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

// SPDX-License-Identifier: Apache-2.0
// SPDX-FileCopyrightText: Copyright The Lance Authors

use std::collections::VecDeque;
use std::pin::Pin;

use arrow::compute::kernels;
use arrow_array::RecordBatch;
use datafusion::physical_plan::{stream::RecordBatchStreamAdapter, SendableRecordBatchStream};
use datafusion_common::DataFusionError;
use futures::{Stream, StreamExt, TryStreamExt};

use lance_core::Result;

/// Wraps a [`SendableRecordBatchStream`] into a stream of RecordBatch chunks of
/// a given size.  This slices but does not copy any buffers.
struct BatchReaderChunker {
    /// The inner stream
    inner: SendableRecordBatchStream,
    /// The batches that have been read from the inner stream but not yet fully yielded
    buffered: VecDeque<RecordBatch>,
    /// The number of rows to yield in each chunk
    output_size: usize,
    /// The position within the first batch in the buffer to start yielding from
    i: usize,
}

impl BatchReaderChunker {
    fn new(inner: SendableRecordBatchStream, output_size: usize) -> Self {
        Self {
            inner,
            buffered: VecDeque::new(),
            output_size,
            i: 0,
        }
    }

    fn buffered_len(&self) -> usize {
        let buffer_total: usize = self.buffered.iter().map(|batch| batch.num_rows()).sum();
        buffer_total - self.i
    }

    async fn fill_buffer(&mut self) -> Result<()> {
        while self.buffered_len() < self.output_size {
            match self.inner.next().await {
                Some(Ok(batch)) => self.buffered.push_back(batch),
                Some(Err(e)) => return Err(e.into()),
                None => break,
            }
        }
        Ok(())
    }

    async fn next(&mut self) -> Option<Result<Vec<RecordBatch>>> {
        match self.fill_buffer().await {
            Ok(_) => {}
            Err(e) => return Some(Err(e)),
        };

        let mut batches = Vec::new();

        let mut rows_collected = 0;

        while rows_collected < self.output_size {
            if let Some(batch) = self.buffered.pop_front() {
                // Skip empty batch
                if batch.num_rows() == 0 {
                    continue;
                }

                let rows_remaining_in_batch = batch.num_rows() - self.i;
                let rows_to_take =
                    std::cmp::min(rows_remaining_in_batch, self.output_size - rows_collected);

                if rows_to_take == rows_remaining_in_batch {
                    // We're taking the whole batch, so we can just move it
                    let batch = if self.i == 0 {
                        batch
                    } else {
                        // We are taking the remainder of the batch, so we need to slice it
                        batch.slice(self.i, rows_to_take)
                    };
                    batches.push(batch);
                    self.i = 0;
                } else {
                    // We're taking a slice of the batch, so we need to copy it
                    batches.push(batch.slice(self.i, rows_to_take));
                    // And then we need to push the remainder back onto the front of the queue
                    self.i += rows_to_take;
                    self.buffered.push_front(batch);
                }

                rows_collected += rows_to_take;
            } else {
                break;
            }
        }

        if batches.is_empty() {
            None
        } else {
            Some(Ok(batches))
        }
    }
}

pub fn chunk_stream(
    stream: SendableRecordBatchStream,
    chunk_size: usize,
) -> Pin<Box<dyn Stream<Item = Result<Vec<RecordBatch>>> + Send>> {
    let chunker = BatchReaderChunker::new(stream, chunk_size);
    futures::stream::unfold(chunker, |mut chunker| async move {
        match chunker.next().await {
            Some(Ok(batches)) => Some((Ok(batches), chunker)),
            Some(Err(e)) => Some((Err(e), chunker)),
            None => None,
        }
    })
    .boxed()
}

pub fn chunk_concat_stream(
    stream: SendableRecordBatchStream,
    chunk_size: usize,
) -> SendableRecordBatchStream {
    let schema = stream.schema().clone();
    let schema_copy = schema.clone();
    let chunked = chunk_stream(stream, chunk_size);
    let chunk_concat = chunked
        .and_then(move |batches| {
            std::future::ready(
                // chunk_stream is zero-copy and so it gives us pieces of batches.  However, the btree
                // index needs 1 batch-per-page and so we concatenate here.
                kernels::concat::concat_batches(&schema, batches.iter()).map_err(|e| e.into()),
            )
        })
        .map_err(DataFusionError::from)
        .boxed();
    Box::pin(RecordBatchStreamAdapter::new(schema_copy, chunk_concat))
}