Line data Source code
1 : //! Local filesystem acting as a remote storage.
2 : //! Multiple API users can use the same "storage" of this kind by using different storage roots.
3 : //!
4 : //! This storage used in tests, but can also be used in cases when a certain persistent
5 : //! volume is mounted to the local FS.
6 :
7 : use std::{
8 : collections::HashSet,
9 : io::ErrorKind,
10 : num::NonZeroU32,
11 : time::{Duration, SystemTime, UNIX_EPOCH},
12 : };
13 :
14 : use anyhow::{bail, ensure, Context};
15 : use bytes::Bytes;
16 : use camino::{Utf8Path, Utf8PathBuf};
17 : use futures::stream::Stream;
18 : use tokio::{
19 : fs,
20 : io::{self, AsyncReadExt, AsyncSeekExt, AsyncWriteExt},
21 : };
22 : use tokio_util::{io::ReaderStream, sync::CancellationToken};
23 : use utils::crashsafe::path_with_suffix_extension;
24 :
25 : use crate::{
26 : Download, DownloadError, Listing, ListingMode, RemotePath, TimeTravelError, TimeoutOrCancel,
27 : REMOTE_STORAGE_PREFIX_SEPARATOR,
28 : };
29 :
30 : use super::{RemoteStorage, StorageMetadata};
31 : use crate::Etag;
32 :
33 : const LOCAL_FS_TEMP_FILE_SUFFIX: &str = "___temp";
34 :
35 : #[derive(Debug, Clone)]
36 : pub struct LocalFs {
37 : storage_root: Utf8PathBuf,
38 : timeout: Duration,
39 : }
40 :
41 : impl LocalFs {
42 : /// Attempts to create local FS storage, along with its root directory.
43 : /// Storage root will be created (if does not exist) and transformed into an absolute path (if passed as relative).
44 217 : pub fn new(mut storage_root: Utf8PathBuf, timeout: Duration) -> anyhow::Result<Self> {
45 217 : if !storage_root.exists() {
46 44 : std::fs::create_dir_all(&storage_root).with_context(|| {
47 0 : format!("Failed to create all directories in the given root path {storage_root:?}")
48 44 : })?;
49 173 : }
50 217 : if !storage_root.is_absolute() {
51 157 : storage_root = storage_root.canonicalize_utf8().with_context(|| {
52 0 : format!("Failed to represent path {storage_root:?} as an absolute path")
53 157 : })?;
54 60 : }
55 :
56 217 : Ok(Self {
57 217 : storage_root,
58 217 : timeout,
59 217 : })
60 217 : }
61 :
62 : // mirrors S3Bucket::s3_object_to_relative_path
63 182 : fn local_file_to_relative_path(&self, key: Utf8PathBuf) -> RemotePath {
64 182 : let relative_path = key
65 182 : .strip_prefix(&self.storage_root)
66 182 : .expect("relative path must contain storage_root as prefix");
67 182 : RemotePath(relative_path.into())
68 182 : }
69 :
70 72 : async fn read_storage_metadata(
71 72 : &self,
72 72 : file_path: &Utf8Path,
73 72 : ) -> anyhow::Result<Option<StorageMetadata>> {
74 72 : let metadata_path = storage_metadata_path(file_path);
75 72 : if metadata_path.exists() && metadata_path.is_file() {
76 8 : let metadata_string = fs::read_to_string(&metadata_path).await.with_context(|| {
77 0 : format!("Failed to read metadata from the local storage at '{metadata_path}'")
78 8 : })?;
79 :
80 8 : serde_json::from_str(&metadata_string)
81 8 : .with_context(|| {
82 0 : format!(
83 0 : "Failed to deserialize metadata from the local storage at '{metadata_path}'",
84 0 : )
85 8 : })
86 8 : .map(|metadata| Some(StorageMetadata(metadata)))
87 : } else {
88 64 : Ok(None)
89 : }
90 72 : }
91 :
92 : #[cfg(test)]
93 12 : async fn list_all(&self) -> anyhow::Result<Vec<RemotePath>> {
94 12 : use std::{future::Future, pin::Pin};
95 36 : fn get_all_files<'a, P>(
96 36 : directory_path: P,
97 36 : ) -> Pin<Box<dyn Future<Output = anyhow::Result<Vec<Utf8PathBuf>>> + Send + Sync + 'a>>
98 36 : where
99 36 : P: AsRef<Utf8Path> + Send + Sync + 'a,
100 36 : {
101 36 : Box::pin(async move {
102 36 : let directory_path = directory_path.as_ref();
103 36 : if directory_path.exists() {
104 36 : if directory_path.is_dir() {
105 36 : let mut paths = Vec::new();
106 36 : let mut dir_contents = fs::read_dir(directory_path).await?;
107 72 : while let Some(dir_entry) = dir_contents.next_entry().await? {
108 36 : let file_type = dir_entry.file_type().await?;
109 36 : let entry_path =
110 36 : Utf8PathBuf::from_path_buf(dir_entry.path()).map_err(|pb| {
111 0 : anyhow::Error::msg(format!(
112 0 : "non-Unicode path: {}",
113 0 : pb.to_string_lossy()
114 0 : ))
115 36 : })?;
116 36 : if file_type.is_symlink() {
117 12 : tracing::debug!("{entry_path:?} is a symlink, skipping")
118 36 : } else if file_type.is_dir() {
119 36 : paths.extend(get_all_files(&entry_path).await?.into_iter())
120 12 : } else {
121 12 : paths.push(entry_path);
122 12 : }
123 12 : }
124 36 : Ok(paths)
125 12 : } else {
126 12 : bail!("Path {directory_path:?} is not a directory")
127 12 : }
128 12 : } else {
129 12 : Ok(Vec::new())
130 12 : }
131 36 : })
132 36 : }
133 12 :
134 12 : Ok(get_all_files(&self.storage_root)
135 36 : .await?
