model-writtenrun 316 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Noah Misch opened the thread by finding missing password restoration, dependency-ordering problems and duplicated pg_dump logic in the existing DDL functions, attaching tests for additional tablespace and permission-error defects in #1.
- After Melanie Plageman requested a release decision in #2, Andrew Dunstan said fixes were underway, but Noah clarified that the design needed either a substantive defense or a revert in #3 and #4.
- Tom Lane supported avoiding independent DDL-generation implementations and suggested exploring pg_dump refactoring, while Nathan Bossart mentioned a pg_dump library and Andrew defended SQL-accessible introspection despite its maintenance cost in #5, #6 and #7.
- Andres Freund found inconsistent privilege checks, missing concurrency protection, repetitive implementation code and inadequate tests in #8; Andrew proposed pg_dump equivalence tests and further fixes in #9.
- Andrew concluded that the required catalog-access permission model and testing changes were too invasive for the imminent release, proposing removal from release 19 while retaining the feature on master in #11.
- Robert Haas distinguished standalone object cloning from full backup restoration, while Noah maintained that migration was an advertised use case and required dependency handling; Robert nevertheless agreed that omitting passwords was incorrect in #15, #16 and #17.
- Tom explored having pg_dump call server functions but identified bulk-performance and statement-splitting obstacles, while Andres proposed sharing SQL generation through a library separate from catalog collection in #13 and #18.
- Robert raised cross-version compatibility and silent divergence concerns, Tom favored dropping the feature if duplication proved unavoidable, and Andrew reiterated the object-cloning use case while suggesting a possible dump-specific mode in #21, #22, #23 and #25.
- Andrew implemented the release-19 removal in db169985c1, then confirmed completion and posted the corrective series developed with Euler Taveira in #26; the revert came directly from the discussion rather than either attachment.
- The uncommitted series in #26 proposed password output, replay-safe role attributes, membership ordering, tablespace fixes, optional database-specific settings, object locking, shared helpers, catalog-based permissions and stronger regression and pg_dump comparison tests.
- Michael Paquier reported that the revert left ddlutils.o in the Makefile and broke release-19 builds in #27, then removed that reference in 4664929e5a; Andrew explained that his checks had used Meson in #28.
- Mario González Troncoso later proposed splitting role DDL into separate components and having pg_dumpall consume server-generated SQL in #29, but Noah said the proposal did not adequately address previous feedback and Mario declined to proceed in #30 and #31.
- At the thread's end, Robert remained skeptical of server-generated output serving pg_dump's version requirements and proposed broader round-trip or DDL-replication testing in #32; the release-19 revert and build repair had landed, while the corrective series and larger redesign remained unresolved.