model-writtenrun 8152 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Thomas Munro questioned the receive-side tuple copy in #1 and supplied #2, letting Gather consume queue memory directly while retaining an explicit copy for GatherMerge buffering.
- Robert Haas raised concerns about separately stored HeapTuple data in #3, and Andres Freund proposed MinimalTuple transport in #4, reasoning that system columns should already be projected below Gather.
- Thomas implemented MinimalTuple transport in #5, reporting over 10% improvement in simple tests, then documented the required projection behavior in #6.
- Soumyadeep Chakraborty tested both approaches in #7 and #9, confirmed gains for scans and joins returning many rows, and identified a remaining sender-side copy; Thomas explained why aggregate queries sending very few tuples would see little benefit in #8.
- Thomas committed #6 unchanged as cdc7169509 and confirmed the landing in #10, leaving direct tuple formation into destination memory for separate work.
- Amit Khandekar opened that follow-up in #11, exposing existing heap or minimal tuple bytes to avoid a sender-side copy, reporting 4–10% improvements, and attaching a testing-only GUC to compare performance without restarts.
- Thomas questioned the incomplete tuple interface and receiver-side shared-memory writes in #12, while Andres advocated a smaller MinimalTuple representation and support for virtual slots and other consumers in #13.
- At the thread’s end, Amit proposed investigating direct tuple formation into queue-provided memory separately from shrinking MinimalTuple in #14; the receive-side change had landed, but the sender-side follow-up remained unresolved.