136 12 : .into_iter()
137 12 : .map(|path| {
138 12 : path.strip_prefix(&self.storage_root)
139 12 : .context("Failed to strip storage root prefix")
140 12 : .and_then(RemotePath::new)
141 12 : .expect(
142 12 : "We list files for storage root, hence should be able to remote the prefix",
143 12 : )
144 12 : })
145 12 : .collect())
146 12 : }
147 :
148 : // recursively lists all files in a directory,
149 : // mirroring the `list_files` for `s3_bucket`
150 187 : async fn list_recursive(&self, folder: Option<&RemotePath>) -> anyhow::Result<Vec<RemotePath>> {
151 187 : let full_path = match folder {
152 179 : Some(folder) => folder.with_base(&self.storage_root),
153 8 : None => self.storage_root.clone(),
154 : };
155 :
156 : // If we were given a directory, we may use it as our starting point.
157 : // Otherwise, we must go up to the first ancestor dir that exists. This is because
158 : // S3 object list prefixes can be arbitrary strings, but when reading
159 : // the local filesystem we need a directory to start calling read_dir on.
160 187 : let mut initial_dir = full_path.clone();
161 187 :
162 187 : // If there's no trailing slash, we have to start looking from one above: even if
163 187 : // `initial_dir` is a directory, we should still list any prefixes in the parent
164 187 : // that start with the same string.
165 187 : if !full_path.to_string().ends_with('/') {
166 26 : initial_dir.pop();
167 161 : }
168 :
169 646 : loop {
170 646 : // Did we make it to the root?
171 646 : if initial_dir.parent().is_none() {
172 0 : anyhow::bail!("list_files: failed to find valid ancestor dir for {full_path}");
173 646 : }
174 646 :
175 646 : match fs::metadata(initial_dir.clone()).await {
176 187 : Ok(meta) if meta.is_dir() => {
177 187 : // We found a directory, break
178 187 : break;
179 : }
180 0 : Ok(_meta) => {
181 0 : // It's not a directory: strip back to the parent
182 0 : initial_dir.pop();
183 0 : }
184 459 : Err(e) if e.kind() == ErrorKind::NotFound => {
185 459 : // It's not a file that exists: strip the prefix back to the parent directory
186 459 : initial_dir.pop();
187 459 : }
188 0 : Err(e) => {
189 0 : // Unexpected I/O error
190 0 : anyhow::bail!(e)
191 : }
192 : }
193 : }
194 : // Note that Utf8PathBuf starts_with only considers full path segments, but
195 : // object prefixes are arbitrary strings, so we need the strings for doing
196 : // starts_with later.
197 187 : let prefix = full_path.as_str();
198 187 :
199 187 : let mut files = vec![];
200 187 : let mut directory_queue = vec![initial_dir];
201 448 : while let Some(cur_folder) = directory_queue.pop() {
202 261 : let mut entries = cur_folder.read_dir_utf8()?;
203 743 : while let Some(Ok(entry)) = entries.next() {
204 482 : let file_name = entry.file_name();
205 482 : let full_file_name = cur_folder.join(file_name);
206 482 : if full_file_name.as_str().starts_with(prefix) {
207 182 : let file_remote_path = self.local_file_to_relative_path(full_file_name.clone());
208 182 : files.push(file_remote_path);
209 182 : if full_file_name.is_dir() {
210 74 : directory_queue.push(full_file_name);
211 108 : }
212 300 : }
213 : }
214 : }
215 :
216 187 : Ok(files)
217 187 : }
218 :
219 2729 : async fn upload0(
220 2729 : &self,
221 2729 : data: impl Stream<Item = std::io::Result<Bytes>> + Send + Sync,
222 2729 : data_size_bytes: usize,
223 2729 : to: &RemotePath,
224 2729 : metadata: Option<StorageMetadata>,
225 2729 : cancel: &CancellationToken,
226 2729 : ) -> anyhow::Result<()> {
227 2729 : let target_file_path = to.with_base(&self.storage_root);
228 2729 : create_target_directory(&target_file_path).await?;
229 : // We need this dance with sort of durable rename (without fsyncs)
230 : // to prevent partial uploads. This was really hit when pageserver shutdown
231 : // cancelled the upload and partial file was left on the fs
232 : // NOTE: Because temp file suffix always the same this operation is racy.
233 : // Two concurrent operations can lead to the following sequence:
234 : // T1: write(temp)
235 : // T2: write(temp) -> overwrites the content
236 : // T1: rename(temp, dst) -> succeeds
237 : // T2: rename(temp, dst) -> fails, temp no longet exists
238 : // This can be solved by supplying unique temp suffix every time, but this situation
239 : // is not normal in the first place, the error can help (and helped at least once)
240 : // to discover bugs in upper level synchronization.
241 2726 : let temp_file_path =
242 2726 : path_with_suffix_extension(&target_file_path, LOCAL_FS_TEMP_FILE_SUFFIX);
243 2714 : let mut destination = io::BufWriter::new(
244 2726 : fs::OpenOptions::new()
245 2726 : .write(true)
246 2726 : .create(true)
247 2726 : .truncate(true)
248 2726 : .open(&temp_file_path)
249 2448 : .await
250 2714 : .with_context(|| {
251 0 : format!("Failed to open target fs destination at '{target_file_path}'")
252 2714 : })?,
253 : );
254 :
255 2714 : let from_size_bytes = data_size_bytes as u64;
256 2714 : let data = tokio_util::io::StreamReader::new(data);
257 2714 : let data = std::pin::pin!(data);
258 2714 : let mut buffer_to_read = data.take(from_size_bytes);
259 2714 :
260 2714 : // alternatively we could just write the bytes to a file, but local_fs is a testing utility
261 2714 : let copy = io::copy_buf(&mut buffer_to_read, &mut destination);
262 :
263 2701 : let bytes_read = tokio::select! {
264 : biased;
265 : _ = cancel.cancelled() => {
266 : let file = destination.into_inner();
267 : // wait for the inflight operation(s) to complete so that there could be a next
268 : // attempt right away and our writes are not directed to their file.
269 : file.into_std().await;
270 :
271 : // TODO: leave the temp or not? leaving is probably less racy. enabled truncate at
272 : // least.
273 : fs::remove_file(temp_file_path).await.context("remove temp_file_path after cancellation or timeout")?;
274 : return Err(TimeoutOrCancel::Cancel.into());
275 : }
276 : read = copy => read,
277 : };
278 :
279 2701 : let bytes_read =
280 2701 : bytes_read.with_context(|| {
281 0 : format!(
282 0 : "Failed to upload file (write temp) to the local storage at '{temp_file_path}'",
283 0 : )
284 2701 : })?;
285 :
286 2701 : if bytes_read < from_size_bytes {
287 4 : bail!("Provided stream was shorter than expected: {bytes_read} vs {from_size_bytes} bytes");
288 2697 : }
289 2697 : // Check if there is any extra data after the given size.
