andygrove commented on PR #2765: URL: https://github.com/apache/datafusion-ballista/pull/2765#issuecomment-6616029001
TPC-H SF100 on a single machine (2 executors × 8 vcores, AQE on), three arms interleaved over 9 runs, each with a fresh cluster: `main`, this PR with `ballista.planner.adaptive.reuse_exchange=false`, and this PR with it on. - q15 runs 6 stages instead of 7 and drops from 3.73 s to 1.99 s (1.87x). - The other 21 queries are unchanged within noise. The sum of medians is 192.6 s on `main`, 194.1 s with reuse off and 192.4 s with reuse on. - Results are identical across all arms. q2 and q11 also repeat work, but not at the exchange level in the AQE plan. The join orders differ in q2, and in q11 the copies end in different aggregates once the joins are broadcast, so there's nothing for this rule to share in those two queries. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
