avantgardnerio commented on code in PR #7192:
URL: https://github.com/apache/arrow-datafusion/pull/7192#discussion_r1294011416
##########
datafusion/core/src/physical_plan/aggregates/priority_queue.rs:
##########
@@ -0,0 +1,235 @@
+use uuid::Uuid;
+use std::collections::{BTreeMap, HashMap};
+use std::pin::Pin;
+use std::sync::Arc;
+use std::task::{Context, Poll};
+use arrow::row::{OwnedRow, RowConverter, SortField};
+use arrow::util::pretty::print_batches;
+use arrow_array::{RecordBatch};
+use arrow_schema::{DataType, SchemaRef, SortOptions};
+use futures::stream::{Stream, StreamExt};
+use hashbrown::HashSet;
+use datafusion_common::DataFusionError;
+use datafusion_execution::TaskContext;
+use datafusion_physical_expr::{PhysicalExpr};
+use crate::physical_plan::aggregates::{aggregate_expressions, AggregateExec,
evaluate_group_by, evaluate_many, group_schema, PhysicalGroupBy};
+use crate::physical_plan::{RecordBatchStream, SendableRecordBatchStream};
+use datafusion_common::Result;
+use datafusion_physical_expr::expressions::{Max, Min};
+
+pub(crate) struct GroupedPriorityQueueAggregateStream {
+ schema: SchemaRef,
+ input: SendableRecordBatchStream,
+ aggregate_arguments: Vec<Vec<Arc<dyn PhysicalExpr>>>,
+ group_by: PhysicalGroupBy,
+ group_converter: RowConverter,
+ value_converter: RowConverter, // TODO: use accumulators
+ group_to_val: HashMap<OwnedRow, OwnedRow>, // TODO: BTreeMap->BinaryHeap,
OwnedRow->Rows
+ val_to_group: BTreeMap<OwnedRow, HashSet<OwnedRow>>,
Review Comment:
I got it on par with `main` using just a `BTreeSet` as far as performance. I
have a branch where I have a BinaryHeap, but to be safe, the map needs to store
indexes into it, which killed performance. The same would be true of a sorted
vec.
I tried using `Arc<Mutex<>>`s as pointers, but that performed even worse.
I rewrote it with raw pointers and it is pretty fast! So it all depends how
much we want performance vs how much we don't want to maintain `unsafe` custom
binary heaps.
--
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]