290 2697 : let mut from = buffer_to_read.into_inner();
291 2697 : let extra_read = from.read(&mut [1]).await?;
292 2695 : ensure!(
293 2695 : extra_read == 0,
294 8 : "Provided stream was larger than expected: expected {from_size_bytes} bytes",
295 : );
296 :
297 2687 : destination.flush().await.with_context(|| {
298 0 : format!(
299 0 : "Failed to upload (flush temp) file to the local storage at '{temp_file_path}'",
300 0 : )
301 2687 : })?;
302 :
303 2687 : fs::rename(temp_file_path, &target_file_path)
304 2391 : .await
305 2681 : .with_context(|| {
306 0 : format!(
307 0 : "Failed to upload (rename) file to the local storage at '{target_file_path}'",
308 0 : )
309 2681 : })?;
310 :
311 2681 : if let Some(storage_metadata) = metadata {
312 : // FIXME: we must not be using metadata much, since this would forget the old metadata
313 : // for new writes? or perhaps metadata is sticky; could consider removing if it's never
314 : // used.
315 4 : let storage_metadata_path = storage_metadata_path(&target_file_path);
316 4 : fs::write(
317 4 : &storage_metadata_path,
318 4 : serde_json::to_string(&storage_metadata.0)
319 4 : .context("Failed to serialize storage metadata as json")?,
320 : )
321 4 : .await
322 4 : .with_context(|| {
323 0 : format!(
324 0 : "Failed to write metadata to the local storage at '{storage_metadata_path}'",
325 0 : )
326 4 : })?;
327 2677 : }
328 :
329 2681 : Ok(())
330 2697 : }
331 : }
332 :
333 : impl RemoteStorage for LocalFs {
334 187 : async fn list(
335 187 : &self,
336 187 : prefix: Option<&RemotePath>,
337 187 : mode: ListingMode,
338 187 : max_keys: Option<NonZeroU32>,
339 187 : cancel: &CancellationToken,
340 187 : ) -> Result<Listing, DownloadError> {
341 187 : let op = async {
342 187 : let mut result = Listing::default();
343 :
344 : // Filter out directories: in S3 directories don't exist, only the keys within them do.
345 187 : let keys = self
346 187 : .list_recursive(prefix)
347 636 : .await
348 187 : .map_err(DownloadError::Other)?;
349 187 : let keys = keys
350 187 : .into_iter()
351 187 : .filter(|k| {
352 182 : let path = k.with_base(&self.storage_root);
353 182 : !path.is_dir()
354 187 : })
355 187 : .collect();
356 187 :
357 187 : if let ListingMode::NoDelimiter = mode {
358 10 : result.keys = keys;
359 10 : } else {
360 177 : let mut prefixes = HashSet::new();
361 255 : for key in keys {
362 : // If the part after the prefix includes a "/", take only the first part and put it in `prefixes`.
363 78 : let relative_key = if let Some(prefix) = prefix {
364 66 : let mut prefix = prefix.clone();
365 66 : // We only strip the dirname of the prefix, so that when we strip it from the start of keys we
366 66 : // end up with full file/dir names.
367 66 : let prefix_full_local_path = prefix.with_base(&self.storage_root);
368 66 : let has_slash = prefix.0.to_string().ends_with('/');
369 66 : let strip_prefix = if prefix_full_local_path.is_dir() && has_slash {
370 34 : prefix
371 : } else {
372 32 : prefix.0.pop();
373 32 : prefix
374 : };
375 :
376 66 : RemotePath::new(key.strip_prefix(&strip_prefix).unwrap()).unwrap()
377 : } else {
378 12 : key
379 : };
380 :
381 78 : let relative_key = format!("{}", relative_key);
382 78 : if relative_key.contains(REMOTE_STORAGE_PREFIX_SEPARATOR) {
383 74 : let first_part = relative_key
384 74 : .split(REMOTE_STORAGE_PREFIX_SEPARATOR)
385 74 : .next()
386 74 : .unwrap()
387 74 : .to_owned();
388 74 : prefixes.insert(first_part);
389 74 : } else {
390 4 : result
391 4 : .keys
392 4 : .push(RemotePath::from_string(&relative_key).unwrap());
393 4 : }
394 : }
395 177 : result.prefixes = prefixes
396 177 : .into_iter()
397 177 : .map(|s| RemotePath::from_string(&s).unwrap())
398 177 : .collect();
399 : }
400 :
401 187 : if let Some(max_keys) = max_keys {
402 0 : result.keys.truncate(max_keys.get() as usize);
403 187 : }
404 187 : Ok(result)
405 187 : };
406 :
407 187 : let timeout = async {
408 506 : tokio::time::sleep(self.timeout).await;
409 0 : Err(DownloadError::Timeout)
410 0 : };
411 :
412 187 : let cancelled = async {
413 563 : cancel.cancelled().await;
414 0 : Err(DownloadError::Cancelled)
415 0 : };
416 :
417 : tokio::select! {
418 : res = op => res,
419 : res = timeout => res,
420 : res = cancelled => res,
421 : }
422 187 : }
423 :
424 2729 : async fn upload(
425 2729 : &self,
426 2729 : data: impl Stream<Item = std::io::Result<Bytes>> + Send + Sync,
427 2729 : data_size_bytes: usize,
428 2729 : to: &RemotePath,
429 2729 : metadata: Option<StorageMetadata>,
430 2729 : cancel: &CancellationToken,
431 2729 : ) -> anyhow::Result<()> {
432 2729 : let cancel = cancel.child_token();
433 2729 :
434 2729 : let op = self.upload0(data, data_size_bytes, to, metadata, &cancel);
435 2729 : let mut op = std::pin::pin!(op);
436 :
437 : // race the upload0 to the timeout; if it goes over, do a graceful shutdown
438 2697 : let (res, timeout) = tokio::select! {
439 : res = &mut op => (res, false),
