parquet/arrow/async_writer/
mod.rs

1// Licensed to the Apache Software Foundation (ASF) under one
2// or more contributor license agreements.  See the NOTICE file
3// distributed with this work for additional information
4// regarding copyright ownership.  The ASF licenses this file
5// to you under the Apache License, Version 2.0 (the
6// "License"); you may not use this file except in compliance
7// with the License.  You may obtain a copy of the License at
8//
9//   http://www.apache.org/licenses/LICENSE-2.0
10//
11// Unless required by applicable law or agreed to in writing,
12// software distributed under the License is distributed on an
13// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
14// KIND, either express or implied.  See the License for the
15// specific language governing permissions and limitations
16// under the License.
17
18//! `async` API for writing [`RecordBatch`]es to Parquet files
19//!
20//! See the [crate-level documentation](crate) for more details.
21//!
22//! The `async` API for writing [`RecordBatch`]es is
23//! similar to the [`sync` API](ArrowWriter), so please
24//! read the documentation there before using this API.
25//!
26//! Here is an example for using [`AsyncArrowWriter`]:
27//!
28//! ```
29//! # #[tokio::main(flavor="current_thread")]
30//! # async fn main() {
31//! #
32//! # use std::sync::Arc;
33//! # use arrow_array::{ArrayRef, Int64Array, RecordBatch, RecordBatchReader};
34//! # use bytes::Bytes;
35//! # use parquet::arrow::{AsyncArrowWriter, arrow_reader::ParquetRecordBatchReaderBuilder};
36//! #
37//! let col = Arc::new(Int64Array::from_iter_values([1, 2, 3])) as ArrayRef;
38//! let to_write = RecordBatch::try_from_iter([("col", col)]).unwrap();
39//!
40//! let mut buffer = Vec::new();
41//! let mut writer = AsyncArrowWriter::try_new(&mut buffer, to_write.schema(), None).unwrap();
42//! writer.write(&to_write).await.unwrap();
43//! writer.close().await.unwrap();
44//!
45//! let buffer = Bytes::from(buffer);
46//! let mut reader = ParquetRecordBatchReaderBuilder::try_new(buffer.clone())
47//!     .unwrap()
48//!     .build()
49//!     .unwrap();
50//! let read = reader.next().unwrap().unwrap();
51//!
52//! assert_eq!(to_write, read);
53//! # }
54//! ```
55//!
56//! There is a blanket implementation of [`AsyncFileWriter`] for all types that
57//! implement [`AsyncWrite`], so writers such as `tokio::fs::File`, or
58//! `object_store::buffered::BufWriter` for writing to object storage, can be
59//! passed to [`AsyncArrowWriter`] directly.
60
61#[cfg(feature = "object_store")]
62mod store;
63#[cfg(feature = "object_store")]
64pub use store::*;
65
66use crate::{
67    arrow::ArrowWriter,
68    arrow::arrow_writer::ArrowWriterOptions,
69    errors::{ParquetError, Result},
70    file::{
71        metadata::{KeyValue, ParquetMetaData, RowGroupMetaData},
72        properties::WriterProperties,
73    },
74};
75use arrow_array::RecordBatch;
76use arrow_schema::SchemaRef;
77use bytes::Bytes;
78use futures::FutureExt;
79use futures::future::BoxFuture;
80use std::mem;
81use tokio::io::{AsyncWrite, AsyncWriteExt};
82
83/// The asynchronous interface used by [`AsyncArrowWriter`] to write parquet files.
84pub trait AsyncFileWriter: Send {
85    /// Write the provided bytes to the underlying writer
86    ///
87    /// The underlying writer CAN decide to buffer the data or write it immediately.
88    /// This design allows the writer implementer to control the buffering and I/O scheduling.
89    ///
90    /// The underlying writer MAY implement retry logic to prevent breaking users write process.
91    fn write(&mut self, bs: Bytes) -> BoxFuture<'_, Result<()>>;
92
93    /// Flush any buffered data to the underlying writer and finish writing process.
94    ///
95    /// After `complete` returns `Ok(())`, caller SHOULD not call write again.
