perf: improve performance of array_union/array_intersect with batched row conversion (apache#20243)

lyne7-sc · web-flow · commit 4e2c0f1feaef · 2026-02-11T16:14:47.000Z
## Which issue does this PR close?  - Closes #. ## Rationale for this change  The current implementation of `array_union` and `array_intersect` performs `RowConverter::convert_columns()` on a per-row basis, which introduces avoidable overhead due to repeated conversions and intermediate allocations. This PR improves performance by: 1. converting all list values to rows in a batch 2. reusing hash sets across iterations 3. removing the `sorted().dedup()` pattern in favor of hash-based set operations ## What changes are included in this PR? Refactored the internal set operation implementation to use batch row conversion and a single-pass construction of result arrays. ### Benchmarks ``` group before optimized ----- ------ --------- array_intersect/high_overlap/10 2.99 1442.0±99.94µs ? ?/sec 1.00 481.6±21.45µs ? ?/sec array_intersect/high_overlap/100 1.90 9.5±0.63ms ? ?/sec 1.00 5.0±0.09ms ? ?/sec array_intersect/high_overlap/50 2.01 5.3±0.41ms ? ?/sec 1.00 2.6±0.05ms ? ?/sec array_intersect/low_overlap/10 3.47 1288.1±72.39µs ? ?/sec 1.00 371.4±14.08µs ? ?/sec array_intersect/low_overlap/100 2.35 9.2±0.43ms ? ?/sec 1.00 3.9±0.08ms ? ?/sec array_intersect/low_overlap/50 2.45 5.1±0.41ms ? ?/sec 1.00 2.1±0.07ms ? ?/sec array_union/high_overlap/10 4.01 1593.1±292.17µs ? ?/sec 1.00 396.9±13.43µs ? ?/sec array_union/high_overlap/100 2.54 9.8±0.18ms ? ?/sec 1.00 3.9±0.11ms ? ?/sec array_union/high_overlap/50 2.65 5.4±0.10ms ? ?/sec 1.00 2.0±0.07ms ? ?/sec array_union/low_overlap/10 3.74 1622.7±96.50µs ? ?/sec 1.00 434.3±17.87µs ? ?/sec array_union/low_overlap/100 2.39 10.3±0.92ms ? ?/sec 1.00 4.3±0.11ms ? ?/sec array_union/low_overlap/50 2.63 5.8±0.27ms ? ?/sec 1.00 2.2±0.11ms ? ?/sec ```  ## Are these changes tested? Yes. Existing SQL logic tests updated to reflect new output order.  ## Are there any user-facing changes? Yes. The output order may differ from the previous implementation. Previously, results were implicitly sorted due to the use of `sorted().dedup()`. The new implementation preserves the order of first appearance within each list. This is a user-visible behavioral change, but it is consistent with typical SQL set operation semantics, which do not guarantee a specific output order.
diff --git a/datafusion/functions-nested/Cargo.toml b/datafusion/functions-nested/Cargo.toml
@@ -92,3 +92,7 @@ name = "array_remove"
 [[bench]]
 harness = false
 name = "array_repeat"
+
+[[bench]]
+harness = false
+name = "array_set_ops"
diff --git a/datafusion/functions-nested/benches/array_set_ops.rs b/datafusion/functions-nested/benches/array_set_ops.rs
@@ -0,0 +1,169 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#[macro_use]
+extern crate criterion;
+
+use arrow::array::{ArrayRef, Int64Array, ListArray};
+use arrow::buffer::OffsetBuffer;
+use arrow::datatypes::{DataType, Field};
+use criterion::{BenchmarkId, Criterion};
+use datafusion_common::config::ConfigOptions;
+use datafusion_expr::{ColumnarValue, ScalarFunctionArgs, ScalarUDFImpl};
+use datafusion_functions_nested::set_ops::{ArrayIntersect, ArrayUnion};
+use rand::SeedableRng;
+use rand::prelude::SliceRandom;
+use rand::rngs::StdRng;
+use std::collections::HashSet;
+use std::hint::black_box;
+use std::sync::Arc;
+
+const NUM_ROWS: usize = 1000;
+const ARRAY_SIZES: &[usize] = &[10, 50, 100];
+const SEED: u64 = 42;
+
+fn criterion_benchmark(c: &mut Criterion) {
+    bench_array_union(c);
+    bench_array_intersect(c);
+}
+
+fn invoke_udf(udf: &impl ScalarUDFImpl, array1: &ArrayRef, array2: &ArrayRef) {
+    black_box(
+        udf.invoke_with_args(ScalarFunctionArgs {
+            args: vec![
+                ColumnarValue::Array(array1.clone()),
+                ColumnarValue::Array(array2.clone()),
+            ],
+            arg_fields: vec![
