model-writtenrun 4856 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Oct - Nov 2022 · 1 entry
- Nikita Glukhov proposed error-safe calls for SQL/JSON without subtransactions, but Tom Lane rejected capability flags and a new calling convention and requested a smaller documented interface with representative input functions #1 #3.
Dec 2022 · 34 entries
- Andrew Dunstan supplied a reduced implementation using IOCallContext, bool/int4 conversions, and COPY NULL_ON_ERROR, after which Tom insisted on preserving complete ErrorData and SQLSTATE through centralized error machinery #9 #10.
- Tom retained FunctionCallInfo.context rather than adding fields to every call frame, then replaced the initial implementation with ErrorReturnContext infrastructure and experimental COPY-based testing #22 #25 #26.
- Robert Haas objected that ereturn misleadingly suggested a return statement, leading Tom to introduce errsave and reserve ereturn for an actually returning wrapper while adding array and record input handling #41 #53.
- Andres Freund's review exposed a leaked record TupleDesc and drove boolean success returns, explicit resource cleanup, shared elog.c helpers, and SQL validation functions replacing the experimental COPY patch #60 #68.
- Discussion about documenting the validation functions led to subdivisions in the system-information documentation, committed without an attached patch as 3b31821953, after David G. Johnston accepted the approach #74 #81.
- The preliminary elog.c refactoring from #68 landed separately as 8305629afe, leaving a reorganized remainder that adopted David's soft-error terminology #89 #90.
- Further review settled on struct Node forward declarations, while the final core series added textual type specifications, the remaining numeric types, and contrib/cube to exercise the interface beyond simple input functions #98 #99 #100.
- The infrastructure portion of #100 landed with a small typedef-list addition as d9f7f5d32f, and its validation functions landed with constant-type-name caching as 1939d26282.
- The datatype portions of #100 were combined into ccff2d20ed, with revised record tests, and Tom announced that the foundation was available for further conversions #105.
- Tom's subsequent datetime work first centralized timezone diagnostics through DateTimeErrorExtra in 2661469d86, an unattached prerequisite to the conversion he had undertaken #103.
- Datetime input functions then gained soft-error handling in c60488b474, covering date, time, timetz, timestamp, timestamptz, and interval without a matching attachment #109.
- Andres suggested replacing formatting.c's separate error mechanism, and Tom implemented that conversion in 4dd687502d after agreeing to the proposal #102 #103.
- Andrew's initial JSON conversion left JSONB problematic, so Tom demonstrated numeric overflow and expanded it with status-returning JsonSemAction callbacks and safe numeric input calls #112 #115.
- The JSON work split into the semantic-action API change 50428a301d and JSON/JSONB conversion c60c9badba, both derived from #115, with Unicode-conversion error handling added after discussion #118 #119.
- JSONB value-construction size-limit errors remained hard because changing pushJsonbValue and its more than 150 callers was judged disproportionate to the immediate benefit #120.
- For domains, Tom and Alvaro Herrera accepted that arbitrary CHECK expressions could still throw hard errors; the resulting domain conversion landed unchanged from #116 as b8c0ffbd2c #110.
- Amul Sul supplied fourteen additional conversions and expanded them to twenty-five, after which Tom rejected changes to five unimplemented input functions and corrected cases that continued processing after a helper reported a soft error #125 #126 #130.
- The bit and network portions of #126 landed with revised MAC parsing as 17407a8eaa, while the geometric portion landed with helper refactoring as 332741e739, leaving some line-construction arithmetic errors hard #130.
- The internal-type portion of #126 landed with additional ACL and OID error handling as 47f3f97fcd, and its cash/UUID portion landed with light edits as 90161dad4d.
- Tom separately converted bpchar, varchar, and bytea in [commit:3b9d2deb67a0d5ee681fc0e07c2dca4e26e96], covering string functions explicitly omitted from Amul's attached series #126.
- Tom extended range conversion into canonicalization functions and changed the exposed range helper signatures; after Amul reviewed the hard-error boundary, both patches from #131 landed together unchanged as d35a1af468 #132 #134.
- XML conversion distinguished malformed input from libxml initialization failures, landing from #136 as 37bef842f5 with additional test output, while its companion cleanup landed as c4939f1215 to stop treating arbitrary exceptions as malformed XML.
- Andrew's jsonpath conversion required changes in both scanner and parser, and Tom's review identified cached regex compilation as a remaining hard-error path #139 #147.
- Andrew also converted contrib/seg, and Tom's diagnostic-name and comment suggestions were incorporated before #145 landed as 878ce16056 #146 #148.
- The revised jsonpath patch used pg_regcomp, but Tom then reproduced a compiled-regex leak and supplied a pg_regfree delta; the combined work landed as e37fe1db6e #149 #159 #161.
- Robert Haas urged broader reg* coverage, but Tom's proposed conversion deliberately left main-grammar, typmod, and some catalog lookup failures hard because fixing them would require substantially wider changes #140 #162.
- The previously overlooked enum input function was converted directly in 442e25d248, whose explanation notes that the initial survey database contained no enumerated types.
- Tom addressed xid, xid8, and cid silently accepting garbage with #163, which landed with simplified validation and a corrected cid diagnostic as eb8312a22a, also extending OID input to hexadecimal and octal forms.
- The full-text conversion carried error contexts through existing parser states and suppressed empty-query NOTICEs in soft-error mode; #165 landed unchanged as 78212f2101.
- The reg* conversion from #162 landed as 858e776c84 with a local adjustment preserving hard errors for invalid validation-function type specifications.
- A separate delta replaced to_reg* implementations with safe input wrappers and added numeric OID acceptance; Andrew endorsed it, and #167 landed unchanged as 3ea7329c9a #169.
- Andrew submitted the remaining ltree/intarray work, Tom requested simpler return interfaces, and Amul supplied a revision carrying escontext through parser state #168 #170 #173.
- Tom handled hstore without an attachment, reporting completion in #172; 197f98a848 also modernized its diagnostics and added missing error tests.
- The final combined submission #173 was split and further simplified into ltree commit 7a05425d96 and intarray commit 3b76622e04, after which Andrew considered the conversion effort complete #174.
Jan 2023 · 1 entry
- The thread ends with the submitted conversions landed, but Robert still objects to incomplete reg* coverage and asks for user documentation distinguishing soft input errors, hard resource failures, and extension-dependent support; the original COPY behavior remains outside these commits #175 #68.