1use std::fs::File;
5use std::ops::Range;
6use std::sync::Arc;
7
8use crate::local::join_local_io;
9#[cfg(windows)]
10use crate::local::read_exact_at;
11#[cfg(unix)]
12use std::os::unix::fs::FileExt;
13
14use bytes::Bytes;
15use futures::{
16 FutureExt,
17 future::{BoxFuture, Shared},
18 stream::{self, StreamExt},
19};
20use lance_core::deepsize::DeepSizeOf;
21use lance_core::{Error, Result, error::CloneableError};
22use object_store::ObjectStoreExt;
23use object_store::{GetOptions, GetResult, ObjectStore, Result as OSResult, path::Path};
24use tokio::sync::OnceCell;
25use tracing::instrument;
26
27use crate::{
28 object_store::DEFAULT_CLOUD_IO_PARALLELISM,
29 traits::{ByteStream, Reader},
30};
31
32trait StaticGetRange {
33 fn path(&self) -> &Path;
34 fn get_range(&self) -> BoxFuture<'static, OSResult<GetResult>>;
35}
36
37struct GetRequest {
40 object_store: Arc<dyn ObjectStore>,
41 path: Path,
42 options: GetOptions,
43}
44
45impl StaticGetRange for Arc<GetRequest> {
46 fn path(&self) -> &Path {
47 &self.path
48 }
49
50 fn get_range(&self) -> BoxFuture<'static, OSResult<GetResult>> {
51 let store_and_path = self.clone();
52 Box::pin(async move {
53 store_and_path
54 .object_store
55 .get_opts(&store_and_path.path, store_and_path.options.clone())
56 .await
57 })
58 }
59}
60
61#[derive(Debug)]
65pub struct CloudObjectReader {
66 pub object_store: Arc<dyn ObjectStore>,
68 pub path: Path,
70 size: OnceCell<usize>,
72
73 block_size: usize,
74 io_parallelism: usize,
75 download_retry_count: usize,
76}
77
78impl DeepSizeOf for CloudObjectReader {
79 fn deep_size_of_children(&self, context: &mut lance_core::deepsize::Context) -> usize {
80 self.path.as_ref().deep_size_of_children(context)
82 }
83}
84
85impl CloudObjectReader {
86 pub fn new(
88 object_store: Arc<dyn ObjectStore>,
89 path: Path,
90 block_size: usize,
91 known_size: Option<usize>,
92 download_retry_count: usize,
93 ) -> Result<Self> {
94 Ok(Self {
95 object_store,
96 path,
97 size: OnceCell::new_with(known_size),
98 block_size,
99 io_parallelism: DEFAULT_CLOUD_IO_PARALLELISM,
100 download_retry_count,
101 })
102 }
103
104 pub fn with_io_parallelism(mut self, io_parallelism: usize) -> Self {
111 self.io_parallelism = io_parallelism;
112 self
113 }
114}
115
116async fn do_with_retry<'a, O>(f: impl Fn() -> BoxFuture<'a, OSResult<O>> + Clone) -> OSResult<O> {
120 let mut retries = 3;
121 loop {
122 let f = f.clone();
123 match f().await {
124 Ok(val) => return Ok(val),
125 Err(err) => {
126 if retries == 0 {
127 return Err(err);
128 }
129 retries -= 1;
130 }
131 }
132 }
133}
134
135async fn do_get_with_outer_retry(
142 download_retry_count: usize,
143 get_request: Arc<GetRequest>,
144 desc: impl Fn() -> String,
145) -> OSResult<Bytes> {
146 let mut retries = download_retry_count;
147 loop {
148 let get_request_clone = get_request.clone();
149 let get_result = do_with_retry(move || get_request_clone.get_range()).await?;
150 match get_result.bytes().await {
151 Ok(bytes) => return Ok(bytes),
152 Err(err) => {
153 if retries == 0 {
154 log::warn!(
155 "Failed to download {} from {} after {} attempts. This may indicate that cloud storage is overloaded or your timeout settings are too restrictive. Error details: {:?}",
156 desc(),
157 get_request.path(),
158 download_retry_count,
159 err
160 );
161 return Err(err);
162 }
163 log::debug!(
164 "Retrying {} from {} (remaining retries: {}). Error details: {:?}",
165 desc(),
166 get_request.path(),
167 retries,
168 err
169 );
170 retries -= 1;
171 }
172 }
173 }
174}
175
176impl Reader for CloudObjectReader {
177 fn path(&self) -> &Path {
178 &self.path
179 }
180
181 fn block_size(&self) -> usize {
182 self.block_size
183 }
184
