perf: Improve materialisation performance of SortPreservingMergeExec (#…

…691)

* test: add benchmarks for SortPreservingMergeExec

* perf: minimise array data extend calls

The `SortPreservingMergeStream` operator merges two streams together by creating an output record batch that is build from the contents of the input. Previously each row of input would be pushed into the output sink even if though the API supports pushing batches of rows.

This commit implements the logic to push batches of rows from inputs where possible.

Performance benchmarks show an improvement of between 3-12%.

```
group                               master                                 pr
-----                               ------                                 --
interleave_batches                  1.04   637.5±51.84µs        ? ?/sec    1.00   615.5±12.13µs        ? ?/sec
merge_batches_no_overlap_large      1.12    454.9±2.90µs        ? ?/sec    1.00   404.9±10.94µs        ? ?/sec
merge_batches_no_overlap_small      1.14    485.1±6.67µs        ? ?/sec    1.00    425.7±9.33µs        ? ?/sec
merge_batches_small_into_large      1.14    263.0±8.85µs        ? ?/sec    1.00    229.7±5.23µs        ? ?/sec
merge_batches_some_overlap_large    1.05    532.5±8.33µs        ? ?/sec    1.00   508.3±14.24µs        ? ?/sec
merge_batches_some_overlap_small    1.06   546.9±12.82µs        ? ?/sec    1.00   516.9±13.20µs        ? ?/sec
```

* test: more test coverage

* refactor: update batch size

datafusion/Cargo.toml

@@ -95,3 +95,7 @@ harness = false
 [[bench]]
 name = "scalar"
 harness = false
+
+[[bench]]
+name = "physical_plan"
+harness = false

datafusion/benches/physical_plan.rs

@@ -0,0 +1,176 @@
+// 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 criterion::{BatchSize, Criterion};
+extern crate arrow;
+extern crate datafusion;
+
+use std::{iter::FromIterator, sync::Arc};
+
+use arrow::{
+    array::{ArrayRef, Int64Array, StringArray},
+    record_batch::RecordBatch,
+};
+use tokio::runtime::Runtime;
+
+use datafusion::physical_plan::{
+    collect,
+    expressions::{col, PhysicalSortExpr},
+    memory::MemoryExec,
+    sort_preserving_merge::SortPreservingMergeExec,
+};
+
+// Initialise the operator using the provided record batches and the sort key
+// as inputs. All record batches must have the same schema.
+fn sort_preserving_merge_operator(batches: Vec<RecordBatch>, sort: &[&str]) {
+    let schema = batches[0].schema();
+
+    let sort = sort
+        .iter()
+        .map(|name| PhysicalSortExpr {
+            expr: col(name, &schema).unwrap(),
+            options: Default::default(),
+        })
+        .collect::<Vec<_>>();
+
+    let exec = MemoryExec::try_new(
+        &batches.into_iter().map(|rb| vec![rb]).collect::<Vec<_>>(),
+        schema,
+        None,
+    )
+    .unwrap();
+    let merge = Arc::new(SortPreservingMergeExec::new(sort, Arc::new(exec), 8192));
+
+    let rt = Runtime::new().unwrap();
+    rt.block_on(collect(merge)).unwrap();
+}
+
+// Produces `n` record batches of row size `m`. Each record batch will have
+// identical contents except for if the `batch_offset` is set. In that case the
+// values for column "d" in each subsequent record batch will be offset in
+// value.
+//
+// The `rows_per_key` value controls how many rows are generated per "key",
+// which is defined as columns a, b and c.
+fn batches(
+    n: usize,
+    m: usize,
+    rows_per_sort_key: usize,
+    batch_offset: usize,
+) -> Vec<RecordBatch> {
+    let mut rbs = Vec::with_capacity(n);
+    let mut curr_batch_offset = 0;
+
+    for _ in 0..n {
+        let mut col_a = Vec::with_capacity(m);
+        let mut col_b = Vec::with_capacity(m);
+        let mut col_c = Vec::with_capacity(m);
+        let mut col_d = Vec::with_capacity(m);
+
+        let mut j = 0;
+        let mut current_rows_per_sort_key = 0;
+
+        for i in 0..m {
+            if current_rows_per_sort_key == rows_per_sort_key {
+                current_rows_per_sort_key = 0;
+                j = i;
+            }
+
+            col_a.push(Some(format!("a-{:?}", j)));
+            col_b.push(Some(format!("b-{:?}", j)));
+            col_c.push(Some(format!("c-{:?}", j)));
+            col_d.push(Some((i + curr_batch_offset) as i64));
+
+            current_rows_per_sort_key += 1;
+        }
+
+        col_a.sort();
+        col_b.sort();
+        col_c.sort();
+
+        let col_a: ArrayRef = Arc::new(StringArray::from_iter(col_a));
+        let col_b: ArrayRef = Arc::new(StringArray::from_iter(col_b));
+        let col_c: ArrayRef = Arc::new(StringArray::from_iter(col_c));
+        let col_d: ArrayRef = Arc::new(Int64Array::from(col_d));
+
+        let rb = RecordBatch::try_from_iter(vec![
+            ("a", col_a),
+            ("b", col_b),
+            ("c", col_c),
+            ("d", col_d),
+        ])
+        .unwrap();
+        rbs.push(rb);
+
+        curr_batch_offset += batch_offset;
+    }
+
+    rbs
+}
+
+fn criterion_benchmark(c: &mut Criterion) {
+    let small_batch = batches(1, 100, 10, 0).remove(0);
+    let large_batch = batches(1, 1000, 1, 0).remove(0);
+
+    let benches = vec![
+        // Two batches with identical rows. They will need to be merged together
+        // with one row from each batch being taken until both batches are
+        // drained.
+        ("interleave_batches", batches(2, 1000, 10, 1)),
+        // Two batches with a small overlapping region of rows for each unique
+        // sort key.
+        ("merge_batches_some_overlap_small", batches(2, 1000, 10, 5)),
+        // Two batches with a large overlapping region of rows for each unique
+        // sort key.
+        (
+            "merge_batches_some_overlap_large",
+            batches(2, 1000, 250, 125),
+        ),
+        // Two batches with no overlapping region of rows for each unique
+        // sort key. For a given unique sort key all rows are drained from one
+        // batch, then all the rows for the same key from the second batch.
+        // This repeats until all rows are drained. There are a small number of
+        // rows (10) for each unique sort key.
+        ("merge_batches_no_overlap_small", batches(2, 1000, 10, 12)),
+        // As above but this time there are a larger number of rows (250) for
+        // each unique sort key - still no overlaps.
+        ("merge_batches_no_overlap_large", batches(2, 1000, 250, 252)),
+        // Merges two batches where one batch is significantly larger than the
+        // other.
+        (
+            "merge_batches_small_into_large",
+            vec![large_batch, small_batch],
+        ),
+    ];
+
+    for (name, input) in benches {
+        c.bench_function(name, move |b| {
+            b.iter_batched(
+                || input.clone(),
+                |input| {
+                    sort_preserving_merge_operator(input, &["a", "b", "c", "d"]);
+                },
+                BatchSize::LargeInput,
+            )
+        });
+    }
+}
+
+criterion_group!(benches, criterion_benchmark);
+criterion_main!(benches);