model-writtenrun 8819 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Jan 2020 ยท 21 entries
- Robert Haas opened the thread with #1, proposing reuse of PostgreSQL's backend JSON parser for frontend backup-manifest validation and posting prerequisite cleanups plus an explicitly incomplete frontend adaptation.
- Tom Lane and Andres Freund favored avoiding a second JSON parser in #2 and #3, while discussion of frontend error recovery led Robert to propose returning errors to callers in #8 rather than requiring backend transaction semantics.
- Tom expanded Robert's wchar.c prerequisite to include encnames.c, MSVC integration, and symlink cleanup in #14, then split it into the verbatim relocation e6afa8918c and header cleanup 3d4cb5d6c1.
- Tom proposed moving backend-only encoding routines into mbutils.c and making ICU lookup return NULL in #17, implementing that discussion directly in 5afaa2e426 without posting another patch.
- David Steele demonstrated successful frontend parsing of several JSON types, an escape, and a multibyte character with the pg_basebackup test harness in #20, while raising the unresolved question of incremental input for large manifests.
- Robert replaced the initial error-macro experiment with JsonParseErrorType return values and removed lex_accept() in #23; David examined the error cases and suggested clearer wrapper naming in #25.
- Mark Dilger independently pursued broader multibyte-header cleanup in #24, moving pg_wchar.h and introducing a backend-only mbutils.h while identifying remaining dependencies through pqformat.h.
- Robert completed the return-code interfaces and separated backend error helpers in #27, then explained the intended json, jsonb, jsonfuncs, and jsonapi boundaries to Alvaro Herrera in #29.
- The manifest discussion exposed filenames that cannot be represented directly in UTF-8: Alvaro proposed a separate base64-encoded path field in #33, Robert provisionally accepted that direction in #35, and subsequent discussion weighed encoding against skipping files, with incomplete-backup concerns raised in #57 and #62.
- Mark extended Robert's series in #45 with explicit input encoding, frontend parser relocation, an INSIST error-propagation macro, and a demonstration executable, reporting little backend performance change in his benchmarks.
- Andrew Dunstan reviewed both authors' work and requested a TAP-tested executable under src/test in #46; Robert rejected hidden returns in the INSIST macro and asked that unrelated multibyte cleanup move to another thread in #63.
- Robert committed the unchanged parser-split and lex_accept-removal components of #27 as 11b5e3e35d and 530609aa42, announcing the initial cleanups' landing in #63.
- Mark narrowed the remaining work to parser relocation and frontend TAP tests in #66, removing the disputed refactoring and multibyte side work; Andrew found a malformed test hunk in #70, which Mark regenerated in #71.
- After Mark's integration testing and Andrew's renewed review in #67, Robert committed the remaining unchanged components of #27 as the return-code conversion 1f3a021730 and backend-helper relocation 73ce2a03f3.
- Robert's review in #72 requested an encoding helper instead of direct table access, a frontend stack-check stub, and fatal handling for impossible parser states; Mark incorporated those changes into #74.
- Mahendra Singh Thalor supplied a compiler-warning fix in #73, which Mark temporarily included in #74, but Mahendra later confirmed Tom had already fixed the issue differently in #78.
- Robert identified unsafe macro expansion and unnecessary allocation branches in #76, and requested meaningful error-output tests in #82; Mark addressed these in #92 and removed the superseded warning patch.
- Andrew supplied Windows build support in #94, identifying both the missing shared-library source entry and additional work needed to build and locate the test executable.
- Robert committed the parser-relocation component of #92 as beb4699091, making only small include and macro-formatting adjustments and leaving the tests pending, as explained in #95.
- Tom reported Windows linker failures in #96, prompting Robert to extract the jsonapi.c build-list hunk from #94 into 006b9dcad4, while reserving the other Windows changes for the test patch in #98.
- The thread ends with frontend parser support committed and Andrew posting the remaining tests in #101, including revised MSVC build and executable-discovery logic; the tests remain uncommitted, require a small buildfarm adjustment, and skip one unresolved Windows command-line quoting case.