model-writtenrun 3495 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Matthias van de Meent proposed shrinking textual node serialization by omitting default-valued fields, resetting query locations and trimming trailing zero bytes, reporting over 50% smaller raw output and roughly one-third smaller pg_rewrite storage #1.
- Peter Eisentraut questioned whether machine storage and debugging should share a format, challenged curated common values being called defaults, and suggested separating location removal; Matthias explained that stored views lack the original query text and ordinary deserialization already discards locations #2 #3.
- David Rowley suggested generated metadata supporting separate human-readable and binary formats, and Matthias explored that direction while treating compact prepared-plan storage as future work #4 #5 #7 #8.
- Following Peter's request for incremental changes, Matthias split the proposal into seven patches covering default omission, catalog location resetting, custom defaults, trailing-zero trimming and list or bitmap encoding, while reporting that binary serialization would require more infrastructure #6 #9.
- Peter identified unsafe conditional-expression macros, proposed matching the existing location-reading toggle, and suggested dedicated location and typmod types to avoid hand-maintained exceptions; he also questioned the readability and maintenance costs of curated defaults #10.
- Peter asked for benefits beyond storage savings, Matthias acknowledged that performance remained unmeasured, and Robert Haas opposed bitmap differential encoding in a format humans must read #12 #13 #14.
- Matthias substantially reworked the series around serialization-time location control, Location and TypMod types, NULL-constant omission, run-length encoding and copied Var fields; bitmap differential encoding, trailing-zero trimming and broad curated defaults were absent from this version #15.
- Two immediate fixes addressed an AddressSanitizer-detected token-reader overread and bitmap serialization that skipped bit zero, with the latter exposed by WRITE_READ_PARSE_PLAN_TREES testing #16 #17.
- Peter's next review prompted simpler field-reading macros, a separate optional patch for enum-zero omission and sign-sensitive floating-point default checks; Matthias submitted those revisions, corrected an accidental unrelated attachment during a rebase, and then fixed restored locations being overwritten with -1 #18 #20 #21 #22.
- Peter supplied a smaller standalone location-output implementation using file-local state and suggested the more specific ParseLoc name; Matthias incorporated both ideas into the full series despite preferring explicitly passed context #23 #24.
- Peter committed the ParseLoc component of #24 as 605721f819, relocating the typedef, correcting its pgindent entry and extending conversion to missed PlannedStmt fields and the newly added MergeSupportFunc node #25.
- While reviewing the remaining location patch, Peter found that converting every catalog caller would be fragile and instead proposed making nodeToString() discard actual locations by default, with nodeToStringWithLocations() reserved for debugging; Matthias accepted this reversal #27 #28.
- Peter committed #27 unchanged apart from rebasing as d20d8fbd3e, completing the two location-related components of Matthias's series #29.
- At the thread's close, Peter declined the remaining NULL, run-length, copied-field and enum optimizations because of debugging costs, and left general default omission and TypMod pending broader agreement and serialization performance measurements; separate debug toggles and a binary storage format remained possible future directions #29.