+                Field::new("arr1", array1.data_type().clone(), false).into(),
+                Field::new("arr2", array2.data_type().clone(), false).into(),
+            ],
+            number_rows: NUM_ROWS,
+            return_field: Field::new("result", array1.data_type().clone(), false).into(),
+            config_options: Arc::new(ConfigOptions::default()),
+        })
+        .unwrap(),
+    );
+}
+
+fn bench_array_union(c: &mut Criterion) {
+    let mut group = c.benchmark_group("array_union");
+    let udf = ArrayUnion::new();
+
+    for (overlap_label, overlap_ratio) in &[("high_overlap", 0.8), ("low_overlap", 0.2)] {
+        for &array_size in ARRAY_SIZES {
+            let (array1, array2) =
+                create_arrays_with_overlap(NUM_ROWS, array_size, *overlap_ratio);
+            group.bench_with_input(
+                BenchmarkId::new(*overlap_label, array_size),
+                &array_size,
+                |b, _| b.iter(|| invoke_udf(&udf, &array1, &array2)),
+            );
+        }
+    }
+
+    group.finish();
+}
+
+fn bench_array_intersect(c: &mut Criterion) {
+    let mut group = c.benchmark_group("array_intersect");
+    let udf = ArrayIntersect::new();
+
+    for (overlap_label, overlap_ratio) in &[("high_overlap", 0.8), ("low_overlap", 0.2)] {
+        for &array_size in ARRAY_SIZES {
+            let (array1, array2) =
+                create_arrays_with_overlap(NUM_ROWS, array_size, *overlap_ratio);
+            group.bench_with_input(
+                BenchmarkId::new(*overlap_label, array_size),
+                &array_size,
+                |b, _| b.iter(|| invoke_udf(&udf, &array1, &array2)),
+            );
+        }
+    }
+
+    group.finish();
+}
+
+fn create_arrays_with_overlap(
+    num_rows: usize,
+    array_size: usize,
+    overlap_ratio: f64,
+) -> (ArrayRef, ArrayRef) {
+    assert!((0.0..=1.0).contains(&overlap_ratio));
+    let overlap_count = ((array_size as f64) * overlap_ratio).round() as usize;
+
+    let mut rng = StdRng::seed_from_u64(SEED);
+
+    let mut values1 = Vec::with_capacity(num_rows * array_size);
+    let mut values2 = Vec::with_capacity(num_rows * array_size);
+
+    for row in 0..num_rows {
+        let base = (row as i64) * (array_size as i64) * 2;
+
+        for i in 0..array_size {
+            values1.push(base + i as i64);
+        }
+
+        let mut positions: Vec<usize> = (0..array_size).collect();
+        positions.shuffle(&mut rng);
+
+        let overlap_positions: HashSet<_> =
+            positions[..overlap_count].iter().copied().collect();
+
+        for i in 0..array_size {
+            if overlap_positions.contains(&i) {
+                values2.push(base + i as i64);
+            } else {
+                values2.push(base + array_size as i64 + i as i64);
+            }
+        }
+    }
+
+    let values1 = Int64Array::from(values1);
+    let values2 = Int64Array::from(values2);
+
+    let field = Arc::new(Field::new("item", DataType::Int64, true));
+
+    let offsets = (0..=num_rows)
+        .map(|i| (i * array_size) as i32)
+        .collect::<Vec<i32>>();
+
+    let array1 = Arc::new(
+        ListArray::try_new(
+            field.clone(),
+            OffsetBuffer::new(offsets.clone().into()),
+            Arc::new(values1),
+            None,
+        )
+        .unwrap(),
+    );
+
+    let array2 = Arc::new(
+        ListArray::try_new(
+            field,
+            OffsetBuffer::new(offsets.into()),
+            Arc::new(values2),
+            None,
+        )
+        .unwrap(),
+    );
+
+    (array1, array2)
+}
+
+criterion_group!(benches, criterion_benchmark);
+criterion_main!(benches);
diff --git a/datafusion/functions-nested/src/set_ops.rs b/datafusion/functions-nested/src/set_ops.rs
@@ -184,11 +184,17 @@ impl ScalarUDFImpl for ArrayUnion {
     )
 )]
 #[derive(Debug, PartialEq, Eq, Hash)]
-pub(super) struct ArrayIntersect {
+pub struct ArrayIntersect {
     signature: Signature,
     aliases: Vec<String>,
 }
 
