Skip to main content

risingwave_object_store/object/opendal_engine/
opendal_object_store.rs

1// Copyright 2023 RisingWave Labs
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
15use std::sync::Arc;
16use std::time::Duration;
17
18use bytes::{Bytes, BytesMut};
19use fail::fail_point;
20use futures::StreamExt;
21use opendal::layers::{RetryLayer, TimeoutLayer};
22use opendal::raw::BoxedStaticFuture;
23use opendal::services::Memory;
24use opendal::{Execute, Executor, Operator, Writer};
25use risingwave_common::config::ObjectStoreConfig;
26use risingwave_common::range::RangeBoundsExt;
27use thiserror_ext::AsReport;
28
29use crate::object::object_metrics::ObjectStoreMetrics;
30use crate::object::{
31    MonitoredObjectStore, ObjectDataStream, ObjectError, ObjectMetadata, ObjectMetadataIter,
32    ObjectRangeBounds, ObjectResult, ObjectStore, OperationType, StreamingUploader, prefix,
33};
34
35/// Opendal object storage.
36#[derive(Clone)]
37pub struct OpendalObjectStore {
38    pub(crate) op: Operator,
39    pub(crate) media_type: MediaType,
40    pub(crate) config: Arc<ObjectStoreConfig>,
41    pub(crate) metrics: Arc<ObjectStoreMetrics>,
42}
43
44#[derive(Clone)]
45pub enum MediaType {
46    Memory,
47    Hdfs,
48    Gcs,
49    Minio,
50    S3,
51    Obs,
52    Oss,
53    Webhdfs,
54    Azblob,
55    Fs,
56}
57
58impl MediaType {
59    pub fn as_str(&self) -> &'static str {
60        match self {
61            MediaType::Memory => "Memory",
62            MediaType::Hdfs => "Hdfs",
63            MediaType::Gcs => "Gcs",
64            MediaType::Minio => "Minio",
65            MediaType::S3 => "S3",
66            MediaType::Obs => "Obs",
67            MediaType::Oss => "Oss",
68            MediaType::Webhdfs => "Webhdfs",
69            MediaType::Azblob => "Azblob",
70            MediaType::Fs => "Fs",
71        }
72    }
73}
74
75impl OpendalObjectStore {
76    /// create opendal memory engine, used for unit tests.
77    pub fn test_new_memory_engine() -> ObjectResult<Self> {
78        // Create memory backend builder.
79        let builder = Memory::default();
80        let op: Operator = Operator::new(builder)?;
81
82        Ok(Self {
83            op,
84            media_type: MediaType::Memory,
85            config: Arc::new(ObjectStoreConfig::default()),
86            metrics: Arc::new(ObjectStoreMetrics::unused()),
87        })
88    }
89}
90
91#[inline]
92fn timestamp_to_secs(t: opendal::raw::Timestamp) -> f64 {
93    t.into_inner().as_second() as f64
94}
95
96#[async_trait::async_trait]
97impl ObjectStore for OpendalObjectStore {
98    type StreamingUploader = OpendalStreamingUploader;
99
100    fn get_object_prefix(&self, obj_id: u64, use_new_object_prefix_strategy: bool) -> String {
101        match self.media_type {
102            MediaType::S3 => prefix::s3::get_object_prefix(obj_id),
103            MediaType::Minio => prefix::s3::get_object_prefix(obj_id),
104            MediaType::Memory => String::default(),
105            MediaType::Hdfs
106            | MediaType::Gcs
107            | MediaType::Obs
108            | MediaType::Oss
109            | MediaType::Webhdfs
110            | MediaType::Azblob
111            | MediaType::Fs => {
112                prefix::opendal_engine::get_object_prefix(obj_id, use_new_object_prefix_strategy)
113            }
114        }
115    }
116
117    async fn upload(&self, path: &str, obj: Bytes) -> ObjectResult<()> {
118        if obj.is_empty() {
119            Err(ObjectError::internal("upload empty object"))
120        } else {
121            self.op.write(path, obj).await?;
122            Ok(())
123        }
124    }
125
126    async fn streaming_upload(&self, path: &str) -> ObjectResult<Self::StreamingUploader> {
127        if self.op.info().capability().write_can_multi {
128            Ok(OpendalStreamingUploader::Native(Box::new(
129                OpendalNativeStreamingUploader::new(
130                    self.op.clone(),
131                    path.to_owned(),
132                    self.config.clone(),
133                    self.metrics.clone(),
134                    self.store_media_type(),
135                )
136                .await?,
137            )))
138        } else {
139            Ok(OpendalStreamingUploader::Buffered(
140                OpendalBufferedStreamingUploader::new(
141                    self.clone()
142                        .monitored(self.metrics.clone(), self.config.clone()),
143                    path.to_owned(),
144                ),
145            ))
146        }
147    }
148
149    async fn read(&self, path: &str, range: impl ObjectRangeBounds) -> ObjectResult<Bytes> {
150        let data = if range.is_full() {
151            self.op.read(path).await?
