Line data Source code
1 : //! Azure Blob Storage wrapper
2 :
3 : use std::borrow::Cow;
4 : use std::collections::HashMap;
5 : use std::env;
6 : use std::fmt::Display;
7 : use std::io;
8 : use std::num::NonZeroU32;
9 : use std::pin::Pin;
10 : use std::str::FromStr;
11 : use std::time::Duration;
12 : use std::time::SystemTime;
13 :
14 : use super::REMOTE_STORAGE_PREFIX_SEPARATOR;
15 : use anyhow::Context;
16 : use anyhow::Result;
17 : use azure_core::request_options::{IfMatchCondition, MaxResults, Metadata, Range};
18 : use azure_core::{Continuable, RetryOptions};
19 : use azure_storage::StorageCredentials;
20 : use azure_storage_blobs::blob::CopyStatus;
21 : use azure_storage_blobs::prelude::ClientBuilder;
22 : use azure_storage_blobs::{blob::operations::GetBlobBuilder, prelude::ContainerClient};
23 : use bytes::Bytes;
24 : use futures::future::Either;
25 : use futures::stream::Stream;
26 : use futures::FutureExt;
27 : use futures_util::StreamExt;
28 : use futures_util::TryStreamExt;
29 : use http_types::{StatusCode, Url};
30 : use scopeguard::ScopeGuard;
31 : use tokio_util::sync::CancellationToken;
32 : use tracing::debug;
33 : use utils::backoff;
34 : use utils::backoff::exponential_backoff_duration_seconds;
35 :
36 : use crate::metrics::{start_measuring_requests, AttemptOutcome, RequestKind};
37 : use crate::DownloadKind;
38 : use crate::{
39 : config::AzureConfig, error::Cancelled, ConcurrencyLimiter, Download, DownloadError,
40 : DownloadOpts, Listing, ListingMode, ListingObject, RemotePath, RemoteStorage, StorageMetadata,
41 : TimeTravelError, TimeoutOrCancel,
42 : };
43 :
44 : pub struct AzureBlobStorage {
45 : client: ContainerClient,
46 : container_name: String,
47 : prefix_in_container: Option<String>,
48 : max_keys_per_list_response: Option<NonZeroU32>,
49 : concurrency_limiter: ConcurrencyLimiter,
50 : // Per-request timeout. Accessible for tests.
51 : pub timeout: Duration,
52 :
53 : // Alternative timeout used for metadata objects which are expected to be small
54 : pub small_timeout: Duration,
55 : }
56 :
57 : impl AzureBlobStorage {
58 10 : pub fn new(
59 10 : azure_config: &AzureConfig,
60 10 : timeout: Duration,
61 10 : small_timeout: Duration,
62 10 : ) -> Result<Self> {
63 10 : debug!(
64 0 : "Creating azure remote storage for azure container {}",
65 : azure_config.container_name
66 : );
67 :
68 : // Use the storage account from the config by default, fall back to env var if not present.
69 10 : let account = azure_config.storage_account.clone().unwrap_or_else(|| {
70 10 : env::var("AZURE_STORAGE_ACCOUNT").expect("missing AZURE_STORAGE_ACCOUNT")
71 10 : });
72 :
73 : // If the `AZURE_STORAGE_ACCESS_KEY` env var has an access key, use that,
74 : // otherwise try the token based credentials.