96    fn complete(&mut self) -> BoxFuture<'_, Result<()>>;
97}
98
99impl AsyncFileWriter for Box<dyn AsyncFileWriter + '_> {
100    fn write(&mut self, bs: Bytes) -> BoxFuture<'_, Result<()>> {
101        self.as_mut().write(bs)
102    }
103
104    fn complete(&mut self) -> BoxFuture<'_, Result<()>> {
105        self.as_mut().complete()
106    }
107}
108
109impl<T: AsyncWrite + Unpin + Send> AsyncFileWriter for T {
110    fn write(&mut self, bs: Bytes) -> BoxFuture<'_, Result<()>> {
111        async move {
112            self.write_all(&bs).await?;
113            Ok(())
114        }
115        .boxed()
116    }
117
118    fn complete(&mut self) -> BoxFuture<'_, Result<()>> {
119        async move {
120            self.flush().await?;
121            self.shutdown().await?;
122            Ok(())
123        }
124        .boxed()
125    }
126}
127
128/// Encodes [`RecordBatch`] to parquet, outputting to an [`AsyncFileWriter`]
129///
130/// ## Memory Usage
131///
132/// This writer eagerly writes data as soon as possible to the underlying [`AsyncFileWriter`],
133/// permitting fine-grained control over buffering and I/O scheduling. However, the columnar
134/// nature of parquet forces data for an entire row group to be buffered in memory, before
135/// it can be flushed. Depending on the data and the configured row group size, this buffering
136/// may be substantial.
137///
138/// Memory usage can be limited by calling [`Self::flush`] to flush the in progress row group,
139/// although this will likely increase overall file size and reduce query performance.
140/// See [ArrowWriter] for more information.
141///
142/// ```no_run
143/// # use tokio::fs::File;
144/// # use arrow_array::RecordBatch;
145/// # use parquet::arrow::AsyncArrowWriter;
146/// # async fn test() {
147/// let mut writer: AsyncArrowWriter<File> = todo!();
148/// let batch: RecordBatch = todo!();
149/// writer.write(&batch).await.unwrap();
150/// // Trigger an early flush if buffered size exceeds 1_000_000
151/// if writer.in_progress_size() > 1_000_000 {
152///     writer.flush().await.unwrap()
153/// }
154/// # }
155/// ```
156pub struct AsyncArrowWriter<W> {
157    /// Underlying sync writer
158    sync_writer: ArrowWriter<Vec<u8>>,
159
160    /// Async writer provided by caller
161    async_writer: W,
162}
163
164impl<W: AsyncFileWriter> AsyncArrowWriter<W> {
165    /// Try to create a new Async Arrow Writer
166    pub fn try_new(
167        writer: W,
168        arrow_schema: SchemaRef,
169        props: Option<WriterProperties>,
170    ) -> Result<Self> {
171        let options = ArrowWriterOptions::new().with_properties(props.unwrap_or_default());
172        Self::try_new_with_options(writer, arrow_schema, options)
173    }
174
175    /// Try to create a new Async Arrow Writer with [`ArrowWriterOptions`]
176    pub fn try_new_with_options(
177        writer: W,
178        arrow_schema: SchemaRef,
179        options: ArrowWriterOptions,
180    ) -> Result<Self> {
181        let sync_writer = ArrowWriter::try_new_with_options(Vec::new(), arrow_schema, options)?;
182
183        Ok(Self {
184            sync_writer,
185            async_writer: writer,
186        })
187    }
188
189    /// Returns metadata for any flushed row groups
190    pub fn flushed_row_groups(&self) -> &[RowGroupMetaData] {
191        self.sync_writer.flushed_row_groups()
192    }
193
194    /// Estimated memory usage, in bytes, of this `ArrowWriter`
195    ///
196    /// See [ArrowWriter::memory_size] for more information.
197    pub fn memory_size(&self) -> usize {
198        self.sync_writer.memory_size()
199    }
200
201    /// Anticipated encoded size of the in progress row group.
202    ///
203    /// See [ArrowWriter::memory_size] for more information.
204    pub fn in_progress_size(&self) -> usize {
205        self.sync_writer.in_progress_size()
206    }
207
208    /// Returns the number of rows buffered in the in progress row group
209    pub fn in_progress_rows(&self) -> usize {
210        self.sync_writer.in_progress_rows()
211    }
212
213    /// Returns the number of bytes written by this instance
214    pub fn bytes_written(&self) -> usize {
215        self.sync_writer.bytes_written()
216    }
217
218    /// Enqueues the provided `RecordBatch` to be written
219    ///
220    /// After every sync write by the inner [ArrowWriter], the inner buffer will be
221    /// checked and flush if at least half full
222    pub async fn write(&mut self, batch: &RecordBatch) -> Result<()> {
223        let before = self.sync_writer.flushed_row_groups().len();
224        self.sync_writer.write(batch)?;
225        if before != self.sync_writer.flushed_row_groups().len() {
226            self.do_write().await?;
227        }
228        Ok(())
229    }
230
231    /// Flushes all buffered rows into a new row group
232    pub async fn flush(&mut self) -> Result<()> {
233        self.sync_writer.flush()?;
234        self.do_write().await?;
235
236        Ok(())
237    }
238
239    /// Append [`KeyValue`] metadata in addition to those in [`WriterProperties`]
240    ///
241    /// This method allows to append metadata after [`RecordBatch`]es are written.
