model-writtenrun 5509 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Apr - Jun 2022 ยท 19 entries
- Alvaro Herrera opened the thread with a reproducer showing that combining publications with different row filters and column lists transmitted values excluded by each publication's own filter, potentially exposing sensitive data #1.
- Tomas Vondra proposed evaluating filters separately for each publication, while Amit Kapila initially defended independent combination of publication properties and raised performance and initial-copy concerns; Alvaro suggested prohibiting unsupported combinations if a complete solution was impractical #2 #3 #6.
- Tomas supplied a prototype that retained a union of columns for relation metadata but selected transmitted values using each publication's row filter, sending NULL for excluded values #11.
- After Amit identified missing initial synchronization, Tomas extended the prototype with per-column CASE expressions in COPY queries, and Alvaro supplied a delta improving comments and optimizing the single-publication case #12 #15 #21.
- A parallel discussion exposed differences between initial copying and streaming when publications select different actions; Amit favored documenting that initial copying ignores publication actions instead of imposing equivalent restrictions on row filters #16 #19 #20 #64.
- The participants explored checking combinations during subscription creation and after publication changes, then recognized that existing relation-cache invalidation lets each walsender detect changed definitions without a new shared-memory signaling mechanism #31 #33 #36 #38 #39.
- Peter Eisentraut argued that column lists primarily describe a static subscriber table shape, and Amit proposed rejecting differing lists for the same table; Hou volunteered to implement that restriction #42 #43 #47.
- Tomas accepted the restrictive direction without being convinced, emphasizing that it excluded useful data-redaction configurations but could later be relaxed compatibly, so his filtering prototype and its fix-up were superseded #45 #48.
- Tomas and Amit settled on enforcement through pgoutput's existing cache refresh, accepting that ALTER PUBLICATION could succeed before replication reported an error and might require subscription repair because decoding uses historical snapshots #52 #53 #54 #55.
- Hou submitted the replacement implementation checking column-list compatibility during subscription creation, initial copying and streaming #56.
- Amit requested exposing publication column lists and row filters through the existing catalog view instead of constructing a complex subscriber query, and Hou split the work into a catalog prerequisite and the restriction patch #57 #59.
- Review led to displaying column names and treating explicit all-column lists like omitted lists; Hou also changed subscription creation from WARNING to ERROR after establishing that pg_dump restores subscriptions with connect=false #60 #61.
- Shi Yu, Takamichi Osumi and Amit requested clearer comments, documentation corrections and tests for the new error paths, which Hou incorporated into the final two-patch submission #63 #65 #66 #67.
- Amit committed the catalog prerequisite from that submission as 0ff20288e1, adding a catalog-version bump and minor formatting while leaving the restriction patch under review #69.
- Tom Lane questioned the post-beta catalog change and compatibility with older publishers; Amit explained why subscription-side validation needed the information and retained the catalog approach, while Hou added a publisher-version check to the remaining restriction patch #71 #72 #73 #75.
- Amit polished comments and formatting, Justin Pryzby corrected the restriction documentation, and Hou incorporated those corrections into the final attachment #76 #77 #78.
- Amit landed the restriction as fd0b9dcebd, reproducing the final attachment's six changed files exactly while expanding the commit explanation and adjusting reviewer credits #78 #79.
- Justin's separate language corrections for the already committed catalog work prompted a wording debate with Peter Smith; Amit chose to retain the existing wording without consensus, leaving those optional corrections uncommitted in this thread #70 #80 #81 #82 #83.
- With both functional changes landed and no objections to closing the reported issue, Amit closed the PostgreSQL 15 open item at the end of the thread #83 #84.