintenance operations #2 #3 #4.
- Joe Conway pointed out that BasicOpenFile() does not automatically clean up descriptors after errors, then supplied a revised patch using OpenTransientFile() and CloseTransientFile() for backend builds #5 #6.
- The coordination discussion explored a durable maintenance flag in the control file, but Andres argued that any stronger interlock should use postmaster.pid; this remained separate from the descriptor fix #7 #8 #10.
- Michael reviewed Joe's patch and requested removal of the stale argument list in its function comment; Joe and Michael agreed to fix only the development branch because practical failures appeared unlikely #9 #11 #12.
- Joe committed #6 with the requested comment cleanup as 4598a99cf2, then reported unnecessary O_RDWR flags in two slru.c readers #14.
- Michael identified another unnecessarily writable open in RestoreSlotFromDisk() and missing CloseTransientFile() error checks, moving that work to thread #40210; its later patch #420038 supplied 82a5649fb9, so that commit is attributed to the other thread #15.
- The original leak report was declared closed, while Michael and Andres continued to disagree over a control-file maintenance flag versus postmaster.pid; Fabien ultimately accepted the checksum patch's postmaster.pid approach without finding its race protection convincing, leaving that broader coordination concern unresolved here #13 #16 #17 #18.