model-writtenrun 7300 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Aug - Sep 2022 · 26 entries
- Jonathan S. Katz opened the thread in August 2022 to decide whether PostgreSQL 15 could ship its already-committed SQL/JSON features; the release management team preferred reverting, while he emphasized their standards and usability value #1.
- Andrew Dunstan began preparing a possible revert, while Jonathan asked that any rescue address design documentation, coercion complexity and the outstanding implementation concerns #3 #4.
- Andres Freund found that subtransaction-selection logic was misleading, routine coercions still started subtransactions, and error recovery raised memory-management concerns; his benchmark showed a substantial cost compared with ERROR ON ERROR #5.
- Nikita Glukhov submitted a domain-return fix and an executor redesign using EEOP_SUBTRANS, explaining that coercion failures and DEFAULT ON EMPTY failures had to reach ON ERROR handling #6.
- Andres questioned the memory cost of subsidiary ExprStates and the risks of changing recursive execution and JIT late in the release cycle; Nikita then incorporated Andres's cleanup, introduced JsonCoercionType and added recursive JIT support #8 #11.
- Amit Langote reviewed missing-coercion representations, redundant parser checks and executor comments, while Michael Paquier warned that the size and complexity of the proposed executor changes were unsuitable after three betas #16 #17 #18.
- Robert Haas argued that expression evaluation should use explicit error-return interfaces instead of subtransactions, and Tom Lane warned that shipping the implementation would constrain future repairs through compatibility obligations #20 #21.
- Andrew proposed salvaging a reduced feature set by disabling DEFAULT behaviors and some return types; Pavel Stehule and Nikita supported safe conversion routines for a limited set of JSON-compatible types #23 #25 #27.
- The discussion clarified that arbitrary and composite return types required broader coercion support, and Tom distinguished a short-term core-type solution from the extensible API that would need a later release #29 #38 #44.
- Nikita abandoned the subtransaction architecture and produced a work-in-progress patch using hardcoded safe conversions; the next revision completed missing conversions, restored mutability checks and restricted DEFAULT ON EMPTY expressions #47 #51 #55.
- Jonathan reviewed the new tests and documentation implications, while Amit questioned constant folding during parse analysis and requested restoration of PASSING-expression initialization in the parent ExprState #57 #58.
- Nikita split the work into input-function and executor patches, restored subsidiary-ExprState elimination, and disabled DEFAULT ON EMPTY; Amit then reviewed the split series and Álvaro Herrera proposed shared internal functions to avoid additional input-function overhead #61 #62 #63.
- The final SQL/JSON rescue revision adopted the internal-function approach, extracted and renamed coercion helpers, and removed DEFAULT ON ERROR despite Nikita's reservations; Amit and Andrew subsequently preferred retaining the syntax and reporting unsupported behavior during parse analysis #67 #70 #71.
- Amit's final detailed review still found missing inline markers, stale comments, questionable test output and a compiler warning, and Andrew raised API concerns about datetime decoding and jsonb input helpers #68 #74.
- Andres, Tom and Robert concluded that there was insufficient time to review the remaining redesign safely, and Jonathan formally requested the revert on behalf of the release management team #73 #75 #76 #78.
- Tom proposed a general API that passes an error-reporting context through FunctionCallInfo.context, allowing incremental adoption by built-in and extension types and future COPY error handling; Robert supported that direction #79 #80 #81.
- Nikita supplied a separate proof of concept with FuncCallError, PG_RETURN_ERROR, a pg_proc capability flag and COPY NULL_ON_ERROR, then described an unindexed archive extending the API to expression nodes and lazy JSON_VALUE coercions, with JIT and several conversion paths still unfinished #88 #89.
- Andrew reverted the SQL/JSON feature family on the development branch and PostgreSQL 15 in 2f2b18bd3f, implementing the thread's decision rather than an attached repair patch, and confirmed completion in #90.
- The remaining compiler-warning discussion identified Amit's unexpectedly old GCC, and the thread closed with Andrew requesting a separate input-function submission and Tom rejecting the proposed pg_proc flag because it would complicate incremental conversion #92 #93 #94 #95.
- At the end of this thread, the PostgreSQL 15 rescue had been abandoned and the soft-error work remained unfinished; the later candidate commits cite this discussion as ancestry, but the supplied cross-thread verdicts identify their matching submissions elsewhere #90 #94 #95.
- Standard JSON constructors returned in March 2023 through 7081ac46ac and IS JSON through 6ee30209a6, with their source components identified in #502994 of thread #38009, replacing the previous discussion-only attribution here.
- JSON(), JSON_SCALAR() and JSON_SERIALIZE() followed in July 2023 through 03734a7fed, sourced from #539031 in thread #48026 with light final cleanup.
- In January 2024, faa2b953ba introduced jsonpath variable callbacks, aaaf9449ec added soft-error expression steps, and 1edb3b491b extended soft errors through record population and domain checks, all from components of #551965 in thread #48026.
- SQL/JSON query functions returned in March 2024 through 6185c9737c, sourced from #557196 in thread #48026 with local executor changes and using the general soft-error architecture rather than the hardcoded conversions and removed DEFAULT behaviors of #67.
- Basic JSON_TABLE support landed in April 2024 through de3600452b from #558821 in thread #48026, deferring NESTED and PLAN support; bb766cde63 then added nested paths from #559360 in that thread with parent-child OUTER and sibling UNION semantics.
- The supplied later history concludes with 86ab7f4c72 adding JSON_TABLE PLAN clauses in July 2026 from the exact matching submission #1147204 in thread #50777, superseding the older PLAN work in thread #48026 and the previous discussion-only attribution here.