ARROW-6803: [Rust] [DataFusion] Performance optimization for single p…

…artition aggregate queries This PR optimizes the case where there is a single partition being aggregated. In this use case there is no need for a secondary aggregation step. I ran benchmarks to confirm that this addresses the performance regression. Closes apache#5590 from andygrove/ARROW-6803 and squashes the following commits: 213fc10 <Andy Grove> Avoid secondary aggregate for single partitions Authored-by: Andy Grove <[email protected]> Signed-off-by: Andy Grove <[email protected]>
Deegue · Oct 8, 2019 · 02d1e97 · 02d1e97
1 parent 583fb7e
commit 02d1e97
Showing 1 changed file with 6 additions and 0 deletions.
diff --git a/rust/datafusion/src/execution/physical_plan/hash_aggregate.rs b/rust/datafusion/src/execution/physical_plan/hash_aggregate.rs
@@ -91,6 +91,12 @@ impl ExecutionPlan for HashAggregateExec {
             })
             .collect();
 
+        if partitions.len() == 1 {
+            // if there is only a single partition then it isn't necessary to perform any
+            // additional logic
+            return Ok(partitions);
+        }
+
         // create partition to combine and aggregate the results
         let final_group: Vec<Arc<dyn PhysicalExpr>> = (0..self.group_expr.len())
             .map(|i| Arc::new(Column::new(i)) as Arc<dyn PhysicalExpr>)