model-writtenrun 3173 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Michael Paquier proposed replacing internal-error codes on user-reachable paths with specific SQLSTATEs, reporting that his patch reduced test-generated internal errors from 85 to 35 while preserving intentional corruption and test failures #1.
- Justin Pryzby pointed to an earlier inventory of user-facing elog calls, and Michael agreed to examine it while deferring the cleanup until PostgreSQL 18 development opened #2 #3.
- Robert Haas approved the proposed cleanup and marked it ready for committer; Michael reiterated the timing decision #4 #5.
- Michael committed the cleanup from #1 as b81a71aa05, dropping an obsolete tablecmds.c hunk and resolving an OpenSSL context conflict during rebasing #6.
- Alexander Lakhin then raised additional user-triggerable internal errors involving satisfies_hash_partition() and BRIN bloom indexes; Michael favored case-by-case treatment, accepting the undocumented function's behavior but finding the ordinary DDL failure more concerning #7 #8.
- Alexander supplied further examples involving object identifiers, snapshot filenames, hash operators, recursive queries and ltree; Michael identified the missing ENAMETOOLONG mapping as actionable, considered the recursive-query case a likely bug, and recommended separate threads for broader investigations #9 #10.
- Daniel Gustafsson supported mapping ENAMETOOLONG to ERRCODE_NAME_TOO_LONG, while Alexander reported the recursive-query bug separately and questioned whether internal-error codes should consistently identify unexpected server conditions #11 #12.
- Michael noted that the recursive-query bug had been fixed elsewhere and reiterated that some tests intentionally provoke internal errors, leaving the broader classification policy unresolved #13.
- Michael reconsidered reusing ERRCODE_NAME_TOO_LONG because it describes database identifiers, and instead submitted a dedicated class-58 file_name_too_long code with its errno mapping; Daniel endorsed this distinction #14 #15.
- After proposing a development-branch-only change, Michael committed #14 unchanged as 2d8ef5e24f and reported checking buildfarm support for ENAMETOOLONG; both attached patches had landed by the thread's end, while the broader error-classification questions remained outside these fixes #16 #17.