model-writtenrun 7446 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Christoph Berg reported reproducible PostgreSQL 14beta1 pgbench test failures on Debian i386, while older distributions and other architectures passed #1.
- Thomas Munro reproduced the failure with GCC in 32-bit mode and traced differing permute() results to getrand(); Tom Lane identified the test's assumption of identical floating-point results across platforms as unjustified #3 #4 #5.
- Fabien COELHO proposed replacing the PRNG or extending the existing random API, while Michael Paquier proposed removing the test; Dean Rasheed supported removal because the separate large-permutation code path that originally justified it no longer existed #6 #7 #8.
- Fabien accepted disabling the test temporarily while pursuing a PRNG fix, then submitted a patch preserving the assertions as comments with a FIXME to keep the portability issue visible #9 #10.
- Dean committed 0f516d039d based on #10, removing the test and its expected-output checks entirely; he explained that the issue affected other random functions too, and any PRNG replacement considered for PostgreSQL 15 should cover all of them and would require new expected results #12.
- Fabien and Andrew Dunstan continued to favor fixing the underlying behavior, with Fabien noting that deleting the test also removed the reminder; no PRNG replacement was submitted in this thread #13 #14.
- The thread ended with the test removal committed but the underlying portability issue unresolved: Christoph's broader Debian build results isolated the original failure to i386, and Tom suspected extended floating-point precision despite compiler options intended to suppress it #15 #16.