+impl Default for ArrayIntersect {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
 impl ArrayIntersect {
     pub fn new() -> Self {
         Self {
@@ -358,69 +364,117 @@ fn generic_set_lists<OffsetSize: OffsetSizeTrait>(
         "{set_op:?} is not implemented for '{l:?}' and '{r:?}'"
     );
 
-    let mut offsets = vec![OffsetSize::usize_as(0)];
-    let mut new_arrays = vec![];
+    // Convert all values to rows in batch for performance.
     let converter = RowConverter::new(vec![SortField::new(l.value_type())])?;
-    for (l_arr, r_arr) in l.iter().zip(r.iter()) {
-        let last_offset = *offsets.last().unwrap();
+    let rows_l = converter.convert_columns(&[Arc::clone(l.values())])?;
+    let rows_r = converter.convert_columns(&[Arc::clone(r.values())])?;
 
-        let (l_values, r_values) = match (l_arr, r_arr) {
-            (Some(l_arr), Some(r_arr)) => (
-                converter.convert_columns(&[l_arr])?,
-                converter.convert_columns(&[r_arr])?,
-            ),
-            _ => {
-                offsets.push(last_offset);
-                continue;
-            }
-        };
+    match set_op {
+        SetOp::Union => generic_set_loop::<OffsetSize, true>(
+            l, r, &rows_l, &rows_r, field, &converter,
+        ),
+        SetOp::Intersect => generic_set_loop::<OffsetSize, false>(
+            l, r, &rows_l, &rows_r, field, &converter,
+        ),
+    }
+}
 
-        let l_iter = l_values.iter().sorted().dedup();
-        let values_set: HashSet<_> = l_iter.clone().collect();
-        let mut rows = if set_op == SetOp::Union {
-            l_iter.collect()
-        } else {
-            vec![]
-        };
+/// Inner loop for set operations, parameterized by const generic to
+/// avoid branching inside the hot loop.
+fn generic_set_loop<OffsetSize: OffsetSizeTrait, const IS_UNION: bool>(
+    l: &GenericListArray<OffsetSize>,
+    r: &GenericListArray<OffsetSize>,
+    rows_l: &arrow::row::Rows,
+    rows_r: &arrow::row::Rows,
+    field: Arc<Field>,
+    converter: &RowConverter,
+) -> Result<ArrayRef> {
+    let l_offsets = l.value_offsets();
+    let r_offsets = r.value_offsets();
+
+    let mut result_offsets = Vec::with_capacity(l.len() + 1);
+    result_offsets.push(OffsetSize::usize_as(0));
+    let initial_capacity = if IS_UNION {
+        // Union can include all elements from both sides
+        rows_l.num_rows()
+    } else {
+        // Intersect result is bounded by the smaller side
+        rows_l.num_rows().min(rows_r.num_rows())
+    };
+
+    let mut final_rows = Vec::with_capacity(initial_capacity);
+
+    // Reuse hash sets across iterations
+    let mut seen = HashSet::new();
+    let mut lookup_set = HashSet::new();
+    for i in 0..l.len() {
+        let last_offset = *result_offsets.last().unwrap();
 
