Skip to content

Apache DataFusion SQL Query Engine

perf: Optimize heap handling in TopK operator (#20556)

Adam Gutglicka026e7da2fe8
diff --git a/datafusion/physical-plan/src/topk/mod.rs b/datafusion/physical-plan/src/topk/mod.rs
index 4b93e6a188..e0b91f2516 100644
--- a/datafusion/physical-plan/src/topk/mod.rs
+++ b/datafusion/physical-plan/src/topk/mod.rs
@@ -724,8 +724,8 @@ impl TopKHeap {
let row = row.as_ref();
// Reuse storage for evicted item if possible
- let new_top_k = if self.inner.len() == self.k {
- let prev_min = self.inner.pop().unwrap();
+ if self.inner.len() == self.k {
+ let mut prev_min = self.inner.peek_mut().unwrap();
// Update batch use
if prev_min.batch_id == batch_entry.id {
@@ -736,15 +736,16 @@ impl TopKHeap {
// update memory accounting
self.owned_bytes -= prev_min.owned_size();
- prev_min.with_new_row(row, batch_id, index)
- } else {
- TopKRow::new(row, batch_id, index)
- };
- self.owned_bytes += new_top_k.owned_size();
+ prev_min.replace_with(row, batch_id, index);
- // put the new row into the heap
- self.inner.push(new_top_k)
+ self.owned_bytes += prev_min.owned_size();
+ } else {
+ let new_row = TopKRow::new(row, batch_id, index);
+ self.owned_bytes += new_row.owned_size();
+ // put the new row into the heap
+ self.inner.push(new_row);
+ };
}
/// Returns the values stored in this heap, from values low to
@@ -911,26 +912,13 @@ impl TopKRow {
}
}
- /// Create a new TopKRow reusing the existing allocation
- fn with_new_row(
- self,
- new_row: impl AsRef<[u8]>,
- batch_id: u32,
- index: usize,
- ) -> Self {
- let Self {
- mut row,
- batch_id: _,
- index: _,
- } = self;
- row.clear();
- row.extend_from_slice(new_row.as_ref());
+ // Replace the existing row capacity with new values
+ fn replace_with(&mut self, new_row: impl AsRef<[u8]>, batch_id: u32, index: usize) {
+ self.row.clear();
+ self.row.extend_from_slice(new_row.as_ref());
- Self {
- row,
- batch_id,
- index,
- }
+ self.batch_id = batch_id;
+ self.index = index;
}
/// Returns the number of bytes owned by this row in the heap (not
diff --git a/datafusion/sqllogictest/test_files/limit.slt b/datafusion/sqllogictest/test_files/limit.slt
index ec8363f51a..ff3c49485a 100644
--- a/datafusion/sqllogictest/test_files/limit.slt
+++ b/datafusion/sqllogictest/test_files/limit.slt
@@ -679,19 +679,19 @@ ON t1.b = t2.b
ORDER BY t1.b desc, c desc, c2 desc;
----
3 98 96
-3 98 89
+3 98 87
3 98 82
3 98 79
3 97 96
-3 97 89
+3 97 87
3 97 82
3 97 79
3 96 96
-3 96 89
+3 96 87
3 96 82
3 96 79
3 95 96
-3 95 89
+3 95 87
3 95 82
3 95 79
diff --git a/datafusion/sqllogictest/test_files/window.slt b/datafusion/sqllogictest/test_files/window.slt
index 8ac8724683..c3e6f39adb 100644
--- a/datafusion/sqllogictest/test_files/window.slt
+++ b/datafusion/sqllogictest/test_files/window.slt
@@ -4387,9 +4387,9 @@ LIMIT 5;
----
78 50
63 38
-3 53
+NULL 19
24 31
-14 94
+24 56
# result should be same with above, when LAG/LEAD algorithm work with pruned data.
# decreasing batch size, causes data to be produced in smaller chunks at the source.
@@ -4406,9 +4406,9 @@ LIMIT 5;
----
78 50
63 38
-3 53
+NULL 19
24 31
-14 94
+24 56
statement ok
set datafusion.execution.batch_size = 100;