152        } else {
153            self.op
154                .read_with(path)
155                .range(range.map(|v| *v as u64))
156                .await?
157        };
158
159        if let Some(len) = range.len()
160            && len != data.len()
161        {
162            return Err(ObjectError::internal(format!(
163                "mismatched size: expected {}, found {} when reading {} at {:?}",
164                len,
165                data.len(),
166                path,
167                range,
168            )));
169        }
170
171        Ok(data.to_bytes())
172    }
173
174    /// Returns a stream reading the object specified in `path`. If given, the stream starts at the
175    /// byte with index `start_pos` (0-based). As far as possible, the stream only loads the amount
176    /// of data into memory that is read from the stream.
177    async fn streaming_read(
178        &self,
179        path: &str,
180        range: impl ObjectRangeBounds,
181    ) -> ObjectResult<ObjectDataStream> {
182        fail_point!("opendal_streaming_read_err", |_| Err(
183            ObjectError::internal("opendal streaming read error")
184        ));
185        let range = range.map(|v| *v as u64);
186
187        // The layer specified first will be executed first.
188        // `TimeoutLayer` must be specified before `RetryLayer`.
189        // Otherwise, it will lead to bad state inside OpenDAL and panic.
190        // See https://docs.rs/opendal/latest/opendal/layers/struct.RetryLayer.html#panics
191        let reader = self
192            .op
193            .clone()
194            .layer(TimeoutLayer::new().with_io_timeout(Duration::from_millis(
195                self.config.retry.streaming_read_attempt_timeout_ms,
196            )))
197            .layer(
198                RetryLayer::new()
199                    .with_min_delay(Duration::from_millis(
200                        self.config.retry.req_backoff_interval_ms,
201                    ))
202                    .with_max_delay(Duration::from_millis(
203                        self.config.retry.req_backoff_max_delay_ms,
204                    ))
205                    .with_max_times(self.config.retry.streaming_read_retry_attempts)
206                    .with_factor(self.config.retry.req_backoff_factor as f32)
207                    .with_jitter(),
208            )
209            .reader_with(path)
210            .await?;
211        let stream = reader.into_bytes_stream(range).await?.map(|item| {
212            item.map(|b| Bytes::copy_from_slice(b.as_ref()))
213                .map_err(|e| {
214                    ObjectError::internal(format!("reader into_stream fail {}", e.as_report()))
215                })
216        });
217
218        Ok(Box::pin(stream))
219    }
220
221    async fn metadata(&self, path: &str) -> ObjectResult<ObjectMetadata> {
222        let opendal_metadata = self.op.stat(path).await?;
223        let key = path.to_owned();
224        let last_modified = match opendal_metadata.last_modified() {
225            Some(t) => timestamp_to_secs(t),
226            None => 0_f64,
227        };
228
229        let total_size = opendal_metadata.content_length() as usize;
230        let metadata = ObjectMetadata {
231            key,
232            last_modified,
233            total_size,
234        };
235        Ok(metadata)
236    }
237
238    async fn delete(&self, path: &str) -> ObjectResult<()> {
239        self.op.delete(path).await?;
240        Ok(())
241    }
242
243    /// Deletes the objects with the given paths permanently from the storage. If an object
244    /// specified in the request is not found, it will be considered as successfully deleted.
245    async fn delete_objects(&self, paths: &[String]) -> ObjectResult<()> {
246        self.op.delete_iter(paths.to_vec()).await?;
247        Ok(())
248    }
249
250    async fn list(
251        &self,
252        prefix: &str,
253        start_after: Option<String>,
254        limit: Option<usize>,
255    ) -> ObjectResult<ObjectMetadataIter> {
256        let mut object_lister = self.op.lister_with(prefix).recursive(true);
257        if let Some(start_after) = start_after {
258            object_lister = object_lister.start_after(&start_after);
259        }
260        let object_lister = object_lister.await?;
261
262        let op = self.op.clone();
263        let stream = object_lister.then(move |object| {
264            let op = op.clone();
265            async move { Self::listed_object_metadata(&op, object?).await }
266        });
267
268        Ok(stream.take(limit.unwrap_or(usize::MAX)).boxed())
269    }
270
271    fn store_media_type(&self) -> &'static str {
272        self.media_type.as_str()
273    }
274}
275
276impl OpendalObjectStore {
277    async fn listed_object_metadata(
278        op: &Operator,
279        object: opendal::Entry,
280    ) -> ObjectResult<ObjectMetadata> {
281        let key = object.path().to_owned();
282        // OpenDAL 0.55 removed list metadata capability flags and reports unknown sizes as 0.
