model-writtenrun 8836 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Deng, Gang reported poor parallel hash join scaling caused by cache-line sharing and proposed testing the match bit before writing it, reducing a 28-worker TPC-DS query from 103.5 to 41 seconds in his tests #1.
- Thomas Munro proposed eliminating match-bit writes entirely for parallel joins because parallel right/full joins were unsupported and the parallel condition could be constant-folded; Tom Lane preferred testing the join type while preserving the serial path’s lack of a runtime branch #2 #3.
- Deng, Gang accepted the parallel-condition approach and explained that repeated writes caused cache-coherence traffic across cores #4 #5.
- Thomas Munro posted the sole indexed patchset, using the parallel condition and an assertion while retaining unconditional serial writes; he deferred join-type specialization because a runtime check was undesirable for a backpatched fix #6.
- Thomas Munro committed the patch unchanged as 3e4818e9dd, with backpatches through PostgreSQL 11, and reported modest improvements on simpler many-core queries without rerunning TPC-DS #7.
- The submitted fix was complete at the thread’s end, while Thomas Munro identified separate unresolved scaling issues involving shared-memory chunk allocation and possible huge-page benefits for future investigation #7.