-        for r_val in r_values.iter().sorted().dedup() {
-            match set_op {
-                SetOp::Union => {
-                    if !values_set.contains(&r_val) {
-                        rows.push(r_val);
-                    }
+        if l.is_null(i) || r.is_null(i) {
+            result_offsets.push(last_offset);
+            continue;
+        }
+
+        let l_start = l_offsets[i].as_usize();
+        let l_end = l_offsets[i + 1].as_usize();
+        let r_start = r_offsets[i].as_usize();
+        let r_end = r_offsets[i + 1].as_usize();
+
+        seen.clear();
+
+        if IS_UNION {
+            for idx in l_start..l_end {
+                let row = rows_l.row(idx);
+                if seen.insert(row) {
+                    final_rows.push(row);
                 }
-                SetOp::Intersect => {
-                    if values_set.contains(&r_val) {
-                        rows.push(r_val);
-                    }
+            }
+            for idx in r_start..r_end {
+                let row = rows_r.row(idx);
+                if seen.insert(row) {
+                    final_rows.push(row);
                 }
             }
-        }
-
-        offsets.push(last_offset + OffsetSize::usize_as(rows.len()));
-        let arrays = converter.convert_rows(rows)?;
-        let array = match arrays.first() {
-            Some(array) => Arc::clone(array),
-            None => {
-                return internal_err!("{set_op}: failed to get array from rows");
+        } else {
+            let l_len = l_end - l_start;
+            let r_len = r_end - r_start;
+
+            // Select shorter side for lookup, longer side for probing
+            let (lookup_rows, lookup_range, probe_rows, probe_range) = if l_len < r_len {
+                (rows_l, l_start..l_end, rows_r, r_start..r_end)
+            } else {
+                (rows_r, r_start..r_end, rows_l, l_start..l_end)
+            };
+            lookup_set.clear();
+            lookup_set.reserve(lookup_range.len());
+
+            // Build lookup table
+            for idx in lookup_range {
+                lookup_set.insert(lookup_rows.row(idx));
             }
-        };
 
-        new_arrays.push(array);
+            // Probe and emit distinct intersected rows
+            for idx in probe_range {
+                let row = probe_rows.row(idx);
+                if lookup_set.contains(&row) && seen.insert(row) {
+                    final_rows.push(row);
+                }
+            }
+        }
+        result_offsets.push(last_offset + OffsetSize::usize_as(seen.len()));
     }
 
-    let offsets = OffsetBuffer::new(offsets.into());
-    let new_arrays_ref: Vec<_> = new_arrays.iter().map(|v| v.as_ref()).collect();
-    let values = if new_arrays_ref.is_empty() {
+    let final_values = if final_rows.is_empty() {
         new_empty_array(&l.value_type())
     } else {
-        compute::concat(&new_arrays_ref)?
+        let arrays = converter.convert_rows(final_rows)?;
+        Arc::clone(&arrays[0])
     };
+
     let arr = GenericListArray::<OffsetSize>::try_new(
         field,
-        offsets,
-        values,
+        OffsetBuffer::new(result_offsets.into()),
+        final_values,
         NullBuffer::union(l.nulls(), r.nulls()),
     )?;
     Ok(Arc::new(arr))
diff --git a/datafusion/sqllogictest/test_files/array.slt b/datafusion/sqllogictest/test_files/array.slt
