Skip to content

Apache DataFusion SQL Query Engine

refactor(hash-aggr): Small fixes and finish hash aggregation refactor (#24961)

Yongting You5b389eb23cb4
diff --git a/datafusion/physical-plan/src/aggregates/mod.rs b/datafusion/physical-plan/src/aggregates/mod.rs
index 69fb09a8b1..ba08c4b003 100644
--- a/datafusion/physical-plan/src/aggregates/mod.rs
+++ b/datafusion/physical-plan/src/aggregates/mod.rs
@@ -858,7 +858,20 @@ pub struct AggregateExec {
/// FILTER (WHERE clause) expression for each aggregate expression
/// The same reason to [`Arc`] it as for [`Self::group_by`].
filter_expr: Arc<[Option<Arc<dyn PhysicalExpr>>]>,
- /// Configuration for limit-based optimizations
+ /// Soft limit for best-effort, limit-based optimizations.
+ ///
+ /// `AggregateExec` has multiple stream implementations and not all of them
+ /// support the optimization, so this is only a hint. The downstream limit
+ /// operator enforces the exact row count either way.
+ ///
+ /// Supported by:
+ /// - [`StreamType::GroupedPriorityQueue`]: retains only the best `limit`
+ /// groups per partition (this stream is selected only when a limit is set)
+ /// - [`StreamType::PartialHash`], [`StreamType::FinalHash`] and the legacy
+ /// [`StreamType::GroupedHash`]: stop reading input once `limit` groups
+ /// have been accumulated
+ ///
+ /// The remaining streams consume all input.
limit_options: Option<LimitOptions>,
/// Input plan, could be a partial aggregate or the input to the aggregate
pub input: Arc<dyn ExecutionPlan>,
@@ -1036,9 +1049,9 @@ impl AggregateExec {
InputOrderMode::Linear
};
- // To keep the ordering optimization simple, grouping sets are not supported.
- // See [`OrderedPartialAggregateStream`] for more details about this optimization.
- if group_by.has_grouping_set() {
+ // Input order mode is also used to advertise plan output ordering, grouping
+ // sets handling, and partial reduce aggregation can't promise that.
+ if group_by.has_grouping_set() || mode == AggregateMode::PartialReduce {
input_order_mode = InputOrderMode::Linear;
}
@@ -1216,11 +1229,14 @@ impl AggregateExec {
//
// # Implementation Note
//
- // `GroupedHashAggregateStream` is being incrementally refactored. See the
- // tracking issue for details.
+ // `GroupedHashAggregateStream` is the legacy implementation of all the
+ // branches below. It has been split into dedicated streams, and the
+ // `enable_migration_aggregate` config option selects between the new
+ // streams and the legacy implementation.
//
- // New features and improvements should go directly into the new implementation.
- // Please coordinate through the tracking issue.
+ // The legacy implementation is deprecated. It is kept for one more
+ // release as a fallback for potential bugs in the new streams, and
+ // will be removed after that.
//
// Issue: <https://github.com/apache/datafusion/issues/22710>
if context
@@ -1270,6 +1286,10 @@ impl AggregateExec {
self, context, partition,
)?));
}
+
+ return internal_err!(
+ "All aggregate cases should be able to handle by the new path"
+ );
}
// Execution paths that have not been migrated use the fallback implementation
@@ -1292,7 +1312,6 @@ impl AggregateExec {
self.mode == AggregateMode::Partial
&& self.input_order_mode == InputOrderMode::Linear
&& !self.group_by.is_true_no_grouping()
- && self.limit_options_supported_by_hash_stream()
}
fn should_use_ordered_partial_aggregate_stream(
@@ -1302,22 +1321,19 @@ impl AggregateExec {
self.mode == AggregateMode::Partial
&& self.input_order_mode != InputOrderMode::Linear
&& !self.group_by.is_true_no_grouping()
- && self.limit_options_supported_by_hash_stream()
}
fn should_use_final_hash_stream(&self, _context: &TaskContext) -> bool {
matches!(
self.mode,
AggregateMode::Final | AggregateMode::FinalPartitioned
- ) && self.limit_options_supported_by_hash_stream()
- && self.input_order_mode == InputOrderMode::Linear
+ ) && self.input_order_mode == InputOrderMode::Linear
&& !self.group_by.is_true_no_grouping()
&& self.group_by.is_single()
}
fn should_use_partial_reduce_hash_stream(&self, _context: &TaskContext) -> bool {
self.mode == AggregateMode::PartialReduce
- && self.limit_options.is_none()
&& self.input_order_mode == InputOrderMode::Linear
&& !self.group_by.is_true_no_grouping()
&& self.group_by.is_single()
@@ -1327,8 +1343,7 @@ impl AggregateExec {
matches!(
self.mode,
AggregateMode::Single | AggregateMode::SinglePartitioned
- ) && self.limit_options.is_none()
- && self.input_order_mode == InputOrderMode::Linear
+ ) && self.input_order_mode == InputOrderMode::Linear
&& !self.group_by.is_true_no_grouping()
}
@@ -1336,8 +1351,7 @@ impl AggregateExec {
matches!(
self.mode,
AggregateMode::Single | AggregateMode::SinglePartitioned
- ) && self.limit_options.is_none()
- && self.input_order_mode != InputOrderMode::Linear
+ ) && self.input_order_mode != InputOrderMode::Linear
&& !self.group_by.is_true_no_grouping()
}
@@ -1345,17 +1359,11 @@ impl AggregateExec {
matches!(
self.mode,
AggregateMode::Final | AggregateMode::FinalPartitioned
- ) && self.limit_options_supported_by_hash_stream()
- && self.input_order_mode != InputOrderMode::Linear
+ ) && self.input_order_mode != InputOrderMode::Linear
&& !self.group_by.is_true_no_grouping()
&& self.group_by.is_single()
}
- /// See comments in `PartialHashAggregateStream` limit optimization section
- fn limit_options_supported_by_hash_stream(&self) -> bool {
- self.limit_options.is_none() || self.is_unordered_unfiltered_group_by_distinct()
- }
-
/// Finds the DataType and SortDirection for this Aggregate, if there is one
pub fn get_minmax_desc(&self) -> Option<(FieldRef, bool)> {
let agg_expr = self.aggr_expr.iter().exactly_one().ok()?;