model-writtenrun 4833 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Paul Ramsey proposed a core random_normal() function with optional distribution parameters and a shared PRNG implementation in #1, motivated by repeatedly needing normally distributed values.
- David G. Johnston requested the conventional mean-first argument order and clearer transformation code in #2, while Justin Pryzby caught an incorrect internal-function binding in #3; Paul incorporated these changes through #4 and #6, declining the suggested micro-optimization.
- Paul addressed CI portability trouble in #7, and Michael Paquier then requested declaration fixes and regression coverage while suggesting an additional random_string() function in #8.
- Mark Dilger exposed negative-standard-deviation and infinity edge cases in #10, prompting discussion with Paul in #11; Joe Conway questioned duplication of tablefunc's normal_rand() in #12, and Paul defended a convenient scalar core counterpart to random() in #13.
- Paul added tests and a separate random_string() attachment in #15, then revised portability and test details through #16 and #18 in response to Michael's review in #17.
- Fabien COELHO reported successful testing in #19, requested reuse of the new generator by pgbench, questioned epsilon-based tail truncation, and challenged random_string()'s naming, placement, and randomness source.
- Discussion broadened to the intended family of random functions in #20 and #21, with Vik Fearing clarifying in #22 that the SQL standard provided no relevant specification.
- Michael supplied a concrete CI compiler failure in #25, and Tom Lane traced it to including a backend-only header from src/common in #27.
- Tom posted a revised random_normal() patch in #28, correcting parallel restrictions and default-argument handling, replacing epsilon rejection with the pgbench-style uniform transformation, and making pgbench call the common generator; he retained sine to preserve pgbench's expected results and left the infinity/NaN behavior unchanged.
- Tom excluded random_string() from #28 because demand and semantics were insufficiently established, and deferred other distributions until someone proposed them.
- Tom's repeated testing exposed frequent failures in the existing random() tests in #29, opening a separate effort to replace the regression tests.
- Dean advocated shared seedable state for pseudorandom functions and future code and documentation consolidation in #30; Tom preferred waiting for a function that required the reorganization in #31, and Dean accepted that scope in #32.
- Tom committed Paul's feature as 38d81760c4 from #28, making only a catalog-version bump, a comment URL correction, and a small regression-test adjustment.
- After the feature landed, Tom proposed replacing random.sql in #33 with faster, more thorough checks designed for roughly one-in-a-billion failure probabilities, and suggested removing the mechanism that ignored random-test failures.
- Dean supplied a Kolmogorov-Smirnov uniformity test in #34, which Tom incorporated into #35 with three attempts to reduce false failures; testing on NetBSD/macppc also led Tom to reduce printed precision for random_normal() results.
- Tom committed the test overhaul as 09d517773f from #35, correcting a comment character and removing the schedule's ignore marker, then separately removed pg_regress's ignore machinery as bd8d453e9b based on the discussion.
- Further buildfarm rounding differences prompted 02d552c4f4 to lower extra_float_digits to -1 without an attached patch; Tom later identified AIX as the affected platform in #38.
- Dean's additional probability analysis in #36 and Paul's thanks in #39 arrived after the commits, rather than serving as pre-commit reviews.
- Andrei Lepikhov raised downstream timeout-testing uses of ignore in #40; Tom defended its removal in #41, and Andrei agreed to seek stable alternatives in #42, with no revert recorded.
- David E. Wheeler revived the thread in October 2025 after pgTAP encountered the removed directive, noting its wiki documentation and requesting formal pg_regress documentation in #43; Tom agreed that the obsolete wiki material should be replaced in #45.
- Michael highlighted undocumented test-environment variables in #46, and David's final message #47 reinforced the documentation need; the four commits remained landed, random_string() remained excluded, and documentation improvements were still discussion only.