185 fn io_parallelism(&self) -> usize {
186 self.io_parallelism
187 }
188
189 fn size(&self) -> BoxFuture<'_, object_store::Result<usize>> {
191 Box::pin(async move {
192 self.size
193 .get_or_try_init(|| async move {
194 let meta =
195 do_with_retry(|| Box::pin(self.object_store.head(&self.path))).await?;
196 Ok(meta.size as usize)
197 })
198 .await
199 .cloned()
200 })
201 }
202
203 #[instrument(level = "debug", skip(self))]
204 fn get_range(&self, range: Range<usize>) -> BoxFuture<'static, OSResult<Bytes>> {
205 let object_store = self.object_store.clone();
206 let path = self.path.clone();
207 let get_range = Range {
208 start: range.start as u64,
209 end: range.end as u64,
210 };
211 Box::pin(async move {
212 let bytes = do_with_retry(move || {
213 let object_store = object_store.clone();
214 let path = path.clone();
215 let get_range = get_range.clone();
216 Box::pin(async move { object_store.get_ranges(&path, &[get_range]).await })
217 })
218 .await?;
219
220 bytes
221 .into_iter()
222 .next()
223 .ok_or_else(|| object_store::Error::Generic {
224 store: "CloudObjectReader",
225 source: "get_ranges returned no bytes".into(),
226 })
227 })
228 }
229
230 #[instrument(level = "debug", skip_all)]
231 fn get_all(&self) -> BoxFuture<'_, OSResult<Bytes>> {
232 let get_request = Arc::new(GetRequest {
233 object_store: self.object_store.clone(),
234 path: self.path.clone(),
235 options: GetOptions::default(),
236 });
237 Box::pin(async move {
238 do_get_with_outer_retry(self.download_retry_count, get_request, || {
239 "read_all".to_string()
240 })
241 .await
242 })
243 }
244
245 fn get_stream(&self) -> BoxFuture<'_, OSResult<ByteStream>> {
246 let get_request = Arc::new(GetRequest {
247 object_store: self.object_store.clone(),
248 path: self.path.clone(),
249 options: GetOptions::default(),
250 });
251 Box::pin(async move {
252 let get_request_clone = get_request.clone();
253 let get_result = do_with_retry(move || get_request_clone.get_range()).await?;
254 Ok(get_result.into_stream())
255 })
256 }
257
258 fn get_range_stream(&self, range: Range<usize>) -> BoxFuture<'_, OSResult<ByteStream>> {
259 let get_request = Arc::new(GetRequest {
260 object_store: self.object_store.clone(),
261 path: self.path.clone(),
262 options: GetOptions {
263 range: Some(
264 Range {
265 start: range.start as u64,
266 end: range.end as u64,
267 }
268 .into(),
269 ),
270 ..Default::default()
271 },
272 });
273 Box::pin(async move {
274 let get_request_clone = get_request.clone();
275 let get_result = do_with_retry(move || get_request_clone.get_range()).await?;
276 Ok(get_result.into_stream())
277 })
278 }
279}
280
281#[derive(Debug)]
282pub struct SmallReaderInner {
283 path: Path,
284 size: usize,
285 state: std::sync::Mutex<SmallReaderState>,
286}
287
288#[derive(Clone, Debug)]
296pub struct SmallReader {
297 inner: Arc<SmallReaderInner>,
298}
299
300enum SmallReaderState {
301 Loading(Shared<BoxFuture<'static, std::result::Result<Bytes, CloneableError>>>),
302 Finished(std::result::Result<Bytes, CloneableError>),
303}
304
305impl std::fmt::Debug for SmallReaderState {
306 fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
307 match self {
308 Self::Loading(_) => write!(f, "Loading"),
309 Self::Finished(Ok(data)) => {
310 write!(f, "Finished({} bytes)", data.len())
311 }
312 Self::Finished(Err(err)) => {
313 write!(f, "Finished({})", err.0)
314 }
315 }
316 }
317}
318
319impl SmallReader {
320 pub fn new(
321 store: Arc<dyn ObjectStore>,
322 path: Path,
323 download_retry_count: usize,
324 size: usize,
325 ) -> Self {
326 let path_ref = path.clone();
327 let state = SmallReaderState::Loading(
328 Box::pin(async move {
329 let object_reader =
