common_datasource/
file_format.rs

1// Copyright 2023 Greptime Team
2//
3// Licensed under the Apache License, Version 2.0 (the "License");
4// you may not use this file except in compliance with the License.
5// You may obtain a copy of the License at
6//
7//     http://www.apache.org/licenses/LICENSE-2.0
8//
9// Unless required by applicable law or agreed to in writing, software
10// distributed under the License is distributed on an "AS IS" BASIS,
11// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12// See the License for the specific language governing permissions and
13// limitations under the License.
14
15pub mod csv;
16pub mod json;
17pub mod orc;
18pub mod parquet;
19#[cfg(test)]
20pub mod tests;
21
22pub const DEFAULT_SCHEMA_INFER_MAX_RECORD: usize = 1000;
23
24use std::collections::HashMap;
25use std::result;
26use std::sync::Arc;
27use std::task::Poll;
28
29use arrow::record_batch::RecordBatch;
30use arrow_schema::{ArrowError, Schema as ArrowSchema};
31use async_trait::async_trait;
32use bytes::{Buf, Bytes};
33use common_recordbatch::DfSendableRecordBatchStream;
34use datafusion::datasource::file_format::file_compression_type::FileCompressionType as DfCompressionType;
35use datafusion::datasource::listing::PartitionedFile;
36use datafusion::datasource::object_store::ObjectStoreUrl;
37use datafusion::datasource::physical_plan::{
38    FileGroup, FileOpenFuture, FileScanConfigBuilder, FileSource, FileStream,
39};
40use datafusion::error::{DataFusionError, Result as DataFusionResult};
41use datafusion::physical_plan::SendableRecordBatchStream;
42use datafusion::physical_plan::metrics::ExecutionPlanMetricsSet;
43use futures::{StreamExt, TryStreamExt};
44use object_store::ObjectStore;
45use snafu::ResultExt;
46use tokio::io::AsyncWriteExt;
47use tokio_util::compat::FuturesAsyncWriteCompatExt;
48
49use self::csv::CsvFormat;
50use self::json::JsonFormat;
51use self::orc::OrcFormat;
52use self::parquet::ParquetFormat;
53use crate::DEFAULT_WRITE_BUFFER_SIZE;
54use crate::buffered_writer::DfRecordBatchEncoder;
55use crate::compressed_writer::{CompressedWriter, IntoCompressedWriter};
56use crate::compression::CompressionType;
57use crate::error::{self, Result};
58use crate::share_buffer::SharedBuffer;
59
60pub const FORMAT_COMPRESSION_TYPE: &str = "compression_type";
61pub const FORMAT_DELIMITER: &str = "delimiter";
62pub const FORMAT_SCHEMA_INFER_MAX_RECORD: &str = "schema_infer_max_record";
63pub const FORMAT_HAS_HEADER: &str = "has_header";
64pub const FORMAT_TYPE: &str = "format";
65pub const FILE_PATTERN: &str = "pattern";
66pub const TIMESTAMP_FORMAT: &str = "timestamp_format";
67pub const TIME_FORMAT: &str = "time_format";
68pub const DATE_FORMAT: &str = "date_format";
69
70#[derive(Debug, Clone, PartialEq, Eq)]
71pub enum Format {
72    Csv(CsvFormat),
73    Json(JsonFormat),
74    Parquet(ParquetFormat),
75    Orc(OrcFormat),
76}
77
78impl Format {
79    pub fn suffix(&self) -> &'static str {
80        match self {
81            Format::Csv(_) => ".csv",
82            Format::Json(_) => ".json",
83            Format::Parquet(_) => ".parquet",
84            &Format::Orc(_) => ".orc",
85        }
86    }
87}
88
89impl TryFrom<&HashMap<String, String>> for Format {
90    type Error = error::Error;
91
92    fn try_from(options: &HashMap<String, String>) -> Result<Self> {
93        let format = options
94            .get(FORMAT_TYPE)
95            .map(|format| format.to_ascii_uppercase())
96            .unwrap_or_else(|| "PARQUET".to_string());
97
98        match format.as_str() {
99            "CSV" => Ok(Self::Csv(CsvFormat::try_from(options)?)),
100            "JSON" => Ok(Self::Json(JsonFormat::try_from(options)?)),
101            "PARQUET" => Ok(Self::Parquet(ParquetFormat::default())),
102            "ORC" => Ok(Self::Orc(OrcFormat)),
103            _ => error::UnsupportedFormatSnafu { format: &format }.fail(),
104        }
105    }
106}
107
108#[async_trait]
109pub trait FileFormat: Send + Sync + std::fmt::Debug {
110    async fn infer_schema(&self, store: &ObjectStore, path: &str) -> Result<ArrowSchema>;
111}
112
113pub trait ArrowDecoder: Send + 'static {
114    /// Decode records from `buf` returning the number of bytes read.