75 10 : let credentials = if let Ok(access_key) = env::var("AZURE_STORAGE_ACCESS_KEY") {
76 10 : StorageCredentials::access_key(account.clone(), access_key)
77 : } else {
78 0 : let token_credential = azure_identity::create_default_credential()
79 0 : .context("trying to obtain Azure default credentials")?;
80 0 : StorageCredentials::token_credential(token_credential)
81 : };
82 :
83 : // we have an outer retry
84 10 : let builder = ClientBuilder::new(account, credentials).retry(RetryOptions::none());
85 10 :
86 10 : let client = builder.container_client(azure_config.container_name.to_owned());
87 :
88 10 : let max_keys_per_list_response =
89 10 : if let Some(limit) = azure_config.max_keys_per_list_response {
90 : Some(
91 4 : NonZeroU32::new(limit as u32)
92 4 : .ok_or_else(|| anyhow::anyhow!("max_keys_per_list_response can't be 0"))?,
93 : )
94 : } else {
95 6 : None
96 : };
97 :
98 10 : Ok(AzureBlobStorage {
99 10 : client,
100 10 : container_name: azure_config.container_name.to_owned(),
101 10 : prefix_in_container: azure_config.prefix_in_container.to_owned(),
102 10 : max_keys_per_list_response,
103 10 : concurrency_limiter: ConcurrencyLimiter::new(azure_config.concurrency_limit.get()),
104 10 : timeout,
105 10 : small_timeout,
106 10 : })
107 10 : }
108 :
109 237 : pub fn relative_path_to_name(&self, path: &RemotePath) -> String {
110 237 : assert_eq!(std::path::MAIN_SEPARATOR, REMOTE_STORAGE_PREFIX_SEPARATOR);
111 237 : let path_string = path.get_path().as_str();
112 237 : match &self.prefix_in_container {
113 237 : Some(prefix) => {
114 237 : if prefix.ends_with(REMOTE_STORAGE_PREFIX_SEPARATOR) {
115 237 : prefix.clone() + path_string
116 : } else {
117 0 : format!("{prefix}{REMOTE_STORAGE_PREFIX_SEPARATOR}{path_string}")
118 : }
119 : }
120 0 : None => path_string.to_string(),
121 : }
122 237 : }
123 :
124 249 : fn name_to_relative_path(&self, key: &str) -> RemotePath {
125 249 : let relative_path =
126 249 : match key.strip_prefix(self.prefix_in_container.as_deref().unwrap_or_default()) {
127 249 : Some(stripped) => stripped,
128 : // we rely on Azure to return properly prefixed paths
129 : // for requests with a certain prefix
130 0 : None => panic!(
131 0 : "Key {key} does not start with container prefix {:?}",
132 0 : self.prefix_in_container
133 0 : ),
134 : };
135 249 : RemotePath(
136 249 : relative_path
137 249 : .split(REMOTE_STORAGE_PREFIX_SEPARATOR)
138 249 : .collect(),
139 249 : )
140 249 : }
141 :
142 11 : async fn download_for_builder(
143 11 : &self,
144 11 : builder: GetBlobBuilder,
145 11 : timeout: Duration,
146 11 : cancel: &CancellationToken,
147 11 : ) -> Result<Download, DownloadError> {
148 11 : let kind = RequestKind::Get;
149 :
150 11 : let _permit = self.permit(kind, cancel).await?;
151 11 : let cancel_or_timeout = crate::support::cancel_or_timeout(self.timeout, cancel.clone());
152 11 : let cancel_or_timeout_ = crate::support::cancel_or_timeout(self.timeout, cancel.clone());
153 11 :
154 11 : let mut etag = None;
155 11 : let mut last_modified = None;
156 11 : let mut metadata = HashMap::new();
157 11 :
158 11 : let started_at = start_measuring_requests(kind);
159 11 :
160 11 : let download = async {
161 11 : let response = builder
162 11 : // convert to concrete Pageable
163 11 : .into_stream()
164 11 : // convert to TryStream
165 11 : .into_stream()
166 11 : .map_err(to_download_error);
167 11 :
168 11 : // apply per request timeout
169 11 : let response = tokio_stream::StreamExt::timeout(response, timeout);
170 11 :
171 11 : // flatten
172 11 : let response = response.map(|res| match res {
173 11 : Ok(res) => res,
174 0 : Err(_elapsed) => Err(DownloadError::Timeout),
175 11 : });
176 11 :
177 11 : let mut response = Box::pin(response);
178 :
