model-writtenrun 8390 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Vladimir Sitnikov reported a reproducible WAL-sender crash found by pgjdbc tests in #1, and Michael Paquier reproduced it with physical START_REPLICATION on a database-connected replication session in #4.
- Kyotaro Horiguchi proposed rejecting that combination and added tests for both incompatible replication directions in #5, but Vladimir warned that rejecting previously working connections could break applications in #6.
- Michael opposed the rejection and identified WAL-reader initialization as the proper repair in #10, while Kyotaro argued that the documented connection modes supported the restriction in #11.
- Michael supplied a replacement patch in #13 that allocates the physical WAL reader when START_REPLICATION executes, removes the fake reader, and retains Kyotaro's test rejecting logical decoding without a database connection.
- Kyotaro accepted the reader cleanup while raising privilege and error-detail concerns in #15; Fujii Masao requested a test of the preserved physical-replication behavior in #14, but Michael explained the available clients' testing limitations in #16.
- Álvaro Herrera considered the replacement cleaner but still favored restricting connection combinations in #19, whereas Andres Freund argued that querying catalogs and streaming on the same connection was useful in #20 and demonstrated SQL support in #26.
- Michael separated the crash repair from the unresolved policy question in #23 and committed #13 unchanged as 879ad9f90e, announcing the fix in #30.
- Jonathan S. Katz reopened the PostgreSQL 13 open item for the remaining policy decision in #31, and Andres supplied use cases involving physical backups and logical replication over one connection in #32.
- The renewed discussion focused on access controls: Kyotaro requested documentation clarification in #34, Fujii wanted physical-replication rejection rules enforced across connection types in #35, and Stephen Frost called for administrator control over permitted operations in #37.
- Robert Haas favored preserving useful existing behavior and adding finer access controls in #39, while Andres confirmed actual users of the combined functionality and supported more granular permissions in #42.
- Álvaro accepted preserving the existing behavior and closed the open item in #44; the crash fix had landed, the proposed prohibition was abandoned, and no further documentation or access-control patch was submitted in this thread.