115    ///
116    /// This method returns `Ok(0)` once `batch_size` objects have been parsed since the
117    /// last call to [`Self::flush`], or `buf` is exhausted.
118    ///
119    /// Any remaining bytes should be included in the next call to [`Self::decode`].
120    fn decode(&mut self, buf: &[u8]) -> result::Result<usize, ArrowError>;
121
122    /// Flushes the currently buffered data to a [`RecordBatch`].
123    ///
124    /// This should only be called after [`Self::decode`] has returned `Ok(0)`,
125    /// otherwise may return an error if part way through decoding a record
126    ///
127    /// Returns `Ok(None)` if no buffered data.
128    fn flush(&mut self) -> result::Result<Option<RecordBatch>, ArrowError>;
129}
130
131impl ArrowDecoder for arrow::csv::reader::Decoder {
132    fn decode(&mut self, buf: &[u8]) -> result::Result<usize, ArrowError> {
133        self.decode(buf)
134    }
135
136    fn flush(&mut self) -> result::Result<Option<RecordBatch>, ArrowError> {
137        self.flush()
138    }
139}
140
141impl ArrowDecoder for arrow::json::reader::Decoder {
142    fn decode(&mut self, buf: &[u8]) -> result::Result<usize, ArrowError> {
143        self.decode(buf)
144    }
145
146    fn flush(&mut self) -> result::Result<Option<RecordBatch>, ArrowError> {
147        self.flush()
148    }
149}
150
151pub fn open_with_decoder<T: ArrowDecoder, F: Fn() -> DataFusionResult<T>>(
152    object_store: Arc<ObjectStore>,
153    path: String,
154    compression_type: CompressionType,
155    decoder_factory: F,
156) -> DataFusionResult<FileOpenFuture> {
157    let mut decoder = decoder_factory()?;
158    Ok(Box::pin(async move {
159        let reader = object_store
160            .reader(&path)
161            .await
162            .map_err(|e| DataFusionError::External(Box::new(e)))?
163            .into_bytes_stream(..)
164            .await
165            .map_err(|e| DataFusionError::External(Box::new(e)))?;
166
167        let mut upstream = compression_type.convert_stream(reader).fuse();
168
169        let mut buffered = Bytes::new();
170
171        let stream = futures::stream::poll_fn(move |cx| {
172            loop {
173                if buffered.is_empty()
174                    && let Some(result) = futures::ready!(upstream.poll_next_unpin(cx))
175                {
176                    buffered = result?;
177                }
178
179                let decoded = decoder.decode(buffered.as_ref())?;
180
181                if decoded == 0 {
182                    break;
183                } else {
184                    buffered.advance(decoded);
185                }
186            }
187
188            Poll::Ready(decoder.flush().transpose())
189        });
190
191        Ok(stream.map_err(Into::into).boxed())
192    }))
193}
194
195pub async fn infer_schemas(
196    store: &ObjectStore,
197    files: &[String],
198    file_format: &dyn FileFormat,
199) -> Result<ArrowSchema> {
200    let mut schemas = Vec::with_capacity(files.len());
201    for file in files {
202        schemas.push(file_format.infer_schema(store, file).await?)
203    }
204    ArrowSchema::try_merge(schemas).context(error::MergeSchemaSnafu)
205}
206
207/// Writes data to a compressed writer if the data is not empty.
208///
209/// Does nothing if `data` is empty; otherwise writes all data and returns any error.