440 : _ = tokio::time::sleep(self.timeout) => {
441 : cancel.cancel();
442 : (op.await, true)
443 : }
444 : };
445 :
446 16 : match res {
447 16 : Err(e) if timeout && TimeoutOrCancel::caused_by_cancel(&e) => {
448 0 : // we caused this cancel (or they happened simultaneously) -- swap it out to
449 0 : // Timeout
450 0 : Err(TimeoutOrCancel::Timeout.into())
451 : }
452 2697 : res => res,
453 : }
454 2697 : }
455 :
456 70 : async fn download(
457 70 : &self,
458 70 : from: &RemotePath,
459 70 : cancel: &CancellationToken,
460 70 : ) -> Result<Download, DownloadError> {
461 70 : let target_path = from.with_base(&self.storage_root);
462 :
463 70 : let file_metadata = file_metadata(&target_path).await?;
464 :
465 52 : let source = ReaderStream::new(
466 52 : fs::OpenOptions::new()
467 52 : .read(true)
468 52 : .open(&target_path)
469 50 : .await
470 52 : .with_context(|| {
471 0 : format!("Failed to open source file {target_path:?} to use in the download")
472 52 : })
473 52 : .map_err(DownloadError::Other)?,
474 : );
475 :
476 52 : let metadata = self
477 52 : .read_storage_metadata(&target_path)
478 4 : .await
479 52 : .map_err(DownloadError::Other)?;
480 :
481 52 : let cancel_or_timeout = crate::support::cancel_or_timeout(self.timeout, cancel.clone());
482 52 : let source = crate::support::DownloadStream::new(cancel_or_timeout, source);
483 52 :
484 52 : let etag = mock_etag(&file_metadata);
485 52 : Ok(Download {
486 52 : metadata,
487 52 : last_modified: file_metadata
488 52 : .modified()
489 52 : .map_err(|e| DownloadError::Other(anyhow::anyhow!(e).context("Reading mtime")))?,
490 52 : etag,
491 52 : download_stream: Box::pin(source),
492 : })
493 70 : }
494 :
495 28 : async fn download_byte_range(
496 28 : &self,
497 28 : from: &RemotePath,
498 28 : start_inclusive: u64,
499 28 : end_exclusive: Option<u64>,
500 28 : cancel: &CancellationToken,
501 28 : ) -> Result<Download, DownloadError> {
502 28 : if let Some(end_exclusive) = end_exclusive {
503 20 : if end_exclusive <= start_inclusive {
504 4 : return Err(DownloadError::Other(anyhow::anyhow!("Invalid range, start ({start_inclusive}) is not less than end_exclusive ({end_exclusive:?})")));
505 16 : };
506 16 : if start_inclusive == end_exclusive.saturating_sub(1) {
507 4 : return Err(DownloadError::Other(anyhow::anyhow!("Invalid range, start ({start_inclusive}) and end_exclusive ({end_exclusive:?}) difference is zero bytes")));
508 12 : }
509 8 : }
510 :
511 20 : let target_path = from.with_base(&self.storage_root);
512 20 : let file_metadata = file_metadata(&target_path).await?;
513 20 : let mut source = tokio::fs::OpenOptions::new()
514 20 : .read(true)
515 20 : .open(&target_path)
516 16 : .await
517 20 : .with_context(|| {
518 0 : format!("Failed to open source file {target_path:?} to use in the download")
519 20 : })
520 20 : .map_err(DownloadError::Other)?;
521 :
522 20 : let len = source
523 20 : .metadata()
524 16 : .await
525 20 : .context("query file length")
526 20 : .map_err(DownloadError::Other)?
527 20 : .len();
528 20 :
529 20 : source
530 20 : .seek(io::SeekFrom::Start(start_inclusive))
531 16 : .await
532 20 : .context("Failed to seek to the range start in a local storage file")
533 20 : .map_err(DownloadError::Other)?;
534 :
535 20 : let metadata = self
536 20 : .read_storage_metadata(&target_path)
537 4 : .await
538 20 : .map_err(DownloadError::Other)?;
539 :
540 20 : let source = source.take(end_exclusive.unwrap_or(len) - start_inclusive);
541 20 : let source = ReaderStream::new(source);
542 20 :
543 20 : let cancel_or_timeout = crate::support::cancel_or_timeout(self.timeout, cancel.clone());
544 20 : let source = crate::support::DownloadStream::new(cancel_or_timeout, source);
545 20 :
546 20 : let etag = mock_etag(&file_metadata);
547 20 : Ok(Download {
548 20 : metadata,
549 20 : last_modified: file_metadata
550 20 : .modified()
551 20 : .map_err(|e| DownloadError::Other(anyhow::anyhow!(e).context("Reading mtime")))?,
552 20 : etag,
553 20 : download_stream: Box::pin(source),
554 : })
555 28 : }
556 :
557 16 : async fn delete(&self, path: &RemotePath, _cancel: &CancellationToken) -> anyhow::Result<()> {
558 16 : let file_path = path.with_base(&self.storage_root);
559 16 : match fs::remove_file(&file_path).await {
560 12 : Ok(()) => Ok(()),
561 : // The file doesn't exist. This shouldn't yield an error to mirror S3's behaviour.
562 : // See https://docs.aws.amazon.com/AmazonS3/latest/API/API_DeleteObject.html
563 : // > If there isn't a null version, Amazon S3 does not remove any objects but will still respond that the command was successful.
564 4 : Err(e) if e.kind() == ErrorKind::NotFound => Ok(()),
565 0 : Err(e) => Err(anyhow::anyhow!(e)),
566 : }
567 16 : }
568 :
569 6 : async fn delete_objects<'a>(
570 6 : &self,
571 6 : paths: &'a [RemotePath],
572 6 : cancel: &CancellationToken,
573 6 : ) -> anyhow::Result<()> {
574 12 : for path in paths {
575 6 : self.delete(path, cancel).await?