242    pub fn append_key_value_metadata(&mut self, kv_metadata: KeyValue) {
243        self.sync_writer.append_key_value_metadata(kv_metadata);
244    }
245
246    /// Close and finalize the writer.
247    ///
248    /// All the data in the inner buffer will be force flushed.
249    ///
250    /// Unlike [`Self::close`] this does not consume self
251    ///
252    /// Attempting to write after calling finish will result in an error
253    pub async fn finish(&mut self) -> Result<ParquetMetaData> {
254        let metadata = self.sync_writer.finish()?;
255
256        // Force to flush the remaining data.
257        self.do_write().await?;
258        self.async_writer.complete().await?;
259
260        Ok(metadata)
261    }
262
263    /// Close and finalize the writer.
264    ///
265    /// All the data in the inner buffer will be force flushed.
266    pub async fn close(mut self) -> Result<ParquetMetaData> {
267        self.finish().await
268    }
269
270    /// Consumes the [`AsyncArrowWriter`] and returns the underlying [`AsyncFileWriter`]
271    ///
272    /// # Notes
273    ///
274    /// This method does **not** flush or finalize the writer, so buffered data
275    /// will be lost if you have not called [`Self::finish`].
276    pub fn into_inner(self) -> W {
277        self.async_writer
278    }
279
280    /// Flush the data written by `sync_writer` into the `async_writer`
281    ///
282    /// # Notes
283    ///
284    /// This method will take the inner buffer from the `sync_writer` and write it into the
285    /// async writer. After the write, the inner buffer will be empty.
286    async fn do_write(&mut self) -> Result<()> {
287        let buffer = mem::take(self.sync_writer.inner_mut());
288
289        self.async_writer
290            .write(Bytes::from(buffer))
291            .await
292            .map_err(|e| ParquetError::External(Box::new(e)))?;
293
294        Ok(())
295    }
296}
297
298#[cfg(test)]
299mod tests {
300    use crate::arrow::arrow_reader::{ParquetRecordBatchReader, ParquetRecordBatchReaderBuilder};
301    use arrow::datatypes::{DataType, Field, Schema};
302    use arrow_array::{ArrayRef, BinaryArray, Int32Array, Int64Array, RecordBatchReader};
303    use bytes::Bytes;
304    use std::sync::Arc;
305
306    use super::*;
307
308    fn get_test_reader() -> ParquetRecordBatchReader {
309        let testdata = arrow::util::test_util::parquet_test_data();
310        // This test file is large enough to generate multiple row groups.
311        let path = format!("{testdata}/alltypes_tiny_pages_plain.parquet");
312        let original_data = Bytes::from(std::fs::read(path).unwrap());
313        ParquetRecordBatchReaderBuilder::try_new(original_data)
314            .unwrap()
315            .build()
316            .unwrap()
317    }
318
319    #[tokio::test]
320    async fn test_async_writer() {
321        let col = Arc::new(Int64Array::from_iter_values([1, 2, 3])) as ArrayRef;
322        let to_write = RecordBatch::try_from_iter([("col", col)]).unwrap();
323
324        let mut buffer = Vec::new();
325        let mut writer = AsyncArrowWriter::try_new(&mut buffer, to_write.schema(), None).unwrap();
326        writer.write(&to_write).await.unwrap();
327        writer.close().await.unwrap();
328
329        let buffer = Bytes::from(buffer);
330        let mut reader = ParquetRecordBatchReaderBuilder::try_new(buffer)
331            .unwrap()
332            .build()
333            .unwrap();
334        let read = reader.next().unwrap().unwrap();
335
336        assert_eq!(to_write, read);
337    }
338
339    // Read the data from the test file and write it by the async writer and sync writer.
340    // And then compares the results of the two writers.