330 CloudObjectReader::new(store, path_ref, 0, None, download_retry_count)
331 .map_err(CloneableError)?;
332 object_reader
333 .get_all()
334 .await
335 .map_err(|err| CloneableError(Error::from(err)))
336 })
337 .boxed()
338 .shared(),
339 );
340 Self {
341 inner: Arc::new(SmallReaderInner {
342 path,
343 size,
344 state: std::sync::Mutex::new(state),
345 }),
346 }
347 }
348}
349
350impl SmallReaderInner {
351 async fn wait(&self) -> OSResult<Bytes> {
352 let future = {
353 let state = self.state.lock().unwrap();
354 match &*state {
355 SmallReaderState::Loading(future) => future.clone(),
356 SmallReaderState::Finished(result) => {
357 return result.clone().map_err(|err| err.0.into());
358 }
359 }
360 };
361
362 let result = future.await;
363 let result_to_return = result.clone().map_err(|err| err.0.into());
364 let mut state = self.state.lock().unwrap();
365 if matches!(*state, SmallReaderState::Loading(_)) {
366 *state = SmallReaderState::Finished(result);
367 }
368 result_to_return
369 }
370}
371
372impl Reader for SmallReader {
373 fn path(&self) -> &Path {
374 &self.inner.path
375 }
376
377 fn block_size(&self) -> usize {
378 64 * 1024
379 }
380
381 fn io_parallelism(&self) -> usize {
382 1024
383 }
384
385 fn size(&self) -> BoxFuture<'_, OSResult<usize>> {
387 let size = self.inner.size;
388 Box::pin(async move { Ok(size) })
389 }
390
391 fn get_range(&self, range: Range<usize>) -> BoxFuture<'static, OSResult<Bytes>> {
392 let inner = self.inner.clone();
393 Box::pin(async move {
394 let bytes = inner.wait().await?;
395 let start = range.start;
396 let end = range.end;
397 if start >= bytes.len() || end > bytes.len() {
398 return Err(object_store::Error::Generic {
399 store: "memory",
400 source: format!(
401 "Invalid range {}..{} for object of size {} bytes",
402 start,
403 end,
404 bytes.len()
405 )
406 .into(),
407 });
408 }
409 Ok(bytes.slice(range))
410 })
411 }
412
413 fn get_all(&self) -> BoxFuture<'_, OSResult<Bytes>> {
414 Box::pin(async move { self.inner.wait().await })
415 }
416}
417
418pub(crate) fn stream_local_range(
419 file: Arc<File>,
420 path: Path,
421 io_tracker: Arc<crate::utils::tracking_store::IOTracker>,
422 range: Range<usize>,
423 chunk_size: usize,
424) -> ByteStream {
425 stream::try_unfold(
426 (file, path, io_tracker, range.start, range.end),
427 move |state| async move {
428 let (file, path, io_tracker, start, end) = state;
429 if start >= end {
430 return Ok(None);
431 }
432
433 let next = (start + chunk_size).min(end);
434 let file_clone = file.clone();
435 let path_clone = path.clone();
436 let num_bytes = (next - start) as u64;
437 let metrics = io_tracker.begin_io("get");
438 let result = join_local_io(tokio::task::spawn_blocking(move || {
439 let mut buf = bytes::BytesMut::with_capacity(next - start);
440 unsafe { buf.set_len(next - start) };
442 #[cfg(unix)]
443 file_clone.read_exact_at(buf.as_mut(), start as u64)?;
444 #[cfg(windows)]
445 read_exact_at(file_clone, buf.as_mut(), start as u64)?;
446 Ok::<_, std::io::Error>(buf.freeze())
447 }))
448 .await;
449 metrics.record(&result, num_bytes);
450 let bytes = result?;
451
452 io_tracker.record_read(
453 "get_range_stream",
454 path_clone,
455 num_bytes,
456 Some(start as u64..next as u64),
457 );
458
459 Ok(Some((bytes, (file, path, io_tracker, next, end))))
460 },
461 )
462 .boxed()
463}
464
465impl DeepSizeOf for SmallReader {
466 fn deep_size_of_children(&self, context: &mut lance_core::deepsize::Context) -> usize {
467 let mut size = self.inner.path.as_ref().deep_size_of_children(context);
468
469 if let Ok(guard) = self.inner.state.try_lock()
470 && let SmallReaderState::Finished(Ok(data)) = &*guard
471 {
472 size += data.len();
473 }
474
475 size
476 }
477}