576 : }
577 6 : Ok(())
578 6 : }
579 :
580 0 : async fn copy(
581 0 : &self,
582 0 : from: &RemotePath,
583 0 : to: &RemotePath,
584 0 : _cancel: &CancellationToken,
585 0 : ) -> anyhow::Result<()> {
586 0 : let from_path = from.with_base(&self.storage_root);
587 0 : let to_path = to.with_base(&self.storage_root);
588 0 : create_target_directory(&to_path).await?;
589 0 : fs::copy(&from_path, &to_path).await.with_context(|| {
590 0 : format!(
591 0 : "Failed to copy file from '{from_path}' to '{to_path}'",
592 0 : from_path = from_path,
593 0 : to_path = to_path
594 0 : )
595 0 : })?;
596 0 : Ok(())
597 0 : }
598 :
599 0 : async fn time_travel_recover(
600 0 : &self,
601 0 : _prefix: Option<&RemotePath>,
602 0 : _timestamp: SystemTime,
603 0 : _done_if_after: SystemTime,
604 0 : _cancel: &CancellationToken,
605 0 : ) -> Result<(), TimeTravelError> {
606 0 : Err(TimeTravelError::Unimplemented)
607 0 : }
608 : }
609 :
610 76 : fn storage_metadata_path(original_path: &Utf8Path) -> Utf8PathBuf {
611 76 : path_with_suffix_extension(original_path, "metadata")
612 76 : }
613 :
614 2729 : async fn create_target_directory(target_file_path: &Utf8Path) -> anyhow::Result<()> {
615 2729 : let target_dir = match target_file_path.parent() {
616 2729 : Some(parent_dir) => parent_dir,
617 0 : None => bail!("File path '{target_file_path}' has no parent directory"),
618 : };
619 2729 : if !target_dir.exists() {
620 381 : fs::create_dir_all(target_dir).await?;
621 2348 : }
622 2726 : Ok(())
623 2726 : }
624 :
625 90 : async fn file_metadata(file_path: &Utf8Path) -> Result<std::fs::Metadata, DownloadError> {
626 90 : tokio::fs::metadata(&file_path).await.map_err(|e| {
627 18 : if e.kind() == ErrorKind::NotFound {
628 18 : DownloadError::NotFound
629 : } else {
630 0 : DownloadError::BadInput(e.into())
631 : }
632 90 : })
633 90 : }
634 :
635 : // Use mtime as stand-in for ETag. We could calculate a meaningful one by md5'ing the contents of files we
636 : // read, but that's expensive and the local_fs test helper's whole reason for existence is to run small tests
637 : // quickly, with less overhead than using a mock S3 server.
638 72 : fn mock_etag(meta: &std::fs::Metadata) -> Etag {
639 72 : let mtime = meta.modified().expect("Filesystem mtime missing");
640 72 : format!("{}", mtime.duration_since(UNIX_EPOCH).unwrap().as_millis()).into()
641 72 : }
642 :
643 : #[cfg(test)]
644 : mod fs_tests {
645 : use super::*;
646 :
647 : use camino_tempfile::tempdir;
648 : use std::{collections::HashMap, io::Write};
649 :
650 12 : async fn read_and_check_metadata(
651 12 : storage: &LocalFs,
652 12 : remote_storage_path: &RemotePath,
653 12 : expected_metadata: Option<&StorageMetadata>,
654 12 : ) -> anyhow::Result<String> {
655 12 : let cancel = CancellationToken::new();
656 12 : let download = storage
657 12 : .download(remote_storage_path, &cancel)
658 27 : .await
659 12 : .map_err(|e| anyhow::anyhow!("Download failed: {e}"))?;
660 12 : ensure!(
661 12 : download.metadata.as_ref() == expected_metadata,
662 0 : "Unexpected metadata returned for the downloaded file"
663 : );
664 :
665 22 : let contents = aggregate(download.download_stream).await?;
666 :
667 12 : String::from_utf8(contents).map_err(anyhow::Error::new)
668 12 : }
669 :
670 : #[tokio::test]
671 4 : async fn upload_file() -> anyhow::Result<()> {
672 4 : let (storage, cancel) = create_storage()?;
673 4 :
674 21 : let target_path_1 = upload_dummy_file(&storage, "upload_1", None, &cancel).await?;
675 4 : assert_eq!(
676 12 : storage.list_all().await?,
677 4 : vec![target_path_1.clone()],
678 4 : "Should list a single file after first upload"
679 4 : );
680 4 :
681 24 : let target_path_2 = upload_dummy_file(&storage, "upload_2", None, &cancel).await?;
682 4 : assert_eq!(
683 12 : list_files_sorted(&storage).await?,
684 4 : vec![target_path_1.clone(), target_path_2.clone()],
685 4 : "Should list a two different files after second upload"
686 4 : );
687 4 :
688 4 : Ok(())
689 4 : }
690 :
691 : #[tokio::test]
692 4 : async fn upload_file_negatives() -> anyhow::Result<()> {
693 4 : let (storage, cancel) = create_storage()?;
694 4 :
695 4 : let id = RemotePath::new(Utf8Path::new("dummy"))?;
696 4 : let content = Bytes::from_static(b"12345");
697 16 : let content = move || futures::stream::once(futures::future::ready(Ok(content.clone())));
698 4 :
699 4 : // Check that you get an error if the size parameter doesn't match the actual
700 4 : // size of the stream.
701 4 : storage
702 4 : .upload(content(), 0, &id, None, &cancel)
703 4 : .await
704 4 : .expect_err("upload with zero size succeeded");
705 4 : storage
706 4 : .upload(content(), 4, &id, None, &cancel)
707 8 : .await
708 4 : .expect_err("upload with too short size succeeded");
709 4 : storage
710 4 : .upload(content(), 6, &id, None, &cancel)
711 8 : .await
712 4 : .expect_err("upload with too large size succeeded");
713 4 :
714 4 : // Correct size is 5, this should succeed.
