HappenLee commented on code in PR #68312:
URL: https://github.com/apache/doris/pull/68312#discussion_r4228151968
##########
be/src/exprs/aggregate/aggregate_function_state_union.h:
##########
@@ -107,6 +108,32 @@ class AggregateStateUnion : public
AggregateFunctionNonFinalBase,
size_t align_of_data() const override { return _function->align_of_data();
}
+private:
+ template <bool selected>
+ void add_batch_impl(size_t batch_size, AggregateDataPtr* places, size_t
place_offset,
+ const IColumn** columns, Arena& arena) const {
+ // Some batch implementations read the first serialized row before
entering their loop.
+ if (batch_size == 0) {
+ return;
+ }
+ // Delegate the whole batch so native state columns (e.g. SUM) can
merge directly,
+ // instead of dispatching through the range interface for every input
row.
+ // Generic implementations need aligned scratch states and manage
their construction,
+ // destruction and exception cleanup. Release only this storage after
the batch;
+ // keep passing the caller's arena for allocations that merged states
may retain.
+ // The function can be shared by evaluators, so scratch must not be a
mutable member.
+ Arena scratch_arena;
+ auto* scratch = scratch_arena.aligned_alloc(batch_size *
_function->size_of_data(),
+
_function->align_of_data());
+ if constexpr (selected) {
+ _function->deserialize_and_merge_vec_selected(places,
place_offset, scratch, columns[0],
Review Comment:
After discussion, we have decided not to address this wide-state batch
deserialization issue in this PR. The current batch/selected behavior is
intentionally left unchanged. The two accompanying changes address unused
native-reader scratch allocation and direct single-state finalization; neither
is claimed to fix this finding.
--
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]
