model-writtenrun 8447 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Bharath Rupireddy proposed extending WALReadFromBuffers() to logical walsenders and local WAL readers, with a separate extension demonstrating how to read unflushed WAL safely, in #1.
- Jingtang Zhang tested the proposal and asked for buffer-hit monitoring in #2; Bharath explained that he used experimental instrumentation and envisaged extending WAL-read statistics later in #3.
- Nitin Jadhav suggested removing a redundant variable and sharing the read logic in #4; Bharath removed the variable in #5 but declined a wrapper that would merely combine two calls.
- Jingtang reported that the WAL writer's advance initialization could erase useful buffered WAL in #6; after a long gap, Bharath defended opportunistic reads that do not impede WAL generation and dropped the demonstration extension as unnecessary documentation in #7.
- Rahila Syed asked about performance and additional readers in #8, while Jingtang approved the revised patch and highlighted its potential with direct I/O in #9; discussion with Jeff Davis distinguished caching benefits from future unflushed-WAL replication, and Bharath supplied earlier direct-I/O measurements in #10, #11, #12, #13 and #14.
- Following the rebase in #15, Bharath added a fix for stale segment descriptors and reported substantially fewer file reads with direct I/O in #17; Soumen Kumar reported successful builds, regression tests and feature testing in #18.
- Michael Paquier requested separate changes for the different reader paths and a logical-message benchmark in #19; Bharath split the series, removed a physical-walsender rewrite that mishandled retries, and supplied a segment-boundary test in #20.
- Michael identified unnecessary descriptor closing and redundant assertions in #21, leading to segment-aware closing in #22; Kirill Reshke benchmarked the revisions and requested shorter comments in #23, which Bharath adopted in #24.
- A CI timeout exposed that logical-walsender tests assumed every read reached a file, so Bharath added distinct WAL-buffer hits to pg_stat_io and updated the tests in #25; the same submission added WAL-summarizer buffer reads with supporting benchmarks.
- Michael argued that hits should count pages rather than successful function calls, requested reproducible benchmark scripts, and questioned integrating the boundary test in #26; Bharath implemented page counting, supplied scripts and stress-testing results, and removed that test in #27.
- Michael committed the statistics component of #27 as 856a6594d8, adding documentation that partial-page reads count as one hit, and announced the landing in #28.
- Bharath resent the unchanged logical-walsender, local-reader and WAL-summarizer patches as the remaining series in #29.
- Hayato Kuroda raised questions about performance when buffers consistently miss and about logical_read_xlog_page() failing to update seg.ws_tli when bypassing file reads in #30.
- Bharath explained that insufficient buffers, lagging consumers and opportunistic page initialization cause buffer misses, and acknowledged that an unset timeline affects error reporting and validation checks, proposing to update seg.ws_tli after buffer-only reads in #31.
- Bharath implemented timeline updates and descriptor closing on timeline changes across all three reader paths in #32; the statistics component remains landed, while the reader patches remain under review and the performance concern about persistent buffer misses has no new benchmark resolution.