179 11 : let Some(part) = response.next().await else {
180 0 : return Err(DownloadError::Other(anyhow::anyhow!(
181 0 : "Azure GET response contained no response body"
182 0 : )));
183 : };
184 11 : let part = part?;
185 9 : if etag.is_none() {
186 9 : etag = Some(part.blob.properties.etag);
187 9 : }
188 9 : if last_modified.is_none() {
189 9 : last_modified = Some(part.blob.properties.last_modified.into());
190 9 : }
191 9 : if let Some(blob_meta) = part.blob.metadata {
192 0 : metadata.extend(blob_meta.iter().map(|(k, v)| (k.to_owned(), v.to_owned())));
193 9 : }
194 :
195 : // unwrap safety: if these were None, bufs would be empty and we would have returned an error already
196 9 : let etag = etag.unwrap();
197 9 : let last_modified = last_modified.unwrap();
198 9 :
199 9 : let tail_stream = response
200 9 : .map(|part| match part {
201 0 : Ok(part) => Either::Left(part.data.map(|r| r.map_err(io::Error::other))),
202 0 : Err(e) => {
203 0 : Either::Right(futures::stream::once(async { Err(io::Error::other(e)) }))
204 : }
205 9 : })
206 9 : .flatten();
207 9 : let stream = part
208 9 : .data
209 9 : .map(|r| r.map_err(io::Error::other))
210 9 : .chain(sync_wrapper::SyncStream::new(tail_stream));
211 9 : //.chain(SyncStream::from_pin(Box::pin(tail_stream)));
212 9 :
213 9 : let download_stream = crate::support::DownloadStream::new(cancel_or_timeout_, stream);
214 9 :
215 9 : Ok(Download {
216 9 : download_stream: Box::pin(download_stream),
217 9 : etag,
218 9 : last_modified,
219 9 : metadata: Some(StorageMetadata(metadata)),
220 9 : })
221 11 : };
222 :
223 11 : let download = tokio::select! {
224 11 : bufs = download => bufs,
225 11 : cancel_or_timeout = cancel_or_timeout => match cancel_or_timeout {
226 0 : TimeoutOrCancel::Timeout => return Err(DownloadError::Timeout),
227 0 : TimeoutOrCancel::Cancel => return Err(DownloadError::Cancelled),
228 : },
229 : };
230 11 : let started_at = ScopeGuard::into_inner(started_at);
231 11 : let outcome = match &download {
232 9 : Ok(_) => AttemptOutcome::Ok,
233 : // At this level in the stack 404 and 304 responses do not indicate an error.
234 : // There's expected cases when a blob may not exist or hasn't been modified since
235 : // the last get (e.g. probing for timeline indices and heatmap downloads).
236 : // Callers should handle errors if they are unexpected.
237 2 : Err(DownloadError::NotFound | DownloadError::Unmodified) => AttemptOutcome::Ok,
238 0 : Err(_) => AttemptOutcome::Err,
239 : };
240 11 : crate::metrics::BUCKET_METRICS
241 11 : .req_seconds
242 11 : .observe_elapsed(kind, outcome, started_at);
243 11 : download
244 11 : }
245 :
246 227 : async fn permit(
247 227 : &self,
248 227 : kind: RequestKind,
249 227 : cancel: &CancellationToken,
250 227 : ) -> Result<tokio::sync::SemaphorePermit<'_>, Cancelled> {
251 227 : let acquire = self.concurrency_limiter.acquire(kind);
252 227 :
253 227 : tokio::select! {
254 227 : permit = acquire => Ok(permit.expect("never closed")),
255 227 : _ = cancel.cancelled() => Err(Cancelled),
256 : }
257 227 : }
258 :
259 0 : pub fn container_name(&self) -> &str {
260 0 : &self.container_name
261 0 : }
262 : }
263 :
264 0 : fn to_azure_metadata(metadata: StorageMetadata) -> Metadata {
265 0 : let mut res = Metadata::new();
266 0 : for (k, v) in metadata.0.into_iter() {
267 0 : res.insert(k, v);
268 0 : }
269 0 : res
270 0 : }
271 :
272 3 : fn to_download_error(error: azure_core::Error) -> DownloadError {
273 3 : if let Some(http_err) = error.as_http_error() {
274 3 : match http_err.status() {
275 1 : StatusCode::NotFound => DownloadError::NotFound,
276 2 : StatusCode::NotModified => DownloadError::Unmodified,
277 0 : StatusCode::BadRequest => DownloadError::BadInput(anyhow::Error::new(error)),
278 0 : _ => DownloadError::Other(anyhow::Error::new(error)),