283        let meta = object.metadata();
284        let mut last_modified = meta.last_modified().map(timestamp_to_secs);
285        let mut total_size = meta.content_length() as usize;
286        if last_modified.is_none() || total_size == 0 {
287            // Propagate stat failures; treating one as EOF can make recovery trust a partial scan.
288            let stat_meta = op.stat(&key).await?;
289            last_modified = stat_meta.last_modified().map(timestamp_to_secs);
290            total_size = stat_meta.content_length() as usize;
291        }
292        Ok(ObjectMetadata {
293            key,
294            last_modified: last_modified.unwrap_or(0_f64),
295            total_size,
296        })
297    }
298
299    pub async fn copy(&self, from_path: &str, to_path: &str) -> ObjectResult<()> {
300        self.op.copy(from_path, to_path).await?;
301        Ok(())
302    }
303}
304
305struct OpendalStreamingUploaderExecute {
306    /// To record metrics for uploading part.
307    metrics: Arc<ObjectStoreMetrics>,
308    media_type: &'static str,
309}
310
311impl OpendalStreamingUploaderExecute {
312    const STREAMING_UPLOAD_TYPE: OperationType = OperationType::StreamingUpload;
313
314    fn new(metrics: Arc<ObjectStoreMetrics>, media_type: &'static str) -> Self {
315        Self {
316            metrics,
317            media_type,
318        }
319    }
320}
321
322impl Execute for OpendalStreamingUploaderExecute {
323    fn execute(&self, f: BoxedStaticFuture<()>) {
324        let operation_type_str = Self::STREAMING_UPLOAD_TYPE.as_str();
325        let media_type = self.media_type;
326
327        let metrics = self.metrics.clone();
328        let _handle = tokio::spawn(async move {
329            let _timer = metrics
330                .operation_latency
331                .with_label_values(&[media_type, operation_type_str])
332                .start_timer();
333
334            f.await
335        });
336    }
337}
338
339pub enum OpendalStreamingUploader {
340    Native(Box<OpendalNativeStreamingUploader>),
341    Buffered(OpendalBufferedStreamingUploader),
342}
343
344/// Store multiple parts in a map, and concatenate them on finish.
345pub struct OpendalNativeStreamingUploader {
346    writer: Writer,
347    /// Buffer for data. It will store at least `UPLOAD_BUFFER_SIZE` bytes of data before wrapping itself
348    /// into a stream and upload to object store as a part.
349    buf: Vec<Bytes>,
350    /// Length of the data that have not been uploaded to object store.
351    not_uploaded_len: usize,
352    /// Whether the writer is valid. The writer is invalid after abort/close.
353    is_valid: bool,
354
355    abort_on_err: bool,
356
357    upload_part_size: usize,
358}
359
360impl OpendalNativeStreamingUploader {
361    pub async fn new(
362        op: Operator,
363        path: String,
364        config: Arc<ObjectStoreConfig>,
365        metrics: Arc<ObjectStoreMetrics>,
366        media_type: &'static str,
367    ) -> ObjectResult<Self> {
368        let monitored_execute = OpendalStreamingUploaderExecute::new(metrics, media_type);
369        let executor = Executor::with(monitored_execute);
370        // Start from the base context so existing layers are replayed exactly once by
371        // `with_context`. Using the composed context here would double-wrap resources such as
372        // the HTTP concurrency semaphore.
