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_HEADERS: &str = "headers";
64pub const FORMAT_HAS_HEADER: &str = "has_header";
65pub const FORMAT_SKIP_BAD_RECORDS: &str = "skip_bad_records";
66pub const FORMAT_STRICT_HEADERS: &str = "strict_headers";
67pub const FORMAT_TYPE: &str = "format";
68pub const FILE_PATTERN: &str = "pattern";
69pub const TIMESTAMP_FORMAT: &str = "timestamp_format";
70pub const TIME_FORMAT: &str = "time_format";
71pub const DATE_FORMAT: &str = "date_format";
72
73#[derive(Debug, Clone, PartialEq, Eq)]
74pub enum Format {
75    Csv(CsvFormat),
76    Json(JsonFormat),
77    Parquet(ParquetFormat),
78    Orc(OrcFormat),
79}
80
81impl Format {
82    pub fn suffix(&self) -> &'static str {
83        match self {
84            Format::Csv(_) => ".csv",
85            Format::Json(_) => ".json",
86            Format::Parquet(_) => ".parquet",
87            &Format::Orc(_) => ".orc",
88        }
89    }
90}
91
92impl TryFrom<&HashMap<String, String>> for Format {
93    type Error = error::Error;
94
95    fn try_from(options: &HashMap<String, String>) -> Result<Self> {
96        let format = options
97            .get(FORMAT_TYPE)
98            .map(|format| format.to_ascii_uppercase())
99            .unwrap_or_else(|| "PARQUET".to_string());
100
101        match format.as_str() {
102            "CSV" => Ok(Self::Csv(CsvFormat::try_from(options)?)),
103            "JSON" => Ok(Self::Json(JsonFormat::try_from(options)?)),
104            "PARQUET" => Ok(Self::Parquet(ParquetFormat::default())),
105            "ORC" => Ok(Self::Orc(OrcFormat)),
106            _ => error::UnsupportedFormatSnafu { format: &format }.fail(),
107        }
108    }
109}
110
111#[async_trait]
112pub trait FileFormat: Send + Sync + std::fmt::Debug {
113    async fn infer_schema(&self, store: &ObjectStore, path: &str) -> Result<ArrowSchema>;
114}
115
116pub trait ArrowDecoder: Send + 'static {
117    /// Decode records from `buf` returning the number of bytes read.
118    ///
119    /// This method returns `Ok(0)` once `batch_size` objects have been parsed since the
120    /// last call to [`Self::flush`], or `buf` is exhausted.
121    ///
122    /// Any remaining bytes should be included in the next call to [`Self::decode`].
123    fn decode(&mut self, buf: &[u8]) -> result::Result<usize, ArrowError>;
124
125    /// Flushes the currently buffered data to a [`RecordBatch`].
126    ///
127    /// This should only be called after [`Self::decode`] has returned `Ok(0)`,
128    /// otherwise may return an error if part way through decoding a record
129    ///
130    /// Returns `Ok(None)` if no buffered data.
131    fn flush(&mut self) -> result::Result<Option<RecordBatch>, ArrowError>;
132}
133
134impl ArrowDecoder for arrow::csv::reader::Decoder {
135    fn decode(&mut self, buf: &[u8]) -> result::Result<usize, ArrowError> {
136        self.decode(buf)
137    }
138
139    fn flush(&mut self) -> result::Result<Option<RecordBatch>, ArrowError> {
140        self.flush()
141    }
142}
143
144impl ArrowDecoder for arrow::json::reader::Decoder {
145    fn decode(&mut self, buf: &[u8]) -> result::Result<usize, ArrowError> {
146        self.decode(buf)
147    }
148
149    fn flush(&mut self) -> result::Result<Option<RecordBatch>, ArrowError> {
150        self.flush()
151    }
152}
153
154pub fn open_with_decoder<T: ArrowDecoder, F: Fn() -> DataFusionResult<T>>(
155    object_store: Arc<ObjectStore>,
156    path: String,
157    compression_type: CompressionType,
158    decoder_factory: F,
159) -> DataFusionResult<FileOpenFuture> {
160    let mut decoder = decoder_factory()?;
161    Ok(Box::pin(async move {
162        let reader = object_store
163            .reader(&path)
164            .await
165            .map_err(|e| DataFusionError::External(Box::new(e)))?
166            .into_bytes_stream(..)
167            .await
168            .map_err(|e| DataFusionError::External(Box::new(e)))?;
169
170        let mut upstream = compression_type.convert_stream(reader).fuse();
171
172        let mut buffered = Bytes::new();
173
174        let stream = futures::stream::poll_fn(move |cx| {
175            loop {
176                if buffered.is_empty()
177                    && let Some(result) = futures::ready!(upstream.poll_next_unpin(cx))
178                {
179                    buffered = result?;
180                }
181
182                let decoded = decoder.decode(buffered.as_ref())?;
183
184                if decoded == 0 {
185                    break;
186                } else {
187                    buffered.advance(decoded);
188                }
189            }
190
191            Poll::Ready(decoder.flush().transpose())
192        });
193
194        Ok(stream.map_err(Into::into).boxed())
195    }))
196}
197
198pub async fn infer_schemas(
199    store: &ObjectStore,
200    files: &[String],
201    file_format: &dyn FileFormat,
202) -> Result<ArrowSchema> {
203    let mut schemas = Vec::with_capacity(files.len());
204    for file in files {
205        schemas.push(file_format.infer_schema(store, file).await?)