279 : }
280 : } else {
281 0 : DownloadError::Other(error.into())
282 : }
283 3 : }
284 :
285 : impl RemoteStorage for AzureBlobStorage {
286 26 : fn list_streaming(
287 26 : &self,
288 26 : prefix: Option<&RemotePath>,
289 26 : mode: ListingMode,
290 26 : max_keys: Option<NonZeroU32>,
291 26 : cancel: &CancellationToken,
292 26 : ) -> impl Stream<Item = Result<Listing, DownloadError>> {
293 26 : // get the passed prefix or if it is not set use prefix_in_bucket value
294 26 : let list_prefix = prefix.map(|p| self.relative_path_to_name(p)).or_else(|| {
295 10 : self.prefix_in_container.clone().map(|mut s| {
296 10 : if !s.ends_with(REMOTE_STORAGE_PREFIX_SEPARATOR) {
297 0 : s.push(REMOTE_STORAGE_PREFIX_SEPARATOR);
298 10 : }
299 10 : s
300 10 : })
301 26 : });
302 26 :
303 26 : async_stream::stream! {
304 26 : let _permit = self.permit(RequestKind::List, cancel).await?;
305 26 :
306 26 : let mut builder = self.client.list_blobs();
307 26 :
308 26 : if let ListingMode::WithDelimiter = mode {
309 26 : builder = builder.delimiter(REMOTE_STORAGE_PREFIX_SEPARATOR.to_string());
310 26 : }
311 26 :
312 26 : if let Some(prefix) = list_prefix {
313 26 : builder = builder.prefix(Cow::from(prefix.to_owned()));
314 26 : }
315 26 :
316 26 : if let Some(limit) = self.max_keys_per_list_response {
317 26 : builder = builder.max_results(MaxResults::new(limit));
318 26 : }
319 26 :
320 26 : let mut next_marker = None;
321 26 :
322 26 : let mut timeout_try_cnt = 1;
323 26 :
324 26 : 'outer: loop {
325 26 : let mut builder = builder.clone();
326 26 : if let Some(marker) = next_marker.clone() {
327 26 : builder = builder.marker(marker);
328 26 : }
329 26 : // Azure Blob Rust SDK does not expose the list blob API directly. Users have to use
330 26 : // their pageable iterator wrapper that returns all keys as a stream. We want to have
331 26 : // full control of paging, and therefore we only take the first item from the stream.
332 26 : let mut response_stream = builder.into_stream();
333 26 : let response = response_stream.next();
334 26 : // Timeout mechanism: Azure client will sometimes stuck on a request, but retrying that request
335 26 : // would immediately succeed. Therefore, we use exponential backoff timeout to retry the request.
336 26 : // (Usually, exponential backoff is used to determine the sleep time between two retries.) We
337 26 : // start with 10.0 second timeout, and double the timeout for each failure, up to 5 failures.
338 26 : // timeout = min(5 * (1.0+1.0)^n, self.timeout).
339 26 : let this_timeout = (5.0 * exponential_backoff_duration_seconds(timeout_try_cnt, 1.0, self.timeout.as_secs_f64())).min(self.timeout.as_secs_f64());
340 26 : let response = tokio::time::timeout(Duration::from_secs_f64(this_timeout), response);
341 45 : let response = response.map(|res| {
342 45 : match res {
343 45 : Ok(Some(Ok(res))) => Ok(Some(res)),
344 26 : Ok(Some(Err(e))) => Err(to_download_error(e)),
345 26 : Ok(None) => Ok(None),
346 26 : Err(_elasped) => Err(DownloadError::Timeout),
347 26 : }
348 45 : });
349 26 : let mut max_keys = max_keys.map(|mk| mk.get());
350 26 : let next_item = tokio::select! {
351 26 : op = response => op,
352 26 : _ = cancel.cancelled() => Err(DownloadError::Cancelled),
353 26 : };
354 26 :
355 26 : if let Err(DownloadError::Timeout) = &next_item {
356 26 : timeout_try_cnt += 1;
357 26 : if timeout_try_cnt <= 5 {
358 26 : continue;
359 26 : }
360 26 : }
361 26 :
362 26 : let next_item = next_item?;
363 26 :
364 26 : if timeout_try_cnt >= 2 {
365 26 : tracing::warn!("Azure Blob Storage list timed out and succeeded after {} tries", timeout_try_cnt);
366 26 : }
367 26 : timeout_try_cnt = 1;
368 26 :
369 26 : let Some(entry) = next_item else {
370 26 : // The list is complete, so yield it.
