model-writtenrun 9141 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Tom Lane opened the thread because nondeterministic collations let texteq()/textne() invoke the same comparison infrastructure as functions marked non-leakproof, making their different markings inconsistent; he preferred making all comparisons leakproof over removing equality's marking and harming query performance #1.
- Tom Lane examined Windows and ICU error paths and found that they did not expose input strings directly, but questioned whether input-dependent encoding errors still violated leakproofness #2 #4 #5.
- Robert Haas distinguished tolerable failures on corrupt data from unacceptable leaks involving valid strings, and Tom Lane argued that valid strings triggering comparison errors would already break basic indexing requirements; Robert Haas and Stephen Frost accepted that reasoning #6 #7 #8 #9.
- Tom Lane summarized consensus to mark the functions leakproof but initially proposed leaving PostgreSQL 12 unchanged to avoid a late catalog-version bump; Peter Eisentraut supported that plan unless another bump became necessary #10 #11.
- Tom Lane posted the sole patch, adding leakproof markings for string comparison functions, documenting the error-reporting constraint in varstr_cmp(), and updating regression expectations; the last message proposed applying it to HEAD and closing the PostgreSQL 12 item without a fix #12.
- Tom Lane subsequently committed that patch with only an added catalog-version increment as c160b8928c; the commit also records back-patching to PostgreSQL 12 because a separate post-beta4 catalog-version bump had become necessary.