1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
use std::any::Any;
use std::pin::Pin;
use std::sync::Arc;
use std::task::{Context, Poll};
use crate::error::{DataFusionError, Result};
use crate::physical_plan::{
ExecutionPlan, Partitioning, RecordBatchStream, SendableRecordBatchStream,
};
use arrow::compute::kernels::concat::concat;
use arrow::datatypes::SchemaRef;
use arrow::error::Result as ArrowResult;
use arrow::record_batch::RecordBatch;
use async_trait::async_trait;
use futures::stream::{Stream, StreamExt};
use log::debug;
#[derive(Debug)]
pub struct CoalesceBatchesExec {
input: Arc<dyn ExecutionPlan>,
target_batch_size: usize,
}
impl CoalesceBatchesExec {
pub fn new(input: Arc<dyn ExecutionPlan>, target_batch_size: usize) -> Self {
Self {
input,
target_batch_size,
}
}
pub fn input(&self) -> &Arc<dyn ExecutionPlan> {
&self.input
}
pub fn target_batch_size(&self) -> usize {
self.target_batch_size
}
}
#[async_trait]
impl ExecutionPlan for CoalesceBatchesExec {
fn as_any(&self) -> &dyn Any {
self
}
fn schema(&self) -> SchemaRef {
self.input.schema()
}
fn children(&self) -> Vec<Arc<dyn ExecutionPlan>> {
vec![self.input.clone()]
}
fn output_partitioning(&self) -> Partitioning {
self.input.output_partitioning()
}
fn with_new_children(
&self,
children: Vec<Arc<dyn ExecutionPlan>>,
) -> Result<Arc<dyn ExecutionPlan>> {
match children.len() {
1 => Ok(Arc::new(CoalesceBatchesExec::new(
children[0].clone(),
self.target_batch_size,
))),
_ => Err(DataFusionError::Internal(
"CoalesceBatchesExec wrong number of children".to_string(),
)),
}
}
async fn execute(&self, partition: usize) -> Result<SendableRecordBatchStream> {
Ok(Box::pin(CoalesceBatchesStream {
input: self.input.execute(partition).await?,
schema: self.input.schema(),
target_batch_size: self.target_batch_size,
buffer: Vec::new(),
buffered_rows: 0,
is_closed: false,
}))
}
}
struct CoalesceBatchesStream {
input: SendableRecordBatchStream,
schema: SchemaRef,
target_batch_size: usize,
buffer: Vec<RecordBatch>,
buffered_rows: usize,
is_closed: bool,
}
impl Stream for CoalesceBatchesStream {
type Item = ArrowResult<RecordBatch>;
fn poll_next(
mut self: Pin<&mut Self>,
cx: &mut Context<'_>,
) -> Poll<Option<Self::Item>> {
if self.is_closed {
return Poll::Ready(None);
}
loop {
let input_batch = self.input.poll_next_unpin(cx);
match input_batch {
Poll::Ready(x) => match x {
Some(Ok(ref batch)) => {
if batch.num_rows() >= self.target_batch_size
&& self.buffer.is_empty()
{
return Poll::Ready(Some(Ok(batch.clone())));
} else if batch.num_rows() == 0 {
} else {
self.buffer.push(batch.clone());
self.buffered_rows += batch.num_rows();
if self.buffered_rows >= self.target_batch_size {
let batch = concat_batches(
&self.schema,
&self.buffer,
self.buffered_rows,
)?;
self.buffer.clear();
self.buffered_rows = 0;
return Poll::Ready(Some(Ok(batch)));
}
}
}
None => {
self.is_closed = true;
if self.buffer.is_empty() {
return Poll::Ready(None);
} else {
let batch = concat_batches(
&self.schema,
&self.buffer,
self.buffered_rows,
)?;
self.buffer.clear();
self.buffered_rows = 0;
return Poll::Ready(Some(Ok(batch)));
}
}
other => return Poll::Ready(other),
},
Poll::Pending => return Poll::Pending,
}
}
}
fn size_hint(&self) -> (usize, Option<usize>) {
self.input.size_hint()
}
}
impl RecordBatchStream for CoalesceBatchesStream {
fn schema(&self) -> SchemaRef {
self.schema.clone()
}
}
pub fn concat_batches(
schema: &SchemaRef,
batches: &[RecordBatch],
row_count: usize,
) -> ArrowResult<RecordBatch> {
if batches.is_empty() {
return Ok(RecordBatch::new_empty(schema.clone()));
}
let mut arrays = Vec::with_capacity(schema.fields().len());
for i in 0..schema.fields().len() {
let array = concat(
&batches
.iter()
.map(|batch| batch.column(i).as_ref())
.collect::<Vec<_>>(),
)?;
arrays.push(array);
}
debug!(
"Combined {} batches containing {} rows",
batches.len(),
row_count
);
RecordBatch::try_new(schema.clone(), arrays)
}
#[cfg(test)]
mod tests {
use super::*;
use crate::physical_plan::{memory::MemoryExec, repartition::RepartitionExec};
use arrow::array::UInt32Array;
use arrow::datatypes::{DataType, Field, Schema};
#[tokio::test(flavor = "multi_thread")]
async fn test_concat_batches() -> Result<()> {
let schema = test_schema();
let partition = create_vec_batches(&schema, 10);
let partitions = vec![partition];
let output_partitions = coalesce_batches(&schema, partitions, 21).await?;
assert_eq!(1, output_partitions.len());
let batches = &output_partitions[0];
assert_eq!(4, batches.len());
assert_eq!(24, batches[0].num_rows());
assert_eq!(24, batches[1].num_rows());
assert_eq!(24, batches[2].num_rows());
assert_eq!(8, batches[3].num_rows());
Ok(())
}
fn test_schema() -> Arc<Schema> {
Arc::new(Schema::new(vec![Field::new("c0", DataType::UInt32, false)]))
}
fn create_vec_batches(schema: &Arc<Schema>, num_batches: usize) -> Vec<RecordBatch> {
let batch = create_batch(schema);
let mut vec = Vec::with_capacity(num_batches);
for _ in 0..num_batches {
vec.push(batch.clone());
}
vec
}
fn create_batch(schema: &Arc<Schema>) -> RecordBatch {
RecordBatch::try_new(
schema.clone(),
vec![Arc::new(UInt32Array::from(vec![1, 2, 3, 4, 5, 6, 7, 8]))],
)
.unwrap()
}
async fn coalesce_batches(
schema: &SchemaRef,
input_partitions: Vec<Vec<RecordBatch>>,
target_batch_size: usize,
) -> Result<Vec<Vec<RecordBatch>>> {
let exec = MemoryExec::try_new(&input_partitions, schema.clone(), None)?;
let exec =
RepartitionExec::try_new(Arc::new(exec), Partitioning::RoundRobinBatch(1))?;
let exec: Arc<dyn ExecutionPlan> =
Arc::new(CoalesceBatchesExec::new(Arc::new(exec), target_batch_size));
let output_partition_count = exec.output_partitioning().partition_count();
let mut output_partitions = Vec::with_capacity(output_partition_count);
for i in 0..output_partition_count {
let mut stream = exec.execute(i).await?;
let mut batches = vec![];
while let Some(result) = stream.next().await {
batches.push(result?);
}
output_partitions.push(batches);
}
Ok(output_partitions)
}
}