341    #[tokio::test]
342    async fn test_async_writer_with_sync_writer() {
343        let reader = get_test_reader();
344
345        let write_props = WriterProperties::builder()
346            .set_max_row_group_row_count(Some(64))
347            .build();
348
349        let mut async_buffer = Vec::new();
350        let mut async_writer = AsyncArrowWriter::try_new(
351            &mut async_buffer,
352            reader.schema(),
353            Some(write_props.clone()),
354        )
355        .unwrap();
356
357        let mut sync_buffer = Vec::new();
358        let mut sync_writer =
359            ArrowWriter::try_new(&mut sync_buffer, reader.schema(), Some(write_props)).unwrap();
360        for record_batch in reader {
361            let record_batch = record_batch.unwrap();
362            async_writer.write(&record_batch).await.unwrap();
363            sync_writer.write(&record_batch).unwrap();
364        }
365        sync_writer.close().unwrap();
366        async_writer.close().await.unwrap();
367
368        assert_eq!(sync_buffer, async_buffer);
369    }
370
371    #[tokio::test]
372    async fn test_async_writer_bytes_written() {
373        let col = Arc::new(Int64Array::from_iter_values([1, 2, 3])) as ArrayRef;
374        let to_write = RecordBatch::try_from_iter([("col", col)]).unwrap();
375
376        let temp = tempfile::tempfile().unwrap();
377
378        let file = tokio::fs::File::from_std(temp.try_clone().unwrap());
379        let mut writer =
380            AsyncArrowWriter::try_new(file.try_clone().await.unwrap(), to_write.schema(), None)
381                .unwrap();
382        writer.write(&to_write).await.unwrap();
383        let _metadata = writer.finish().await.unwrap();
384        // After `finish` this should include the metadata and footer
385        let reported = writer.bytes_written();
386
387        // Get actual size from file metadata
388        let actual = file.metadata().await.unwrap().len() as usize;
389
390        assert_eq!(reported, actual);
391    }
392
393    #[tokio::test]
394    async fn test_async_writer_file() {
395        let col = Arc::new(Int64Array::from_iter_values([1, 2, 3])) as ArrayRef;
396        let col2 = Arc::new(BinaryArray::from_iter_values(vec![
397            vec![0; 500000],
398            vec![0; 500000],
399            vec![0; 500000],
400        ])) as ArrayRef;
401        let to_write = RecordBatch::try_from_iter([("col", col), ("col2", col2)]).unwrap();
402
403        let temp = tempfile::tempfile().unwrap();
404
405        let file = tokio::fs::File::from_std(temp.try_clone().unwrap());
406        let mut writer = AsyncArrowWriter::try_new(file, to_write.schema(), None).unwrap();
407        writer.write(&to_write).await.unwrap();
408        writer.close().await.unwrap();
409
410        let mut reader = ParquetRecordBatchReaderBuilder::try_new(temp)
411            .unwrap()
412            .build()
413            .unwrap();
414        let read = reader.next().unwrap().unwrap();
415
416        assert_eq!(to_write, read);
417    }
418
419    #[tokio::test]
420    async fn in_progress_accounting() {
421        // define schema
422        let schema = Schema::new(vec![Field::new("a", DataType::Int32, false)]);
423
424        // create some data
425        let a = Int32Array::from_value(0_i32, 512);
426
427        // build a record batch
428        let batch = RecordBatch::try_new(Arc::new(schema), vec![Arc::new(a)]).unwrap();
429
430        let temp = tempfile::tempfile().unwrap();
431        let file = tokio::fs::File::from_std(temp.try_clone().unwrap());
432        let mut writer = AsyncArrowWriter::try_new(file, batch.schema(), None).unwrap();
433
434        // starts empty
435        assert_eq!(writer.in_progress_size(), 0);
436        assert_eq!(writer.in_progress_rows(), 0);
437        assert_eq!(writer.bytes_written(), 4); // Initial Parquet header
438        writer.write(&batch).await.unwrap();
439
440        // updated on write
441        let initial_size = writer.in_progress_size();
442        assert!(initial_size > 0);
443        assert_eq!(writer.in_progress_rows(), batch.num_rows());
444        let initial_memory = writer.memory_size();
445        // memory estimate is larger than estimated encoded size
446        assert!(
447            initial_size <= initial_memory,
448            "{initial_size} <= {initial_memory}"
449        );
450
451        // updated on second write
452        writer.write(&batch).await.unwrap();
453        assert!(writer.in_progress_size() > initial_size);
454        assert_eq!(writer.in_progress_rows(), batch.num_rows() * 2);
455        assert!(writer.memory_size() > initial_memory);
456        assert!(
457            writer.in_progress_size() <= writer.memory_size(),
458            "in_progress_size {} <= memory_size {}",
459            writer.in_progress_size(),
460            writer.memory_size()
461        );
462
463        // in progress tracking is cleared, but the overall data written is updated
464        let pre_flush_bytes_written = writer.bytes_written();
465        writer.flush().await.unwrap();
466        assert_eq!(writer.in_progress_size(), 0);
467        assert_eq!(writer.memory_size(), 0);
468        assert_eq!(writer.in_progress_rows(), 0);
469        assert!(writer.bytes_written() > pre_flush_bytes_written);
470
471        writer.close().await.unwrap();
472    }
473}
parquet/arrow/async_writer/mod.rs

parquet/arrow/async_writer/
mod.rs
