model-writtenrun 8431 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Cary Huang proposed sequence decoding through output plugins in #1, logging every nextval call to expose current values while leaving built-in publisher/subscriber support for later work.
- Andres Freund questioned the extra WAL and decoding overhead, requested tests and objected to exposing sequence tuples in #3, while Michael Paquier emphasized the performance cost of changing WAL batching in #4.
- Cary initially defended exact subscriber values in #5, then posted #6 with regression tests and cleanup, restoring batched logging but recording current rather than future sequence values.
- Andres explained in #7 that failover requires safe future values rather than exact synchronization, and Craig Ringer requested rollback and concurrent-transaction tests in #8; Cary accepted the gap-tolerant design in #9.
- Cary's final attachment #10 restored the original WAL behavior, made sequence output optional in test_decoding and added rollback/savepoint tests, but reported that rollback suppressed decoded advances even though the sequence itself advanced.
- The thread ended with Andres recommending nontransactional handling modeled on logical messages in #11, leaving that correctness issue and Cary's concurrent-test questions unresolved, with no landing announced here.
- The later infrastructure commit 0da92dc530 came from #497855 in thread #44344, introducing transactional/nontransactional handling and relfilenode tracking while crediting Cary's original patch.
- The test_decoding commit 80901b3291 came from the later #498904 in thread #44344, adding dedicated sequence callbacks and an include-sequences option.
- Built-in replication subsequently landed as 75b1521dae from #502813 in thread #44344, adding publication, protocol and subscriber support beyond this thread's attachments.