715 12 : storage.upload(content(), 5, &id, None, &cancel).await?;
716 4 :
717 4 : Ok(())
718 4 : }
719 :
720 44 : fn create_storage() -> anyhow::Result<(LocalFs, CancellationToken)> {
721 44 : let storage_root = tempdir()?.path().to_path_buf();
722 44 : LocalFs::new(storage_root, Duration::from_secs(120)).map(|s| (s, CancellationToken::new()))
723 44 : }
724 :
725 : #[tokio::test]
726 4 : async fn download_file() -> anyhow::Result<()> {
727 4 : let (storage, cancel) = create_storage()?;
728 4 : let upload_name = "upload_1";
729 22 : let upload_target = upload_dummy_file(&storage, upload_name, None, &cancel).await?;
730 4 :
731 16 : let contents = read_and_check_metadata(&storage, &upload_target, None).await?;
732 4 : assert_eq!(
733 4 : dummy_contents(upload_name),
734 4 : contents,
735 4 : "We should upload and download the same contents"
736 4 : );
737 4 :
738 4 : let non_existing_path = "somewhere/else";
739 4 : match storage.download(&RemotePath::new(Utf8Path::new(non_existing_path))?, &cancel).await {
740 4 : Err(DownloadError::NotFound) => {} // Should get NotFound for non existing keys
741 4 : other => panic!("Should get a NotFound error when downloading non-existing storage files, but got: {other:?}"),
742 4 : }
743 4 : Ok(())
744 4 : }
745 :
746 : #[tokio::test]
747 4 : async fn download_file_range_positive() -> anyhow::Result<()> {
748 4 : let (storage, cancel) = create_storage()?;
749 4 : let upload_name = "upload_1";
750 21 : let upload_target = upload_dummy_file(&storage, upload_name, None, &cancel).await?;
751 4 :
752 4 : let full_range_download_contents =
753 13 : read_and_check_metadata(&storage, &upload_target, None).await?;
754 4 : assert_eq!(
755 4 : dummy_contents(upload_name),
756 4 : full_range_download_contents,
757 4 : "Download full range should return the whole upload"
758 4 : );
759 4 :
760 4 : let uploaded_bytes = dummy_contents(upload_name).into_bytes();
761 4 : let (first_part_local, second_part_local) = uploaded_bytes.split_at(3);
762 4 :
763 4 : let first_part_download = storage
764 4 : .download_byte_range(
765 4 : &upload_target,
766 4 : 0,
767 4 : Some(first_part_local.len() as u64),
768 4 : &cancel,
769 4 : )
770 12 : .await?;
771 4 : assert!(
772 4 : first_part_download.metadata.is_none(),
773 4 : "No metadata should be returned for no metadata upload"
774 4 : );
775 4 :
776 4 : let first_part_remote = aggregate(first_part_download.download_stream).await?;
777 4 : assert_eq!(
778 4 : first_part_local, first_part_remote,
779 4 : "First part bytes should be returned when requested"
780 4 : );
781 4 :
782 4 : let second_part_download = storage
783 4 : .download_byte_range(
784 4 : &upload_target,
785 4 : first_part_local.len() as u64,
786 4 : Some((first_part_local.len() + second_part_local.len()) as u64),
787 4 : &cancel,
788 4 : )
789 12 : .await?;
790 4 : assert!(
791 4 : second_part_download.metadata.is_none(),
792 4 : "No metadata should be returned for no metadata upload"
793 4 : );
794 4 :
795 4 : let second_part_remote = aggregate(second_part_download.download_stream).await?;
796 4 : assert_eq!(
797 4 : second_part_local, second_part_remote,
798 4 : "Second part bytes should be returned when requested"
799 4 : );
800 4 :
801 4 : let suffix_bytes = storage
802 4 : .download_byte_range(&upload_target, 13, None, &cancel)
803 12 : .await?
804 4 : .download_stream;
805 4 : let suffix_bytes = aggregate(suffix_bytes).await?;
806 4 : let suffix = std::str::from_utf8(&suffix_bytes)?;
807 4 : assert_eq!(upload_name, suffix);
808 4 :
809 4 : let all_bytes = storage
810 4 : .download_byte_range(&upload_target, 0, None, &cancel)
811 12 : .await?
812 4 : .download_stream;
813 4 : let all_bytes = aggregate(all_bytes).await?;
814 4 : let all_bytes = std::str::from_utf8(&all_bytes)?;
815 4 : assert_eq!(dummy_contents("upload_1"), all_bytes);
816 4 :
817 4 : Ok(())
818 4 : }
819 :
820 : #[tokio::test]
821 4 : async fn download_file_range_negative() -> anyhow::Result<()> {
822 4 : let (storage, cancel) = create_storage()?;
823 4 : let upload_name = "upload_1";
824 24 : let upload_target = upload_dummy_file(&storage, upload_name, None, &cancel).await?;
825 4 :
826 4 : let start = 1_000_000_000;
827 4 : let end = start + 1;
828 4 : match storage
829 4 : .download_byte_range(
830 4 : &upload_target,
831 4 : start,
832 4 : Some(end), // exclusive end
833 4 : &cancel,
834 4 : )
835 4 : .await
836 4 : {
837 4 : Ok(_) => panic!("Should not allow downloading wrong ranges"),
838 4 : Err(e) => {
839 4 : let error_string = e.to_string();
840 4 : assert!(error_string.contains("zero bytes"));
841 4 : assert!(error_string.contains(&start.to_string()));
842 4 : assert!(error_string.contains(&end.to_string()));
843 4 : }
844 4 : }
845 4 :
846 4 : let start = 10000;
847 4 : let end = 234;
848 4 : assert!(start > end, "Should test an incorrect range");
849 4 : match storage
850 4 : .download_byte_range(&upload_target, start, Some(end), &cancel)
851 4 : .await
852 4 : {
853 4 : Ok(_) => panic!("Should not allow downloading wrong ranges"),
854 4 : Err(e) => {
855 4 : let error_string = e.to_string();
856 4 : assert!(error_string.contains("Invalid range"));
857 4 : assert!(error_string.contains(&start.to_string()));
858 4 : assert!(error_string.contains(&end.to_string()));
859 4 : }
860 4 : }
861 4 :
