model-writtenrun 8871 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Pierre Ducroquet opened the thread with two performance experiments: simplifying PostmasterIsAlive for older PostgreSQL versions and caching the WAL flush position to reduce spinlock contention among logical-decoding walsenders, reporting large decoding speedups in #1.
- Julien Rouhaud requested InvalidXLogRecPtr initialization and comments explaining the cache in #2; Pierre incorporated those changes in #3 and reported approximately 5% improvement when the decoder had to process the changes fully, while Julien subsequently questioned a redundant initialization check in #4.
- Pierre requested a PostgreSQL 12 backport in #5; Tom Lane supported that scope but identified broken patch whitespace in #6, prompting the corrected submission in #7, where Pierre clarified that the PostmasterIsAlive experiment only helped reproduce the problem on older releases and that broader backpatching carried excessive risk.
- Marc Cousin supplied reproducible benchmarks in #9, reporting approximately 20% faster pgbench runs and substantially lower walsender CPU consumption with the caching patch under activity spread across multiple databases.
- Tom Lane committed #7 as e369f37086 with comment cleanup and an else-if avoiding a redundant initialization fetch, backpatched the change to PostgreSQL 12, and announced completion in #10; the auxiliary PostmasterIsAlive experiment was left out.