371 26 : break;
372 26 : };
373 26 :
374 26 : let mut res = Listing::default();
375 26 : next_marker = entry.continuation();
376 26 : let prefix_iter = entry
377 26 : .blobs
378 26 : .prefixes()
379 52 : .map(|prefix| self.name_to_relative_path(&prefix.name));
380 26 : res.prefixes.extend(prefix_iter);
381 26 :
382 26 : let blob_iter = entry
383 26 : .blobs
384 26 : .blobs()
385 197 : .map(|k| ListingObject{
386 197 : key: self.name_to_relative_path(&k.name),
387 197 : last_modified: k.properties.last_modified.into(),
388 197 : size: k.properties.content_length,
389 197 : }
390 26 : );
391 26 :
392 26 : for key in blob_iter {
393 26 : res.keys.push(key);
394 26 :
395 26 : if let Some(mut mk) = max_keys {
396 26 : assert!(mk > 0);
397 26 : mk -= 1;
398 26 : if mk == 0 {
399 26 : yield Ok(res); // limit reached
400 26 : break 'outer;
401 26 : }
402 26 : max_keys = Some(mk);
403 26 : }
404 26 : }
405 26 : yield Ok(res);
406 26 :
407 26 : // We are done here
408 26 : if next_marker.is_none() {
409 26 : break;
410 26 : }
411 26 : }
412 26 : }
413 26 : }
414 :
415 3 : async fn head_object(
416 3 : &self,
417 3 : key: &RemotePath,
418 3 : cancel: &CancellationToken,
419 3 : ) -> Result<ListingObject, DownloadError> {
420 3 : let kind = RequestKind::Head;
421 3 : let _permit = self.permit(kind, cancel).await?;
422 :
423 3 : let started_at = start_measuring_requests(kind);
424 3 :
425 3 : let blob_client = self.client.blob_client(self.relative_path_to_name(key));
426 3 : let properties_future = blob_client.get_properties().into_future();
427 3 :
428 3 : let properties_future = tokio::time::timeout(self.small_timeout, properties_future);
429 :
430 3 : let res = tokio::select! {
431 3 : res = properties_future => res,
432 3 : _ = cancel.cancelled() => return Err(TimeoutOrCancel::Cancel.into()),
433 : };
434 :
435 3 : if let Ok(inner) = &res {
436 3 : // do not incl. timeouts as errors in metrics but cancellations
437 3 : let started_at = ScopeGuard::into_inner(started_at);
438 3 : crate::metrics::BUCKET_METRICS
439 3 : .req_seconds
440 3 : .observe_elapsed(kind, inner, started_at);
441 3 : }
442 :
443 3 : let data = match res {
444 2 : Ok(Ok(data)) => Ok(data),
445 1 : Ok(Err(sdk)) => Err(to_download_error(sdk)),
446 0 : Err(_timeout) => Err(DownloadError::Timeout),
447 1 : }?;
448 :
449 2 : let properties = data.blob.properties;
450 2 : Ok(ListingObject {
451 2 : key: key.to_owned(),
452 2 : last_modified: SystemTime::from(properties.last_modified),
453 2 : size: properties.content_length,
454 2 : })
455 3 : }
456 :
457 93 : async fn upload(
458 93 : &self,
459 93 : from: impl Stream<Item = std::io::Result<Bytes>> + Send + Sync + 'static,
460 93 : data_size_bytes: usize,
461 93 : to: &RemotePath,
462 93 : metadata: Option<StorageMetadata>,
463 93 : cancel: &CancellationToken,
464 93 : ) -> anyhow::Result<()> {
465 93 : let kind = RequestKind::Put;
466 93 : let _permit = self.permit(kind, cancel).await?;
467 :
468 93 : let started_at = start_measuring_requests(kind);
469 93 :
470 93 : let op = async {
471 93 : let blob_client = self.client.blob_client(self.relative_path_to_name(to));
472 93 :
473 93 : let from: Pin<Box<dyn Stream<Item = std::io::Result<Bytes>> + Send + Sync + 'static>> =