862 4 : Ok(())
863 4 : }
864 :
865 : #[tokio::test]
866 4 : async fn delete_file() -> anyhow::Result<()> {
867 4 : let (storage, cancel) = create_storage()?;
868 4 : let upload_name = "upload_1";
869 24 : let upload_target = upload_dummy_file(&storage, upload_name, None, &cancel).await?;
870 4 :
871 4 : storage.delete(&upload_target, &cancel).await?;
872 12 : assert!(storage.list_all().await?.is_empty());
873 4 :
874 4 : storage
875 4 : .delete(&upload_target, &cancel)
876 4 : .await
877 4 : .expect("Should allow deleting non-existing storage files");
878 4 :
879 4 : Ok(())
880 4 : }
881 :
882 : #[tokio::test]
883 4 : async fn file_with_metadata() -> anyhow::Result<()> {
884 4 : let (storage, cancel) = create_storage()?;
885 4 : let upload_name = "upload_1";
886 4 : let metadata = StorageMetadata(HashMap::from([
887 4 : ("one".to_string(), "1".to_string()),
888 4 : ("two".to_string(), "2".to_string()),
889 4 : ]));
890 4 : let upload_target =
891 26 : upload_dummy_file(&storage, upload_name, Some(metadata.clone()), &cancel).await?;
892 4 :
893 4 : let full_range_download_contents =
894 20 : read_and_check_metadata(&storage, &upload_target, Some(&metadata)).await?;
895 4 : assert_eq!(
896 4 : dummy_contents(upload_name),
897 4 : full_range_download_contents,
898 4 : "We should upload and download the same contents"
899 4 : );
900 4 :
901 4 : let uploaded_bytes = dummy_contents(upload_name).into_bytes();
902 4 : let (first_part_local, _) = uploaded_bytes.split_at(3);
903 4 :
904 4 : let partial_download_with_metadata = storage
905 4 : .download_byte_range(
906 4 : &upload_target,
907 4 : 0,
908 4 : Some(first_part_local.len() as u64),
909 4 : &cancel,
910 4 : )
911 20 : .await?;
912 4 : let first_part_remote = aggregate(partial_download_with_metadata.download_stream).await?;
913 4 : assert_eq!(
914 4 : first_part_local,
915 4 : first_part_remote.as_slice(),
916 4 : "First part bytes should be returned when requested"
917 4 : );
918 4 :
919 4 : assert_eq!(
920 4 : partial_download_with_metadata.metadata,
921 4 : Some(metadata),
922 4 : "We should get the same metadata back for partial download"
923 4 : );
924 4 :
925 4 : Ok(())
926 4 : }
927 :
928 : #[tokio::test]
929 4 : async fn list() -> anyhow::Result<()> {
930 4 : // No delimiter: should recursively list everything
931 4 : let (storage, cancel) = create_storage()?;
932 22 : let child = upload_dummy_file(&storage, "grandparent/parent/child", None, &cancel).await?;
933 4 : let child_sibling =
934 24 : upload_dummy_file(&storage, "grandparent/parent/child_sibling", None, &cancel).await?;
935 24 : let uncle = upload_dummy_file(&storage, "grandparent/uncle", None, &cancel).await?;
936 4 :
937 4 : let listing = storage
938 4 : .list(None, ListingMode::NoDelimiter, None, &cancel)
939 4 : .await?;
940 4 : assert!(listing.prefixes.is_empty());
941 4 : assert_eq!(
942 4 : listing.keys.into_iter().collect::<HashSet<_>>(),
943 4 : HashSet::from([uncle.clone(), child.clone(), child_sibling.clone()])
944 4 : );
945 4 :
946 4 : // Delimiter: should only go one deep
947 4 : let listing = storage
948 4 : .list(None, ListingMode::WithDelimiter, None, &cancel)
949 4 : .await?;
950 4 :
951 4 : assert_eq!(
952 4 : listing.prefixes,
953 4 : [RemotePath::from_string("timelines").unwrap()].to_vec()
954 4 : );
955 4 : assert!(listing.keys.is_empty());
956 4 :
957 4 : // Delimiter & prefix with a trailing slash
958 4 : let listing = storage
959 4 : .list(
960 4 : Some(&RemotePath::from_string("timelines/some_timeline/grandparent/").unwrap()),
961 4 : ListingMode::WithDelimiter,
962 4 : None,
963 4 : &cancel,
964 4 : )
965 4 : .await?;
966 4 : assert_eq!(
967 4 : listing.keys,
968 4 : [RemotePath::from_string("uncle").unwrap()].to_vec()
969 4 : );
970 4 : assert_eq!(
971 4 : listing.prefixes,
972 4 : [RemotePath::from_string("parent").unwrap()].to_vec()
973 4 : );
974 4 :
975 4 : // Delimiter and prefix without a trailing slash
976 4 : let listing = storage
977 4 : .list(
978 4 : Some(&RemotePath::from_string("timelines/some_timeline/grandparent").unwrap()),
979 4 : ListingMode::WithDelimiter,
980 4 : None,
981 4 : &cancel,
982 4 : )
983 4 : .await?;
984 4 : assert_eq!(listing.keys, [].to_vec());
985 4 : assert_eq!(
986 4 : listing.prefixes,
987 4 : [RemotePath::from_string("grandparent").unwrap()].to_vec()
988 4 : );
989 4 :
990 4 : // Delimiter and prefix that's partway through a path component
991 4 : let listing = storage
992 4 : .list(
993 4 : Some(&RemotePath::from_string("timelines/some_timeline/grandp").unwrap()),
994 4 : ListingMode::WithDelimiter,
995 4 : None,
996 4 : &cancel,
997 4 : )
998 4 : .await?;
999 4 : assert_eq!(listing.keys, [].to_vec());
1000 4 : assert_eq!(
1001 4 : listing.prefixes,
1002 4 : [RemotePath::from_string("grandparent").unwrap()].to_vec()
1003 4 : );
1004 4 :
1005 4 : Ok(())
1006 4 : }
1007 :
1008 : #[tokio::test]
1009 4 : async fn list_part_component() -> anyhow::Result<()> {
1010 4 : // No delimiter: should recursively list everything
1011 4 : let (storage, cancel) = create_storage()?;
1012 4 :
1013 4 : // Imitates what happens in a tenant path when we have an unsharded path and a sharded path, and do a listing
1014 4 : // of the unsharded path: although there is a "directory" at the unsharded path, it should be handled as
1015 4 : // a freeform prefix.
