model-writtenrun 9681 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Dec 2018 - Feb 2019 ยท 23 entries
- Andrew Gierth proposed replacing expensive floating-point text conversion with an adaptation of Ryu, retaining printf-like formatting and making shortest precise output the default whenever extra_float_digits is positive #1.
- Andres Freund's initial review questioned leftover code, generated-table references, and naming; the early revisions cleaned these up and repaired Windows build integration and type usage #2 #8 #10 #12.
- Thomas Munro argued against unnecessary changes to imported code, while Tom Lane questioned its long-term maintenance cost; Andres Freund emphasized the practical conversion bottleneck, and Gierth supplied benchmarks showing substantial gains #9 #14 #15 #21.
- After a month, Gierth rebased the patch and asked for decisions on exact output by default, compatibility with extra_float_digits, float-to-numeric casts, and permission to retain upstream declaration placement #22.
- Robert Haas emphasized preserving stored values across dump and restore, while Andres Freund supported allowing textual representations to change when they preserve the same floating-point values and proposed suppressing declaration warnings only for the imported files #28 #29 #31 #32.
- Testing prompted by portability concerns exposed double rounding in float4in's use of strtod; Gierth proposed strtof, and Tom Lane requested a configure-tested fallback for older systems lacking it #40 #41.
- Donald Dong posted a standalone float4in test program, but its frontend executable could not link backend functions; Gierth recommended SQL tests or a loadable module, and this harness was not incorporated into the committed changes #44 #45.
- Gierth then combined the separately developed strtof prerequisite with Ryu, renamed the public functions by their purpose, restored declaration-after-statement style, and adjusted regression tests while preserving rounded output for numerically unstable cases #46.
- Windows rounding and subnormal failures led to another revision that avoided exact-midpoint decimal outputs, strengthened strtof compatibility handling, and updated contrib expectations #48 #49.
- Dong tested the changed precision behavior and requested documentation; Gierth explained that every positive extra_float_digits value intentionally requests shortest precise output rather than a fixed number of additional digits #50 #51 #52.
- The documentation revision established the new default and IEEE-format assumptions, while float-to-numeric conversion changes were deferred because immutable casts could appear in indexes preserved by pg_upgrade #54.
- Gierth incorporated an upstream small-integer fast path and optimized fixed-point formatting, reporting that conversion was now faster than bigint output in his tests #55.
- Tom Lane's HP-UX test exposed an unavailable HUGE_VALF constant, which Gierth guarded; the final attached revision also adopted Noah's configure approach, documented buffer sizes and entry points, and simplified formatting code #56 #57 #58.
- Lane subsequently confirmed that the portability fix compiled and passed core regression tests on his HPPA system #59.
- The input prerequisite landed as f397e08599, attributable to the newer strtof-specific submission #417062 in thread #39983; the Ryu output portion of #58 landed separately as 02ddd49932, including a correction to its int128 feature guard.
- After the landing, Gierth realized that changed dump text broke cross-version upgrade comparisons and proposed another GUC to restore historical formatting #60.
- Andrew Dunstan demonstrated an alternative by temporarily letting pg_dump obtain extra_float_digits from an environment variable; setting it to zero made cross-version tests pass back to PostgreSQL 9.2 without another server GUC #61.
- Dunstan also investigated Windows regression failures and confirmed that the machines matched the ordinary float8 expected output, information Gierth said enabled a straightforward float4 fix #62 #64 #65.
- Gierth rejected an implicit environment dependency and preferred an explicit --extra-float-digits option; Lane supported that interface, and Dunstan confirmed that the buildfarm could invoke it selectively for upgrades to PostgreSQL 12 #67 #68 #70 #72.
- A separate post-commit portability discussion concerned Ryu's use of boolean values as integers and compiler warnings when PostgreSQL's fallback bool was plain char; Lane proposed changing that fallback to unsigned char #73 #74 #75.
- Dunstan posted the pg_dump and pg_dumpall option patch, after which Lane requested documentation and Gierth supplied wording explaining that routine backups should retain maximum precision #76 #77 #81.
- Dunstan committed that option as af25bc03e1, retaining the implementation from #76 and adding the agreed documentation and minor comment edits, then reported completion #83.
- At the thread's close, Gierth reported testing and pushing the unsigned-char bool fallback, with the locust buildfarm member responding successfully; the main output change and dump option had landed, while float-to-numeric conversion remained deferred #84 #54.