474 93 : Box::pin(from);
475 93 :
476 93 : let from = NonSeekableStream::new(from, data_size_bytes);
477 93 :
478 93 : let body = azure_core::Body::SeekableStream(Box::new(from));
479 93 :
480 93 : let mut builder = blob_client.put_block_blob(body);
481 :
482 93 : if let Some(metadata) = metadata {
483 0 : builder = builder.metadata(to_azure_metadata(metadata));
484 93 : }
485 :
486 93 : let fut = builder.into_future();
487 93 : let fut = tokio::time::timeout(self.timeout, fut);
488 93 :
489 93 : match fut.await {
490 93 : Ok(Ok(_response)) => Ok(()),
491 0 : Ok(Err(azure)) => Err(azure.into()),
492 0 : Err(_timeout) => Err(TimeoutOrCancel::Timeout.into()),
493 : }
494 93 : };
495 :
496 93 : let res = tokio::select! {
497 93 : res = op => res,
498 93 : _ = cancel.cancelled() => return Err(TimeoutOrCancel::Cancel.into()),
499 : };
500 :
501 93 : let outcome = match res {
502 93 : Ok(_) => AttemptOutcome::Ok,
503 0 : Err(_) => AttemptOutcome::Err,
504 : };
505 93 : let started_at = ScopeGuard::into_inner(started_at);
506 93 : crate::metrics::BUCKET_METRICS
507 93 : .req_seconds
508 93 : .observe_elapsed(kind, outcome, started_at);
509 93 :
510 93 : res
511 93 : }
512 :
513 11 : async fn download(
514 11 : &self,
515 11 : from: &RemotePath,
516 11 : opts: &DownloadOpts,
517 11 : cancel: &CancellationToken,
518 11 : ) -> Result<Download, DownloadError> {
519 11 : let blob_client = self.client.blob_client(self.relative_path_to_name(from));
520 11 :
521 11 : let mut builder = blob_client.get();
522 :
523 11 : if let Some(ref etag) = opts.etag {
524 3 : builder = builder.if_match(IfMatchCondition::NotMatch(etag.to_string()))
525 8 : }
526 :
527 11 : if let Some((start, end)) = opts.byte_range() {
528 5 : builder = builder.range(match end {
529 3 : Some(end) => Range::Range(start..end),
530 2 : None => Range::RangeFrom(start..),
531 : });
532 6 : }
533 :
534 11 : let timeout = match opts.kind {
535 0 : DownloadKind::Small => self.small_timeout,
536 11 : DownloadKind::Large => self.timeout,
537 : };
538 :
539 11 : self.download_for_builder(builder, timeout, cancel).await
540 11 : }
541 :
542 86 : async fn delete(&self, path: &RemotePath, cancel: &CancellationToken) -> anyhow::Result<()> {
543 86 : self.delete_objects(std::array::from_ref(path), cancel)
544 86 : .await
545 86 : }
546 :
547 93 : async fn delete_objects<'a>(
548 93 : &self,
549 93 : paths: &'a [RemotePath],
550 93 : cancel: &CancellationToken,
551 93 : ) -> anyhow::Result<()> {
552 93 : let kind = RequestKind::Delete;
553 93 : let _permit = self.permit(kind, cancel).await?;
554 93 : let started_at = start_measuring_requests(kind);
555 93 :
556 93 : let op = async {
557 : // TODO batch requests are not supported by the SDK
558 : // https://github.com/Azure/azure-sdk-for-rust/issues/1068
559 205 : for path in paths {
560 112 : #[derive(Debug)]
561 112 : enum AzureOrTimeout {
562 : AzureError(azure_core::Error),
563 : Timeout,
564 : Cancel,
565 112 : }
566 112 : impl Display for AzureOrTimeout {
567 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
568 0 : write!(f, "{self:?}")
569 0 : }
570 : }
571 112 : let warn_threshold = 3;
