model-writtenrun 2445 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Dec 2023 · 1 entry
- Jelte Fennema-Nio proposed backtrace_functions='*' to obtain backtraces for unexpected development errors without repeatedly identifying function names, while noting that the initial implementation also traced ordinary LOG messages #1.
Feb - Jun 2024 · 15 entries
- Daniel Gustafsson supplied a revision restricting wildcard matches to ERROR and higher while preserving existing named-function behavior and clarifying its documentation; Jelte instead preferred a configurable threshold shared by both forms #2 #3.
- Jelte introduced backtrace_functions_min_level with an ERROR default, then corrected inclusive-threshold wording and added a dedicated severity-options array following Daniel's review #5 #6 #7.
- Alvaro Herrera objected to accepting ineffective partial wildcards and redundant mixtures of '*' with function names, while Bharath Rupireddy supplied and tested a competing parser rewrite using SplitIdentifierString() #9 #12.
- Jelte adopted Alvaro's standalone-wildcard restriction and left Bharath's larger parser rewrite out for separate discussion; subsequent revisions addressed documentation feedback and restored GUC_NOT_IN_SAMPLE after a CI failure #14 #16 #19 #22.
- Peter Eisentraut questioned the relationship with the recently added backtrace_on_internal_error GUC and proposed a unified mode setting, prompting Jelte to design one mode control with additional severity and function-name filters #18 #23 #25.
- Jelte implemented that broader design with log_backtrace, backtrace_min_level, and an empty backtrace_functions value meaning no function filter, replacing the explicit wildcard and adding a separate PGErrorVerbosity cleanup #26.
- Jelte added an optional patch enabling internal-error backtraces by default; Bharath reviewed the interactions and reported increased test runtime and log volume, while Jelte explained why the empty function list deliberately matched everything #27 #28 #29.
- Concern about releasing a narrowly named GUC made the issue a PostgreSQL 17 open item, and Michael Paquier proposed a smaller two-value backtrace_mode conversion while deferring broader functionality to PostgreSQL 18 #30 #31 #32.
- Peter argued that the future design lacked consensus, and Robert Haas rejected replacing a boolean with an enum without adding functionality, also suggesting error-code filtering as another possible direction #37 #39 #40.
- Tom Lane warned that renaming a released GUC could break stored settings and upgrades, initially recommending a revert while also discussing unconditional internal-error backtraces as a way to avoid prematurely fixing a configuration interface #41 #43.
- Peter found 85 backtraces in a standard test run with internal-error tracing enabled, and Jelte argued that enabling tracing without an off switch was too large a post-freeze change, especially for extensions that use elog freely #46 #47.
- Michael analyzed the test failures, separated incorrect internal-error codes from legitimate test cases, and moved toward reverting; Andres Freund's security and performance objections to unconditional backtraces reinforced the need for a configurable design #50 #52.
- Robert, Tom, and Michael explicitly supported reverting, and Peter carried out that discussion decision in 592a228372, removing backtrace_on_internal_error for PostgreSQL 17 without applying an attached patch #51 #53 #55 #56.
- Robert cautioned that the postponed work still needed agreement on its design, rather than merely another CommitFest entry #59.
- Jelte resumed the broader proposal in June 2024, rebasing after the revert, dropping the default-on patch, and adopting internal_error terminology; he defended intersecting filters despite requiring multiple settings to reproduce the old named-function behavior #60.
Feb - Mar 2025 · 3 entries
- After another long gap, Jelte rebased the two-patch series in February 2025, and Daniel committed only its unchanged PGErrorVerbosity cleanup as 95ef3d9029, while questioning the three interdependent controls in the remaining feature #61 #62.
- Peter questioned manually registering unused typedefs because buildfarm collection would not discover them, and Tom explained that manual curation could preserve such entries but agreed that unused typedefs might be unnecessary #63 #64.
- At the thread's end, Jelte resent only the uncommitted backtrace feature, rebased and with its unused LogBacktraceVerbosity typedef removed; the cleanup had landed, but the broader controls remained under review with the GUC-interaction concern unresolved #65 #62.