373        let ctx = op.base_context().with_executor(executor);
374        let op = op.with_context(ctx);
375
376        let writer = op
377            .clone()
378            .layer(TimeoutLayer::new().with_io_timeout(Duration::from_millis(
379                config.retry.streaming_upload_attempt_timeout_ms,
380            )))
381            .layer(
382                RetryLayer::new()
383                    .with_min_delay(Duration::from_millis(config.retry.req_backoff_interval_ms))
384                    .with_max_delay(Duration::from_millis(config.retry.req_backoff_max_delay_ms))
385                    .with_max_times(config.retry.streaming_upload_retry_attempts)
386                    .with_factor(config.retry.req_backoff_factor as f32)
387                    .with_jitter(),
388            )
389            .writer_with(&path)
390            .concurrent(config.opendal_upload_concurrency)
391            .await?;
392        Ok(Self {
393            writer,
394            buf: vec![],
395            not_uploaded_len: 0,
396            is_valid: true,
397            abort_on_err: config.opendal_writer_abort_on_err,
398            upload_part_size: config.upload_part_size,
399        })
400    }
401
402    async fn flush(&mut self) -> ObjectResult<()> {
403        let data: Vec<Bytes> = self.buf.drain(..).collect();
404        debug_assert_eq!(
405            data.iter().map(|b| b.len()).sum::<usize>(),
406            self.not_uploaded_len
407        );
408        if let Err(err) = self.writer.write(data).await {
409            self.is_valid = false;
410            if self.abort_on_err {
411                self.writer.abort().await?;
412            }
413            return Err(err.into());
414        }
415        self.not_uploaded_len = 0;
416        Ok(())
417    }
418}
419
420impl StreamingUploader for OpendalNativeStreamingUploader {
421    async fn write_bytes(&mut self, data: Bytes) -> ObjectResult<()> {
422        assert!(self.is_valid);
423        self.not_uploaded_len += data.len();
424        self.buf.push(data);
425        if self.not_uploaded_len >= self.upload_part_size {
426            self.flush().await?;
427        }
428        Ok(())
429    }
430
431    async fn finish(mut self) -> ObjectResult<()> {
432        assert!(self.is_valid);
433        if self.not_uploaded_len > 0 {
434            self.flush().await?;
435        }
436
437        assert!(self.buf.is_empty());
438        assert_eq!(self.not_uploaded_len, 0);
439
440        self.is_valid = false;
441        match self.writer.close().await {
442            Ok(_) => (),
443            Err(err) => {
444                if self.abort_on_err {
445                    self.writer.abort().await?;
446                }
447                return Err(err.into());
448            }
449        };
450
451        Ok(())
452    }
453
454    // Not absolutely accurate. Some bytes may be in the infight request.
455    fn get_memory_usage(&self) -> u64 {
456        self.not_uploaded_len as u64
457    }
458}
459
460pub struct OpendalBufferedStreamingUploader {
461    store: MonitoredObjectStore<OpendalObjectStore>,
462    path: String,
463    buf: BytesMut,
464}
465
466impl OpendalBufferedStreamingUploader {
467    fn new(store: MonitoredObjectStore<OpendalObjectStore>, path: String) -> Self {
468        Self {
469            store,
470            path,
471            buf: BytesMut::new(),
472        }
473    }
474}
475
476impl StreamingUploader for OpendalBufferedStreamingUploader {
477    async fn write_bytes(&mut self, data: Bytes) -> ObjectResult<()> {
478        self.buf.extend_from_slice(&data);
479        Ok(())
480    }
481
482    async fn finish(self) -> ObjectResult<()> {
483        self.store.upload(&self.path, self.buf.freeze()).await
484    }
485
486    fn get_memory_usage(&self) -> u64 {
487        self.buf.len() as u64
488    }
489}
490
491impl StreamingUploader for OpendalStreamingUploader {
492    async fn write_bytes(&mut self, data: Bytes) -> ObjectResult<()> {
493        match self {
494            Self::Native(uploader) => uploader.write_bytes(data).await,
495            Self::Buffered(uploader) => uploader.write_bytes(data).await,
496        }
497    }
498
499    async fn finish(self) -> ObjectResult<()> {
500        match self {
501            Self::Native(uploader) => (*uploader).finish().await,
502            Self::Buffered(uploader) => uploader.finish().await,
503        }
504    }
505
506    fn get_memory_usage(&self) -> u64 {
507        match self {
508            Self::Native(uploader) => uploader.get_memory_usage(),
509            Self::Buffered(uploader) => uploader.get_memory_usage(),
510        }
511    }
512}
513
514#[cfg(test)]
515mod tests {
516    use futures::TryStreamExt;
517
518    use super::*;
519
520    async fn list_all(prefix: &str, store: &OpendalObjectStore) -> Vec<ObjectMetadata> {
521        store
522            .list(prefix, None, None)
523            .await
524            .unwrap()
525            .try_collect::<Vec<_>>()
526            .await
527            .unwrap()
528    }
529
530    #[tokio::test]
531    async fn test_memory_upload() {
532        let block = Bytes::from("123456");
533        let store = OpendalObjectStore::test_new_memory_engine().unwrap();
534        store.upload("/abc", block).await.unwrap();
535
536        // No such object.