572 112 : let max_retries = 5;
573 112 : backoff::retry(
574 112 : || async {
575 112 : let blob_client = self.client.blob_client(self.relative_path_to_name(path));
576 112 :
577 112 : let request = blob_client.delete().into_future();
578 :
579 112 : let res = tokio::time::timeout(self.timeout, request).await;
580 :
581 112 : match res {
582 90 : Ok(Ok(_v)) => Ok(()),
583 22 : Ok(Err(azure_err)) => {
584 22 : if let Some(http_err) = azure_err.as_http_error() {
585 22 : if http_err.status() == StatusCode::NotFound {
586 22 : return Ok(());
587 0 : }
588 0 : }
589 0 : Err(AzureOrTimeout::AzureError(azure_err))
590 : }
591 0 : Err(_elapsed) => Err(AzureOrTimeout::Timeout),
592 : }
593 224 : },
594 112 : |err| match err {
595 0 : AzureOrTimeout::AzureError(_) | AzureOrTimeout::Timeout => false,
596 0 : AzureOrTimeout::Cancel => true,
597 112 : },
598 112 : warn_threshold,
599 112 : max_retries,
600 112 : "deleting remote object",
601 112 : cancel,
602 112 : )
603 112 : .await
604 112 : .ok_or_else(|| AzureOrTimeout::Cancel)
605 112 : .and_then(|x| x)
606 112 : .map_err(|e| match e {
607 0 : AzureOrTimeout::AzureError(err) => anyhow::Error::from(err),
608 0 : AzureOrTimeout::Timeout => TimeoutOrCancel::Timeout.into(),
609 0 : AzureOrTimeout::Cancel => TimeoutOrCancel::Cancel.into(),
610 112 : })?;
611 : }
612 93 : Ok(())
613 93 : };
614 :
615 93 : let res = tokio::select! {
616 93 : res = op => res,
617 93 : _ = cancel.cancelled() => return Err(TimeoutOrCancel::Cancel.into()),
618 : };
619 :
620 93 : let started_at = ScopeGuard::into_inner(started_at);
621 93 : crate::metrics::BUCKET_METRICS
622 93 : .req_seconds
623 93 : .observe_elapsed(kind, &res, started_at);
624 93 : res
625 93 : }
626 :
627 0 : fn max_keys_per_delete(&self) -> usize {
628 0 : super::MAX_KEYS_PER_DELETE_AZURE
629 0 : }
630 :
631 1 : async fn copy(
632 1 : &self,
633 1 : from: &RemotePath,
634 1 : to: &RemotePath,
635 1 : cancel: &CancellationToken,
636 1 : ) -> anyhow::Result<()> {
637 1 : let kind = RequestKind::Copy;
638 1 : let _permit = self.permit(kind, cancel).await?;
639 1 : let started_at = start_measuring_requests(kind);
640 1 :
641 1 : let timeout = tokio::time::sleep(self.timeout);
642 1 :
643 1 : let mut copy_status = None;
644 1 :
645 1 : let op = async {
646 1 : let blob_client = self.client.blob_client(self.relative_path_to_name(to));
647 :
648 1 : let source_url = format!(
649 1 : "{}/{}",
650 1 : self.client.url()?,
651 1 : self.relative_path_to_name(from)
652 : );
653 :
654 1 : let builder = blob_client.copy(Url::from_str(&source_url)?);
655 1 : let copy = builder.into_future();
656 :
657 1 : let result = copy.await?;
658 :
659 1 : copy_status = Some(result.copy_status);
660 : loop {
661 1 : match copy_status.as_ref().expect("we always set it to Some") {
662 : CopyStatus::Aborted => {
663 0 : anyhow::bail!("Received abort for copy from {from} to {to}.");
664 : }
665 : CopyStatus::Failed => {
666 0 : anyhow::bail!("Received failure response for copy from {from} to {to}.");
667 : }
668 1 : CopyStatus::Success => return Ok(()),
669 0 : CopyStatus::Pending => (),
670 0 : }
671 0 : // The copy is taking longer. Waiting a second and then re-trying.