206    }
207    ArrowSchema::try_merge(schemas).context(error::MergeSchemaSnafu)
208}
209
210/// Writes data to a compressed writer if the data is not empty.
211///
212/// Does nothing if `data` is empty; otherwise writes all data and returns any error.
213async fn write_to_compressed_writer(
214    compressed_writer: &mut CompressedWriter,
215    data: &[u8],
216) -> Result<()> {
217    if !data.is_empty() {
218        compressed_writer
219            .write_all(data)
220            .await
221            .context(error::AsyncWriteSnafu)?;
222    }
223    Ok(())
224}
225
226/// Streams [SendableRecordBatchStream] to a file with optional compression support.
227/// Data is buffered and flushed according to the given `threshold`.
228/// Ensures that writer resources are cleanly released and that an empty file is not
229/// created if no rows are written.
230///
231/// Returns the total number of rows successfully written.
232pub async fn stream_to_file<E>(
233    mut stream: SendableRecordBatchStream,
234    store: ObjectStore,
235    path: &str,
236    threshold: usize,
237    concurrency: usize,
238    compression_type: CompressionType,
239    encoder_factory: impl Fn(SharedBuffer) -> E,
240) -> Result<usize>
241where
242    E: DfRecordBatchEncoder,
243{
244    // Create the file writer with OpenDAL's built-in buffering
245    let writer = store
246        .writer_with(path)
247        .concurrent(concurrency)
248        .chunk(DEFAULT_WRITE_BUFFER_SIZE.as_bytes() as usize)
249        .await
250        .with_context(|_| error::WriteObjectSnafu { path })?
251        .into_futures_async_write()
252        .compat_write();
253
254    // Apply compression if needed
255    let mut compressed_writer = writer.into_compressed_writer(compression_type);
256
257    // Create a buffer for the encoder
258    let buffer = SharedBuffer::with_capacity(threshold);
259    let mut encoder = encoder_factory(buffer.clone());
260
261    let mut rows = 0;
262
263    // Process each record batch
264    while let Some(batch) = stream.next().await {
265        let batch = batch.context(error::ReadRecordBatchSnafu)?;
266
267        // Write batch using the encoder
268        encoder.write(&batch)?;
269        rows += batch.num_rows();
270
271        loop {
272            let chunk = {
273                let mut buffer_guard = buffer.buffer.lock().unwrap();
274                if buffer_guard.len() < threshold {
275                    break;
276                }
277                buffer_guard.split_to(threshold)
278            };
279            write_to_compressed_writer(&mut compressed_writer, &chunk).await?;
280        }
281    }
282
283    // If no row's been written, just simply close the underlying writer
284    // without flush so that no file will be actually created.
285    if rows != 0 {
286        // Final flush of any remaining data
287        let final_data = {
288            let mut buffer_guard = buffer.buffer.lock().unwrap();
289            buffer_guard.split()
290        };
291        write_to_compressed_writer(&mut compressed_writer, &final_data).await?;
292    }
293
294    // Shutdown compression and close writer
295    compressed_writer.shutdown().await?;
296
297    Ok(rows)
298}
299
300/// Creates a [FileStream] for reading data from a file with optional column projection
301/// and compression support.
302///
303/// Returns [SendableRecordBatchStream].
304pub async fn file_to_stream(
305    store: &ObjectStore,
306    filename: &str,
307    file_source: Arc<dyn FileSource>,
308    projection: Option<Vec<usize>>,
309    compression_type: CompressionType,
310) -> Result<DfSendableRecordBatchStream> {
311    let df_compression: DfCompressionType = compression_type.into();
312    let config =
313        FileScanConfigBuilder::new(ObjectStoreUrl::local_filesystem(), file_source.clone())
314            .with_file_group(FileGroup::new(vec![PartitionedFile::new(
315                filename.to_string(),
316                0,
317            )]))
318            .with_projection_indices(projection)?
319            .with_file_compression_type(df_compression)
320            .build();
321
322    let store = Arc::new(object_store_opendal::OpendalStore::new(store.clone()));
323    let file_opener = config.file_source().create_file_opener(store, &config, 0)?;
324    let stream = FileStream::new(&config, 0, file_opener, &ExecutionPlanMetricsSet::new())?;
325
326    Ok(Box::pin(stream))
327}
common_datasource/file_format.rs

common_datasource/
file_format.rs
