adriangb commented on PR #23390: URL: https://github.com/apache/datafusion/pull/23390#issuecomment-4920862352
## Final benchmark picture (head 775c059d88, adaptive targeting on by default) Two runs prove the two halves of the design contract: **1. Default run (merge-base vs this branch, no memory limit).** With an infinite pool the adaptive path derives no target and the normalizer wrapper is never installed, so this is a same-code-path regression check: | suite | main | branch | detail | |---|---|---|---| | tpch | 760.3ms | 756.8ms | 22/22 no change | | tpcds | 10072ms | 10152ms (+0.8%) | 3 faster / 3 slower / 93 no change, movers <=1.09x with overlapping error bars | | clickbench_partitioned | 19738ms | 19992ms (+1.3%) | 2 faster / 8 slower / 33 no change; flagged queries have near-identical minimums (e.g. Q29 "1.40x" comes from one ±26ms-stddev outlier run, mins 40.7 vs 41.9ms) | **2. Same-SHA A/B under `DATAFUSION_RUNTIME_MEMORY_LIMIT=2G` (baseline = adaptive disabled, changed = default).** This is the regime where adaptive activates (derived target ~10.9MiB at 12 partitions): | suite | adaptive off | adaptive on | detail | |---|---|---|---| | tpch | 775.7ms | 781.2ms (+0.7%) | 1 slower (Q2, 1.09x) / 21 no change | | sort_tpch | 3668ms | 3707ms (+1.0%) | 10/10 comparable queries no change; Q3 (VARCHAR-key sort) fails on BOTH sides — pre-existing 2G capacity limit, unaffected by this PR | | clickbench_partitioned | 30116ms | 30573ms (+1.5%) | 5 faster / 13 slower / 25 no change, movers <=1.16x both directions | | external_aggr | 1492.7ms | 1513.3ms | 8/8 no change (its per-query small pools fall under the adaptive minimum and keep exact behavior) | Summary: with no memory limit the feature is provably inert; under a memory limit on standard (small-batch) data it costs ~0.7-1.5% total; and for the workloads it exists for (`large_values`: 100MB-536MB decoded batches), the earlier comments show it turning ResourcesExhausted into completed queries with zero configuration. 🤖 Generated with [Claude Code](https://claude.com/claude-code) https://claude.ai/code/session_01UcPTREZVLXsSZDRCae33gm -- 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]
