polars_io/parquet/write/
batched_writer.rs

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
use std::io::Write;
use std::sync::Mutex;

use arrow::record_batch::RecordBatch;
use polars_core::prelude::*;
use polars_core::POOL;
use polars_parquet::read::ParquetError;
use polars_parquet::write::{
    array_to_columns, CompressedPage, Compressor, DynIter, DynStreamingIterator, Encoding,
    FallibleStreamingIterator, FileWriter, Page, ParquetType, RowGroupIterColumns,
    SchemaDescriptor, WriteOptions,
};
use rayon::prelude::*;

pub struct BatchedWriter<W: Write> {
    // A mutex so that streaming engine can get concurrent read access to
    // compress pages.
    pub(super) writer: Mutex<FileWriter<W>>,
    pub(super) parquet_schema: SchemaDescriptor,
    pub(super) encodings: Vec<Vec<Encoding>>,
    pub(super) options: WriteOptions,
    pub(super) parallel: bool,
}

impl<W: Write> BatchedWriter<W> {
    pub fn encode_and_compress<'a>(
        &'a self,
        df: &'a DataFrame,
    ) -> impl Iterator<Item = PolarsResult<RowGroupIterColumns<'static, PolarsError>>> + 'a {
        let rb_iter = df.iter_chunks(CompatLevel::newest(), false);
        rb_iter.filter_map(move |batch| match batch.len() {
            0 => None,
            _ => {
                let row_group = create_eager_serializer(
                    batch,
                    self.parquet_schema.fields(),
                    self.encodings.as_ref(),
                    self.options,
                );

                Some(row_group)
            },
        })
    }

    /// Write a batch to the parquet writer.
    ///
    /// # Panics
    /// The caller must ensure the chunks in the given [`DataFrame`] are aligned.
    pub fn write_batch(&mut self, df: &DataFrame) -> PolarsResult<()> {
        let row_group_iter = prepare_rg_iter(
            df,
            &self.parquet_schema,
            &self.encodings,
            self.options,
            self.parallel,
        );
        // Lock before looping so that order is maintained under contention.
        let mut writer = self.writer.lock().unwrap();
        for group in row_group_iter {
            writer.write(group?)?;
        }
        Ok(())
    }

    pub fn get_writer(&self) -> &Mutex<FileWriter<W>> {
        &self.writer
    }

    pub fn write_row_groups(
        &self,
        rgs: Vec<RowGroupIterColumns<'static, PolarsError>>,
    ) -> PolarsResult<()> {
        // Lock before looping so that order is maintained.
        let mut writer = self.writer.lock().unwrap();
        for group in rgs {
            writer.write(group)?;
        }
        Ok(())
    }

    /// Writes the footer of the parquet file. Returns the total size of the file.
    pub fn finish(&self) -> PolarsResult<u64> {
        let mut writer = self.writer.lock().unwrap();
        let size = writer.end(None)?;
        Ok(size)
    }
}

// Note that the df should be rechunked
fn prepare_rg_iter<'a>(
    df: &'a DataFrame,
    parquet_schema: &'a SchemaDescriptor,
    encodings: &'a [Vec<Encoding>],
    options: WriteOptions,
    parallel: bool,
) -> impl Iterator<Item = PolarsResult<RowGroupIterColumns<'static, PolarsError>>> + 'a {
    let rb_iter = df.iter_chunks(CompatLevel::newest(), false);
    rb_iter.filter_map(move |batch| match batch.len() {
        0 => None,
        _ => {
            let row_group =
                create_serializer(batch, parquet_schema.fields(), encodings, options, parallel);

            Some(row_group)
        },
    })
}

fn pages_iter_to_compressor(
    encoded_columns: Vec<DynIter<'static, PolarsResult<Page>>>,
    options: WriteOptions,
) -> Vec<PolarsResult<DynStreamingIterator<'static, CompressedPage, PolarsError>>> {
    encoded_columns
        .into_iter()
        .map(|encoded_pages| {
            // iterator over pages
            let pages = DynStreamingIterator::new(
                Compressor::new_from_vec(
                    encoded_pages.map(|result| {
                        result.map_err(|e| {
                            ParquetError::FeatureNotSupported(format!("reraised in polars: {e}",))
                        })
                    }),
                    options.compression,
                    vec![],
                )
                .map_err(PolarsError::from),
            );

            Ok(pages)
        })
        .collect::<Vec<_>>()
}

fn array_to_pages_iter(
    array: &ArrayRef,
    type_: &ParquetType,
    encoding: &[Encoding],
    options: WriteOptions,
) -> Vec<PolarsResult<DynStreamingIterator<'static, CompressedPage, PolarsError>>> {
    let encoded_columns = array_to_columns(array, type_.clone(), options, encoding).unwrap();
    pages_iter_to_compressor(encoded_columns, options)
}

fn create_serializer(
    batch: RecordBatch,
    fields: &[ParquetType],
    encodings: &[Vec<Encoding>],
    options: WriteOptions,
    parallel: bool,
) -> PolarsResult<RowGroupIterColumns<'static, PolarsError>> {
    let func = move |((array, type_), encoding): ((&ArrayRef, &ParquetType), &Vec<Encoding>)| {
        array_to_pages_iter(array, type_, encoding, options)
    };

    let columns = if parallel {
        POOL.install(|| {
            batch
                .columns()
                .par_iter()
                .zip(fields)
                .zip(encodings)
                .flat_map(func)
                .collect::<Vec<_>>()
        })
    } else {
        batch
            .columns()
            .iter()
            .zip(fields)
            .zip(encodings)
            .flat_map(func)
            .collect::<Vec<_>>()
    };

    let row_group = DynIter::new(columns.into_iter());

    Ok(row_group)
}

/// This serializer encodes and compresses all eagerly in memory.
/// Used for separating compute from IO.
fn create_eager_serializer(
    batch: RecordBatch,
    fields: &[ParquetType],
    encodings: &[Vec<Encoding>],
    options: WriteOptions,
) -> PolarsResult<RowGroupIterColumns<'static, PolarsError>> {
    let func = move |((array, type_), encoding): ((&ArrayRef, &ParquetType), &Vec<Encoding>)| {
        array_to_pages_iter(array, type_, encoding, options)
    };

    let columns = batch
        .columns()
        .iter()
        .zip(fields)
        .zip(encodings)
        .flat_map(func)
        .collect::<Vec<_>>();

    let row_group = DynIter::new(columns.into_iter());

    Ok(row_group)
}