537        store.read("/ab", 0..3).await.unwrap_err();
538
539        let bytes = store.read("/abc", 4..6).await.unwrap();
540        assert_eq!(String::from_utf8(bytes.to_vec()).unwrap(), "56".to_owned());
541
542        store.delete("/abc").await.unwrap();
543
544        // No such object.
545        store.read("/abc", 0..3).await.unwrap_err();
546    }
547
548    #[tokio::test]
549    async fn test_memory_read_out_of_range_returns_error() {
550        let block = Bytes::from("123456");
551        let store = OpendalObjectStore::test_new_memory_engine().unwrap();
552        store.upload("/abc", block).await.unwrap();
553
554        // OpenDAL 0.58 reports an out-of-range read instead of panicking.
555        store.read("/abc", 4..44).await.unwrap_err();
556    }
557
558    #[tokio::test]
559    async fn test_memory_metadata() {
560        let block = Bytes::from("123456");
561        let path = "/abc".to_owned();
562        let obj_store = OpendalObjectStore::test_new_memory_engine().unwrap();
563        obj_store.upload("/abc", block).await.unwrap();
564
565        let err = obj_store.metadata("/not_exist").await.unwrap_err();
566        assert!(err.is_object_not_found_error());
567        let metadata = obj_store.metadata("/abc").await.unwrap();
568        assert_eq!(metadata.total_size, 6);
569        obj_store.delete(&path).await.unwrap();
570    }
571
572    #[tokio::test]
573    async fn test_buffered_streaming_upload() {
574        let store = OpendalObjectStore::test_new_memory_engine().unwrap();
575        let mut uploader = OpendalBufferedStreamingUploader::new(
576            store
577                .clone()
578                .monitored(store.metrics.clone(), store.config.clone()),
579            "/abc".to_owned(),
580        );
581
582        uploader.write_bytes(Bytes::from("123")).await.unwrap();
583        assert_eq!(uploader.get_memory_usage(), 3);
584        uploader.write_bytes(Bytes::from("456")).await.unwrap();
585        assert_eq!(uploader.get_memory_usage(), 6);
586        uploader.finish().await.unwrap();
587
588        let read_obj = store.read("/abc", ..).await.unwrap();
589        assert_eq!(read_obj, Bytes::from("123456"));
590    }
591
592    #[tokio::test]
593    async fn test_empty_buffered_streaming_upload() {
594        let store = OpendalObjectStore::test_new_memory_engine().unwrap();
595        let uploader = OpendalBufferedStreamingUploader::new(
596            store
597                .clone()
598                .monitored(store.metrics.clone(), store.config.clone()),
599            "/abc".to_owned(),
600        );
601
602        uploader.finish().await.unwrap_err();
603    }
604
605    #[tokio::test]
606    async fn test_listed_object_metadata_propagates_stat_failure() {
607        let store = OpendalObjectStore::test_new_memory_engine().unwrap();
608        store
609            .upload("sst", Bytes::from_static(b"data"))
610            .await
611            .unwrap();
612        // Capture the listed entry before deletion so the subsequent metadata lookup fails.
613        // Memory entries have no last_modified, so our adapter must stat even a listed file.
614        let mut objects = store.op.lister_with("").recursive(true).await.unwrap();
615        let object = objects.next().await.unwrap().unwrap();
616        store.delete("sst").await.unwrap();
617        let error = OpendalObjectStore::listed_object_metadata(&store.op, object)
618            .await
619            .unwrap_err();
620        assert!(error.is_object_not_found_error());
621    }
622
623    #[tokio::test]
624    async fn test_memory_delete_objects_and_list_object() {
625        let block1 = Bytes::from("123456");
626        let block2 = Bytes::from("987654");
627        let store = OpendalObjectStore::test_new_memory_engine().unwrap();
628        store.upload("abc", Bytes::from("123456")).await.unwrap();
629        store.upload("prefix/abc", block1).await.unwrap();
630
631        store.upload("prefix/xyz", block2).await.unwrap();
632
633        assert_eq!(list_all("", &store).await.len(), 3);
634        assert_eq!(list_all("prefix/", &store).await.len(), 2);
635        let str_list = [String::from("prefix/abc"), String::from("prefix/xyz")];
636
637        store.delete_objects(&str_list).await.unwrap();
638
639        assert!(store.read("prefix/abc/", ..).await.is_err());
640        assert!(store.read("prefix/xyz/", ..).await.is_err());
641        assert_eq!(list_all("", &store).await.len(), 1);
642        assert_eq!(list_all("prefix/", &store).await.len(), 0);
643    }
644}