210async fn write_to_compressed_writer(
211    compressed_writer: &mut CompressedWriter,
212    data: &[u8],
213) -> Result<()> {
214    if !data.is_empty() {
215        compressed_writer
216            .write_all(data)
217            .await
218            .context(error::AsyncWriteSnafu)?;
219    }
220    Ok(())
221}
222
223/// Streams [SendableRecordBatchStream] to a file with optional compression support.
224/// Data is buffered and flushed according to the given `threshold`.
225/// Ensures that writer resources are cleanly released and that an empty file is not
226/// created if no rows are written.
227///
228/// Returns the total number of rows successfully written.
229pub async fn stream_to_file<E>(
230    mut stream: SendableRecordBatchStream,
231    store: ObjectStore,
232    path: &str,
233    threshold: usize,
234    concurrency: usize,
235    compression_type: CompressionType,
236    encoder_factory: impl Fn(SharedBuffer) -> E,
237) -> Result<usize>
238where
239    E: DfRecordBatchEncoder,
240{
241    // Create the file writer with OpenDAL's built-in buffering
242    let writer = store
243        .writer_with(path)
244        .concurrent(concurrency)
245        .chunk(DEFAULT_WRITE_BUFFER_SIZE.as_bytes() as usize)
246        .await
247        .with_context(|_| error::WriteObjectSnafu { path })?
248        .into_futures_async_write()
249        .compat_write();
250
251    // Apply compression if needed
252    let mut compressed_writer = writer.into_compressed_writer(compression_type);
253
254    // Create a buffer for the encoder
255    let buffer = SharedBuffer::with_capacity(threshold);
256    let mut encoder = encoder_factory(buffer.clone());
257
258    let mut rows = 0;
259
260    // Process each record batch
261    while let Some(batch) = stream.next().await {
262        let batch = batch.context(error::ReadRecordBatchSnafu)?;
263
264        // Write batch using the encoder
265        encoder.write(&batch)?;
266        rows += batch.num_rows();
267
268        loop {
269            let chunk = {
270                let mut buffer_guard = buffer.buffer.lock().unwrap();
271                if buffer_guard.len() < threshold {
272                    break;
273                }
274                buffer_guard.split_to(threshold)
275            };
276            write_to_compressed_writer(&mut compressed_writer, &chunk).await?;
277        }
278    }
279
280    // If no row's been written, just simply close the underlying writer
281    // without flush so that no file will be actually created.
282    if rows != 0 {
283        // Final flush of any remaining data
284        let final_data = {
285            let mut buffer_guard = buffer.buffer.lock().unwrap();
286            buffer_guard.split()
287        };
288        write_to_compressed_writer(&mut compressed_writer, &final_data).await?;
289    }
290
291    // Shutdown compression and close writer
292    compressed_writer.shutdown().await?;
293
294    Ok(rows)
295}
296
297/// Creates a [FileStream] for reading data from a file with optional column projection
298/// and compression support.
299///
300/// Returns [SendableRecordBatchStream].
301pub async fn file_to_stream(
302    store: &ObjectStore,
303    filename: &str,
304    file_source: Arc<dyn FileSource>,
305    projection: Option<Vec<usize>>,
306    compression_type: CompressionType,
307) -> Result<DfSendableRecordBatchStream> {
308    let df_compression: DfCompressionType = compression_type.into();
309    let config =
310        FileScanConfigBuilder::new(ObjectStoreUrl::local_filesystem(), file_source.clone())
311            .with_file_group(FileGroup::new(vec![PartitionedFile::new(
312                filename.to_string(),
313                0,
314            )]))
315            .with_projection_indices(projection)?
316            .with_file_compression_type(df_compression)
317            .build();
318
319    let store = Arc::new(object_store::compat::OpendalStore::new(store.clone()));
320    let file_opener = config.file_source().create_file_opener(store, &config, 0)?;
321    let stream = FileStream::new(&config, 0, file_opener, &ExecutionPlanMetricsSet::new())?;
322
323    Ok(Box::pin(stream))
324}
common_datasource/file_format.rs

common_datasource/
file_format.rs
