model-writtenrun 5596 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Feb - Mar 2018 · 2 entries
- Euler Taveira proposed publisher-side, per-table WHERE filtering in #1, with independent executor and grammar cleanups and no initial-copy filtering yet; #5 clarified that filters belonged to individual tables rather than whole publications.
- Craig Ringer raised immutable, historically safe expression execution in #3, while Erik Rijkers reported memory growth in cascading replication tests in #4 and #7.
Oct 2018 - Feb 2019 · 2 entries
- Euler added initial COPY filtering, psql display and pg_dump support in #15, then corrected the remote relation identifier after Erik's tests in #18, which Erik verified in #19.
- Hironobu Suzuki reported ignored ALTER PUBLICATION SET filters in #20, and Petr Jelinek explained hidden function dependencies and replica-identity limitations in #21 and #29; the discussion rejected tying row filtering to RLS policies in #23 and #24.
Aug - Sep 2019 · 2 entries
- After a long gap, Alexey Kondratov rebased the series and added bidirectional-replication tests in #42, while Euler independently revised his earlier branch in #43 to fix catalog/dependency behavior and prohibit user functions.
- The two branches differed in tuple-slot handling, as explained in #52, and Euler corrected his version in #53; Movead Li's inheritance/COPY report in #54 then led to another revision in #56.
Nov 2019 · 1 entry
- Amit Langote reviewed and rebased the series in #58 and #59, but substantial progress paused until Önder Kalacı revived it in #66, demonstrating per-tuple planning overhead and unsafe custom operators while adapting partition support.
Dec 2020 - Jun 2021 · 5 entries
- Andres Freund argued for cached, safely restricted expressions in #69, and Euler reorganized the implementation and publish_via_partition_root behavior in #70, dropping redundant executor and tuple-count prerequisites.
- Michael Paquier committed the unchanged atttypmod cleanup from #72 as 4ad31bb2ef, leaving the filtering feature for further review.
- Rahila Syed's review in #81 and #82 led to documentation and naming improvements in #84, whose unchanged grammar prerequisite Peter Eisentraut committed as 91d1f2d302.
- Peter Eisentraut supplied compatibility and style fixups in #85, and Euler merged server, psql and pg_dump support in #86, documenting that old subscribers would still ignore filters during initial COPY and keeping measurement patches for testing only.
- Andres raised remaining execution-safety and per-tuple planning concerns in #93; Amit revived the patch with dedicated pgoutput executor setup in #97, and Euler updated it in #99.
Jul 2021 · 3 entries
- Peter Smith supplied expression-cache and test improvements in #100, Greg Nancarrow demonstrated and fixed repeated tuple-slot allocation costs in #102, and Euler integrated these optimizations in #105 before follow-up lifetime and node-copy fixes in #124.
- Alvaro Herrera highlighted stale or missing subscriber rows when UPDATE crossed a filter boundary in #127 and proposed UPDATE conversion in #158, while Greg laid out the transition cases in #169 and Amit supported publisher-side handling in #167.
- TOAST correctness became a separate blocker when Hou demonstrated unsafe historic detoasting after a table drop in #187, and Dilip explained the need for available replica-identity values in #188 and #194.
Aug - Oct 2021 · 4 entries
- Peter maintained the patch through upstream changes and revised cache invalidation in #195 and #208, while Euler tried an explicit old/new/both update-filter option on an earlier base in #209.
- Peter added replica-identity and expression-walker validation in #217 and #219, then collected the pending cache and completion work in #220; Ajin Cherian supplied four-way UPDATE handling in #221 so rows could become INSERTs, DELETEs, UPDATEs or no change.
- Dilip's reviews of tuple availability and deformation led Ajin to reuse cached and virtual slots in #253, #257 and #259, while schema-publication changes subsequently required staged rebases in #267 and #268.
- Tomas Vondra argued that multiple subscribed publications should combine filters with OR in #235, and Peter implemented that decision in #308 together with Greg's fixes for crashes and regular-expression lifetime failures reported by Tang in #296 and #304.
Nov 2021 · 2 entries
- Validation gaps exposed by Vignesh and Hou led to #347, which moved replica-identity checks to actual UPDATE/DELETE execution and cached the result in relcache, matching existing replication checks rather than trying to validate every DDL change.
- Vignesh supplied transformed-WHERE comparison for ALTER PUBLICATION SET in #374, folded into #378, while Hou's mixed-operation example in #345 led Peter to separate expression caches by publication action in #389.
Dec 2021 · 6 entries
- Euler produced a competing consolidation in #395 and #397, rewriting validation and temporarily leaving out UPDATE transformations and runtime validation; Ajin and Peter reconciled useful changes with the established series in #413 and #417.
- Tang showed in #407 that unrestricted ALL TABLES and schema publications must override narrower filters, and #417 and #430 implemented that behavior while consolidating initial-copy SQL and expression validation.
- Greg questioned surprising combinations of action-specific filters in #435 and #443, but Amit retained the flexibility with clearer documentation in #447, and Tang reported little cost from evaluating both row versions in #446.
- Euler proposed HeapTuple-based UPDATE handling in #454, but #463 restored the slot-based implementation to retain the required tuple-reuse behavior while folding in most of Euler's explanatory fixes.
- Vignesh demonstrated that refreshing a subscription after changing its filter does not recopy existing rows in #458, and Amit retained and documented that behavior in #462 to avoid conflicts with existing subscriber data.
- Hou added single-pass WHERE transformation with correct source text in #481, and the series later unified filtering around tuple-slot interfaces in #535 after Alvaro and Amit's API reviews.
Jan 2022 · 2 entries
- Tang's action-specific replica-identity bug report in #575 led to separate UPDATE and DELETE validity handling in #585, while Alvaro's cache-design review led #615 to collect complete publication actions and cache validity flags instead of invalid column numbers.
- Andres identified executor-lifetime and per-row EState overhead in #618, prompting #621 to cache executor state and allocate it under the decoding context; the temporarily included cache-invalidation prerequisite was removed after its separate landing in #646.
Feb - Apr 2022 · 8 entries
- The final review rounds tightened expression type, collation and function checks in #642 and #649, fixed inherited-table COPY and duplicate filters in #658, and rejected unusable partition-root filters in #662.
- Andres refined the performance methodology in #644, and Tang's late benchmarks in #661 found roughly one to four percent worst-case decoding overhead with benefits when more rows were filtered.
- Amit committed the complete feature from #665 as 52e4f0cd47, adding diagnostics, a catalog bump and an AlterPublicationOptions syscache rewrite; unsafe user-defined or mutable expressions and broader replica-identity redesign remained outside this feature.
- Peter reported an intermittent buildfarm timing failure in #671, supplied the missing subscriber catchup wait in #674, and Amit committed it unchanged as cfb4e209ec.
- Tomas found missing catalog-state and pg_dump tests in #675, Shi yu expanded them in #678, and Amit committed that patch unchanged as ceb57afd3c after Euler's review.
- Peter's cache-leak reproducer in #683 exposed a missing release in the landed AlterPublicationOptions code, leading to Hou's minimal fix in #684 and Alvaro's broader restructuring in #686.
- Amit's concurrency review refined that repair through #693, which Alvaro committed with light comments, formatting and diagnostic edits as ed0fbc8e5a.
- At the thread's end, Alvaro asked about concurrent ALTER PUBLICATION locking in #696, and Amit demonstrated serialization by tuple-update locking in #697; the feature and attached follow-up fixes had landed, with no attached series still awaiting review.