1016 4 : let _child_a =
1017 24 : upload_dummy_file(&storage, "grandparent/tenant-01/child", None, &cancel).await?;
1018 4 : let _child_b =
1019 24 : upload_dummy_file(&storage, "grandparent/tenant/child", None, &cancel).await?;
1020 4 :
1021 4 : // Delimiter and prefix that's partway through a path component
1022 4 : let listing = storage
1023 4 : .list(
1024 4 : Some(
1025 4 : &RemotePath::from_string("timelines/some_timeline/grandparent/tenant").unwrap(),
1026 4 : ),
1027 4 : ListingMode::WithDelimiter,
1028 4 : None,
1029 4 : &cancel,
1030 4 : )
1031 4 : .await?;
1032 4 : assert_eq!(listing.keys, [].to_vec());
1033 4 :
1034 4 : let mut found_prefixes = listing.prefixes.clone();
1035 4 : found_prefixes.sort();
1036 4 : assert_eq!(
1037 4 : found_prefixes,
1038 4 : [
1039 4 : RemotePath::from_string("tenant").unwrap(),
1040 4 : RemotePath::from_string("tenant-01").unwrap(),
1041 4 : ]
1042 4 : .to_vec()
1043 4 : );
1044 4 :
1045 4 : Ok(())
1046 4 : }
1047 :
1048 : #[tokio::test]
1049 4 : async fn overwrite_shorter_file() -> anyhow::Result<()> {
1050 4 : let (storage, cancel) = create_storage()?;
1051 4 :
1052 4 : let path = RemotePath::new("does/not/matter/file".into())?;
1053 4 :
1054 4 : let body = Bytes::from_static(b"long file contents is long");
1055 4 : {
1056 4 : let len = body.len();
1057 4 : let body =
1058 4 : futures::stream::once(futures::future::ready(std::io::Result::Ok(body.clone())));
1059 16 : storage.upload(body, len, &path, None, &cancel).await?;
1060 4 : }
1061 4 :
1062 8 : let read = aggregate(storage.download(&path, &cancel).await?.download_stream).await?;
1063 4 : assert_eq!(body, read);
1064 4 :
1065 4 : let shorter = Bytes::from_static(b"shorter body");
1066 4 : {
1067 4 : let len = shorter.len();
1068 4 : let body =
1069 4 : futures::stream::once(futures::future::ready(std::io::Result::Ok(shorter.clone())));
1070 12 : storage.upload(body, len, &path, None, &cancel).await?;
1071 4 : }
1072 4 :
1073 8 : let read = aggregate(storage.download(&path, &cancel).await?.download_stream).await?;
1074 4 : assert_eq!(shorter, read);
1075 4 : Ok(())
1076 4 : }
1077 :
1078 : #[tokio::test]
1079 4 : async fn cancelled_upload_can_later_be_retried() -> anyhow::Result<()> {
1080 4 : let (storage, cancel) = create_storage()?;
1081 4 :
1082 4 : let path = RemotePath::new("does/not/matter/file".into())?;
1083 4 :
1084 4 : let body = Bytes::from_static(b"long file contents is long");
1085 4 : {
1086 4 : let len = body.len();
1087 4 : let body =
1088 4 : futures::stream::once(futures::future::ready(std::io::Result::Ok(body.clone())));
1089 4 : let cancel = cancel.child_token();
1090 4 : cancel.cancel();
1091 4 : let e = storage
1092 4 : .upload(body, len, &path, None, &cancel)
1093 12 : .await
1094 4 : .unwrap_err();
1095 4 :
1096 4 : assert!(TimeoutOrCancel::caused_by_cancel(&e));
1097 4 : }
1098 4 :
1099 4 : {
1100 4 : let len = body.len();
1101 4 : let body =
1102 4 : futures::stream::once(futures::future::ready(std::io::Result::Ok(body.clone())));
1103 12 : storage.upload(body, len, &path, None, &cancel).await?;
1104 4 : }
1105 4 :
1106 8 : let read = aggregate(storage.download(&path, &cancel).await?.download_stream).await?;
1107 4 : assert_eq!(body, read);
1108 4 :
1109 4 : Ok(())
1110 4 : }
1111 :
1112 48 : async fn upload_dummy_file(
1113 48 : storage: &LocalFs,
1114 48 : name: &str,
1115 48 : metadata: Option<StorageMetadata>,
1116 48 : cancel: &CancellationToken,
1117 48 : ) -> anyhow::Result<RemotePath> {
1118 48 : let from_path = storage
1119 48 : .storage_root
1120 48 : .join("timelines")
1121 48 : .join("some_timeline")
1122 48 : .join(name);
1123 48 : let (file, size) = create_file_for_upload(&from_path, &dummy_contents(name)).await?;
1124 :
1125 48 : let relative_path = from_path
1126 48 : .strip_prefix(&storage.storage_root)
1127 48 : .context("Failed to strip storage root prefix")
1128 48 : .and_then(RemotePath::new)
1129 48 : .with_context(|| {
1130 0 : format!(
1131 0 : "Failed to resolve remote part of path {:?} for base {:?}",
1132 0 : from_path, storage.storage_root
1133 0 : )
1134 48 : })?;
1135 :
1136 48 : let file = tokio_util::io::ReaderStream::new(file);
1137 48 :
1138 48 : storage
1139 48 : .upload(file, size, &relative_path, metadata, cancel)
1140 232 : .await?;
1141 48 : Ok(relative_path)
1142 48 : }
1143 :
1144 48 : async fn create_file_for_upload(
1145 48 : path: &Utf8Path,
1146 48 : contents: &str,
1147 48 : ) -> anyhow::Result<(fs::File, usize)> {
1148 48 : std::fs::create_dir_all(path.parent().unwrap())?;
1149 48 : let mut file_for_writing = std::fs::OpenOptions::new()
1150 48 : .write(true)
1151 48 : .create_new(true)
1152 48 : .open(path)?;
1153 48 : write!(file_for_writing, "{}", contents)?;
1154 48 : drop(file_for_writing);
1155 48 : let file_size = path.metadata()?.len() as usize;
1156 48 : Ok((
1157 48 : fs::OpenOptions::new().read(true).open(&path).await?,
1158 48 : file_size,
1159 : ))
1160 48 : }
1161 :
1162 72 : fn dummy_contents(name: &str) -> String {
1163 72 : format!("contents for {name}")
1164 72 : }
1165 :
1166 4 : async fn list_files_sorted(storage: &LocalFs) -> anyhow::Result<Vec<RemotePath>> {
1167 12 : let mut files = storage.list_all().await?;
1168 4 : files.sort_by(|a, b| a.0.cmp(&b.0));
1169 4 : Ok(files)
1170 4 : }
1171 :
1172 44 : async fn aggregate(
1173 44 : stream: impl Stream<Item = std::io::Result<Bytes>>,
1174 44 : ) -> anyhow::Result<Vec<u8>> {
1175 44 : use futures::stream::StreamExt;
1176 44 : let mut out = Vec::new();
1177 44 : let mut stream = std::pin::pin!(stream);
1178 88 : while let Some(res) = stream.next().await {
1179 44 : out.extend_from_slice(&res?[..]);
1180 : }
1181 44 : Ok(out)
1182 44 : }
1183 : }
|