model-writtenrun 5644 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Oct - Nov 2019 · 2 entries
- Jeff Janes reported that logical replication of a nearly inactive table still transmitted BEGIN/COMMIT pairs for unrelated transactions, and proposed delaying BEGIN until useful data appeared in #1.
- Euler Taveira implemented that approach in #2, drawing on test_decoding, and Jeff confirmed sharply reduced traffic while observing frequent keepalives in #3.
Mar 2020 · 2 entries
- Dilip Kumar initially approved the patch in #4, but Amit Kapila questioned restart_lsn advancement in #5, leading to discussion of local acknowledgement and transaction-count, time, or WAL-distance thresholds in #6, #7, and #14.
- Craig Ringer explained that existing feedback already advances replication positions but synchronous commits need timely responses in #19; Andres Freund proposed extending the progress callback in #20, and Craig favored reusing keepalives rather than changing the wire protocol in #21.
Jul 2020 · 1 entry
- Rahila Syed reported passing checks and initially reassuring synchronous-replication measurements in #24, but the patch became stale and was returned with feedback in #25.
Apr - May 2021 · 2 entries
- Ajin Cherian revived Euler's work in April 2021 with streaming compatibility and logical-message handling in #26, then incorporated Peter Smith's testing, state-name improvements, and cleanup suggestions through #32.
- After the rebase in #33, Vignesh requested tests and diagnostic logging in #34, and Ajin expanded the implementation to prepared transactions with per-transaction state and protocol changes in #35.
Jul - Sep 2021 · 4 entries
- Peter Smith, Takamichi Osumi, and Greg Nancarrow reviewed the expanded patch, leading to test-protocol, assertion, documentation, and formatting improvements through #47; all three subsequently found it satisfactory in #48, #49, and #50.
- Peter's measurements in #51 showed substantial traffic savings and modest throughput gains, but Amit identified unresolved synchronous-feedback and prepared-transaction restart issues, plus protocol compatibility concerns, in #52.
- Ajin split ordinary and prepared transactions into separate patches in #53, then withheld prepared-transaction suppression and introduced a progress-callback flag for synchronous keepalives in #55.
- Peter's reviews simplified the deferred keepalive mechanism and naming through #59, after which work resumed with a January 2022 rebase in #60.
Jan - Mar 2022 · 15 entries
- Osumi found missing cleanup for prepared-transaction state in #61 and #65, prompting COMMIT PREPARED and ROLLBACK PREPARED fixes in #63 and #66.
- Osumi reconsidered direct keepalive timing and ways to preserve knowledge of skipped prepares in #67 and #68, while Amit rejected extra disk I/O for a network optimization and suggested documenting the prepared-transaction limitation in #70.
- Following Amit's review in #69, Ajin added deferred STREAM START and empty-stream suppression in #77, while removing unnecessary prepared-transaction state allocation.
- Amit and Osumi's timing discussion culminated in passing the decoded end position directly to WalSndKeepalive(), which Ajin implemented and manually checked in #80 instead of waiting for WalSndWaitForWal().
- Reviews by Wang Wei in #78 and Peter in #79 led to clearer stream-state tracking and helper naming in #81.
- Shi Yu demonstrated that BEGIN was sent before row filtering and that skipped streamed prepares caused missing-prepared-transaction errors in #82; Ajin fixed the row-filter ordering in #84 but confirmed the deeper streamed-prepare problem in #83.
- Ajin separated streaming suppression into a second patch in #85, corrected the ordinary patch's logical-message streaming guard after Shi's review in #87, and rebased both parts in #89.
- Amit recommended dropping streaming suppression until its prepared-transaction problem was solved in #90, and Ajin removed it, documented both exclusions, and added a basic test in #91.
- Amit supplied a cleaned-up version in #92 and requested a test that actually proves suppression; Zhijie Hou took over the final revisions, adding DEBUG-log verification and correcting an initially wrong attachment in #93 and #94.
- Amit supplied a separate cleanup delta in #95, which Hou folded into #96 along with subscriber-content verification and a keepalive reply-flag adjustment; Shi's bandwidth measurements in #97 confirmed savings in both synchronous and asynchronous modes.
- Hou explained why a keepalive for each skipped synchronous transaction was still necessary in #98, then adapted deferred BEGIN handling to newly added sequence decoding in #99.
- Amit identified the need to flush keepalives promptly in #100, and Hou extracted shared pending-write processing and removed the lag-tracker timing dependency from keepalive delivery in #101.
- Sawada's review in #102 corrected the publisher log offset and minor code details, Amit requested a shorter helper name in #104, and Hou produced the final submission in #105.
- Shi reported improved synchronous throughput when testing #101 in #106, and Sawada approved the updated patch in #107.
- Amit committed #105 as d5a9d86d8f with only a test-comment wording correction and a small commit-message edit, then announced the push in #108; ordinary empty transactions were handled, while prepared and streamed transaction suppression remained deliberately excluded.