feat(aggregate): cost-aware partial-aggregation skip (opt-in)

datafusion/common/src/config.rs

@@ -648,6 +648,33 @@ config_namespace! {
         /// aggregation ratio check and trying to switch to skipping aggregation mode
         pub skip_partial_aggregation_probe_rows_threshold: usize, default = 100_000
 
+        /// (experimental) When true, apply a *secondary* skip rule on top
+        /// of `skip_partial_aggregation_probe_ratio_threshold`: skip
+        /// partial aggregation when the measured ratio is at least
+        /// `skip_partial_aggregation_cost_min_ratio` (default 0.5).
+        /// Targets ClickBench Q18-shape queries where the ratio (~0.56)
+        /// sits just below the fixed 0.8 threshold so partial agg keeps
+        /// running, but the absolute work (heavy variable-length keys,
+        /// complex aggregates) makes it net-negative.
+        ///
+        /// Empirical motivation: lowering the global ratio threshold to
+        /// 0.6 fixes Q18 (1.73× faster) but risks regressing low-cost
+        /// queries at similar ratios. This flag exposes the lower
+        /// threshold as a separate, opt-in knob. Whether the cost-aware
+        /// signal (`partial_agg_probe_ns_per_row` metric) can replace
+        /// this static threshold is an open question — for now the
+        /// metric is reported alongside so callers can evaluate.
+        pub skip_partial_aggregation_use_cost_model: bool, default = true
+
+        /// Number of input rows used in the A/B sampling window after the
+        /// initial partial probe completes. During this window the operator
+        /// routes input through the passthrough (`transform_to_states`)
+        /// path so the probe can measure `passthrough_ns/row` and compare
+        /// it against the previously measured `partial_ns/row`. Default
+        /// 10000 — large enough to amortise per-row noise, small enough to
+        /// be cheap if the decision turns out to be "keep partial".
+        pub skip_partial_aggregation_ab_sampling_rows: usize, default = 10_000
+
         /// Should DataFusion use row number estimates at the input to decide
         /// whether increasing parallelism is beneficial or not. By default,
         /// only exact row numbers (not estimates) are used for this decision.