model-writtenrun 7097 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Dec 2023 - May 2024 · 11 entries
- Jelte Fennema-Nio proposed ParameterSet so connection poolers could inspect setting changes without parsing SQL, reuse ParameterStatus values and eventually control settings unavailable through SQL, with backend, libpq and pipeline-test support in #1.
- Jeff Davis raised older-server compatibility in #3, while Tom Lane questioned mixing nontransactional changes with ordinary GUCs in #4; Jelte consequently added negotiation and version-reporting prerequisites in #12.
- Jacob Burroughs proposed a protocol-only GUC flag and a list of ordinary settings convertible into protocol-controlled settings in #15, which Jelte implemented in #16 and split into smaller patches in #17.
- Robert Haas opposed dynamically converting arbitrary GUCs in #18, and Tom suggested dedicated GUC contexts in #24; Jelte adopted those contexts, restored transactional behavior for ordinary settings and prohibited protocol-setting changes inside transactions in #26.
- Jelte removed a source-precedence rule that prevented overriding SET values, added configurable report_parameters and relaxed unknown managed-parameter names to warnings in #27, while explaining that configurable reporting mattered more to him than managed parameters in #29.
- Jelte added psql support and tests in #31, then clarified backward negotiation after Robert's compatibility objections; Robert acknowledged that explanation but favored feature-specific APIs that hide negotiation details in #34.
- After API cleanup and conflict rebases, Jelte added max_protocol_version as an escape hatch for older servers and incompatible poolers in #41, while Robert continued questioning the public namespace and coupling between client and pooler protocol state in #42 and #44.
- Peter Eisentraut distinguished generally applicable protocol-version changes from optional extensions in #45, but the design remained unsettled, and Robert ruled out a PostgreSQL 17 landing for lack of consensus and sufficient review in #52.
- Jelte accepted postponement in #53 and feature-specific APIs in #54, while subsequent discussion addressed startup round trips, optional features and the burden of translating between protocol versions in poolers.
- Robert argued that ordinary GUCs did not need the new setting mechanism in #64, and discussion of pooler reuse led Jelte to distinguish compatible version changes from independently adjustable protocol features in #75.
- Tom explicitly advocated infrastructure separate from transactional GUCs in #87 and #91, and Robert supported separate reporting so client-to-pooler settings could differ from pooler-to-server settings in #98.
Jun - Sep 2024 · 10 entries
- Jelte's rewrite in #100 introduced dedicated protocol-parameter infrastructure and negotiation/set/completion messages, retained report_parameters as its demonstration feature, and dropped the earlier generic GUC-setting and managed-parameter approach.
- The rewritten series in #100 retained protocol 3.0 as libpq's default for compatibility and proposed protocol 3.2, skipping 3.1 because existing PgBouncer versions incorrectly advertised support for it.
- Robert endorsed maximum/minimum negotiation semantics in #105, then found parsing mistakes, incorrect parameter offsets, weak error reporting and incomplete message-flow documentation in #111.
- Jelte addressed those findings and moved tracing fixes to another thread in #112, then incorporated Robert's further review from #114 into #115, separating basic version negotiation from unfinished parameter infrastructure.
- Robert committed the backend negotiation component of #115 unchanged as 516b87502d, ensuring that the server never negotiates a version above the client's request, and announced it in #116.
- Heikki Linnakangas advocated requesting newer versions during development to exercise negotiation in #117 and #121, while Robert emphasized the compatibility costs and left the release-default decision open in #120 and #123.
- Jacob Champion demonstrated that changing PQprotocolVersion would break existing callers in #118 and #124, leading Robert and Heikki to support a separate API in #125 and #126.
- The API dispute expanded into guarantees for minor protocol changes: Jacob sought explicit intermediary compatibility in #150, and Jelte proposed retaining protocol-3.0 behavior after startup unless new features were activated in #155.
- Robert urged an incremental API change in #159, Jelte accepted the name PQfullProtocolVersion in #160, and Jacob supplied the extracted replacement implementation in #161.
- Robert approved Jacob's patch in #162 and committed it unchanged as cdb6b0fdb0, preserving PQprotocolVersion while exposing the full version through a new function, as announced in #163.
Dec 2024 · 1 entry
- Returning after several months, Jelte added the missing Windows export for the committed function, restored the tracing patch, introduced min_protocol_version and added version tests in #164, while explicitly leaving earlier feedback on protocol parameters and report_parameters unresolved.
Feb 2025 · 3 entries
- After rebasing in #165, Jelte ended this thread by moving the preparatory work to the longer-cancel-keys discussion in #166, omitting the standalone latest-version pipeline-test change and both protocol-parameter patches from that continuation.
- The transferred negotiation work subsequently landed as 5070349102 from the later #588045 in thread #49258, which already included duplicate-message detection and revised parser error handling, so its source attribution belongs to that continuation.
- The options and protocol-version work subsequently landed together as 285613c60a from #588045 in thread #49258, with explicit supported-version parsing and revised tests and documentation; the original protocol-parameter and configurable-reporting feature remained unfinished in the supplied thread.