672 0 : // TODO estimate time based on copy_progress and adjust time based on that
673 0 : tokio::time::sleep(Duration::from_millis(1000)).await;
674 0 : let properties = blob_client.get_properties().into_future().await?;
675 0 : let Some(status) = properties.blob.properties.copy_status else {
676 0 : tracing::warn!("copy_status for copy is None!, from={from}, to={to}");
677 0 : return Ok(());
678 : };
679 0 : copy_status = Some(status);
680 : }
681 1 : };
682 :
683 1 : let res = tokio::select! {
684 1 : res = op => res,
685 1 : _ = cancel.cancelled() => return Err(anyhow::Error::new(TimeoutOrCancel::Cancel)),
686 1 : _ = timeout => {
687 0 : let e = anyhow::Error::new(TimeoutOrCancel::Timeout);
688 0 : let e = e.context(format!("Timeout, last status: {copy_status:?}"));
689 0 : Err(e)
690 : },
691 : };
692 :
693 1 : let started_at = ScopeGuard::into_inner(started_at);
694 1 : crate::metrics::BUCKET_METRICS
695 1 : .req_seconds
696 1 : .observe_elapsed(kind, &res, started_at);
697 1 : res
698 1 : }
699 :
700 0 : async fn time_travel_recover(
701 0 : &self,
702 0 : _prefix: Option<&RemotePath>,
703 0 : _timestamp: SystemTime,
704 0 : _done_if_after: SystemTime,
705 0 : _cancel: &CancellationToken,
706 0 : ) -> Result<(), TimeTravelError> {
707 0 : // TODO use Azure point in time recovery feature for this
708 0 : // https://learn.microsoft.com/en-us/azure/storage/blobs/point-in-time-restore-overview
709 0 : Err(TimeTravelError::Unimplemented)
710 0 : }
711 : }
712 :
713 : pin_project_lite::pin_project! {
714 : /// Hack to work around not being able to stream once with azure sdk.
715 : ///
716 : /// Azure sdk clones streams around with the assumption that they are like
717 : /// `Arc<tokio::fs::File>` (except not supporting tokio), however our streams are not like
718 : /// that. For example for an `index_part.json` we just have a single chunk of [`Bytes`]
719 : /// representing the whole serialized vec. It could be trivially cloneable and "semi-trivially"
720 : /// seekable, but we can also just re-try the request easier.
721 : #[project = NonSeekableStreamProj]
722 : enum NonSeekableStream<S> {
723 : /// A stream wrappers initial form.
724 : ///
725 : /// Mutex exists to allow moving when cloning. If the sdk changes to do less than 1
726 : /// clone before first request, then this must be changed.
727 : Initial {
728 : inner: std::sync::Mutex<Option<tokio_util::compat::Compat<tokio_util::io::StreamReader<S, Bytes>>>>,
729 : len: usize,
730 : },
731 : /// The actually readable variant, produced by cloning the Initial variant.
732 : ///
733 : /// The sdk currently always clones once, even without retry policy.
734 : Actual {
735 : #[pin]
736 : inner: tokio_util::compat::Compat<tokio_util::io::StreamReader<S, Bytes>>,
737 : len: usize,
738 : read_any: bool,
739 : },
740 : /// Most likely unneeded, but left to make life easier, in case more clones are added.
741 : Cloned {
742 : len_was: usize,
743 : }
744 : }
745 : }
746 :
747 : impl<S> NonSeekableStream<S>
748 : where
749 : S: Stream<Item = std::io::Result<Bytes>> + Send + Sync + 'static,
750 : {
751 93 : fn new(inner: S, len: usize) -> NonSeekableStream<S> {
752 : use tokio_util::compat::TokioAsyncReadCompatExt;
753 :
754 93 : let inner = tokio_util::io::StreamReader::new(inner).compat();
755 93 : let inner = Some(inner);
756 93 : let inner = std::sync::Mutex::new(inner);
757 93 : NonSeekableStream::Initial { inner, len }
758 93 : }
759 : }
760 :
761 : impl<S> std::fmt::Debug for NonSeekableStream<S> {
762 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
763 0 : match self {
764 0 : Self::Initial { len, .. } => f.debug_struct("Initial").field("len", len).finish(),
765 0 : Self::Actual { len, .. } => f.debug_struct("Actual").field("len", len).finish(),
766 0 : Self::Cloned { len_was, .. } => f.debug_struct("Cloned").field("len", len_was).finish(),
767 : }
768 0 : }
769 : }
770 :
771 : impl<S> futures::io::AsyncRead for NonSeekableStream<S>
772 : where
773 : S: Stream<Item = std::io::Result<Bytes>>,
774 : {
775 93 : fn poll_read(
776 93 : self: std::pin::Pin<&mut Self>,
777 93 : cx: &mut std::task::Context<'_>,
778 93 : buf: &mut [u8],
779 93 : ) -> std::task::Poll<std::io::Result<usize>> {
780 93 : match self.project() {
781 : NonSeekableStreamProj::Actual {
782 93 : inner, read_any, ..
783 93 : } => {
784 93 : *read_any = true;
785 93 : inner.poll_read(cx, buf)
786 : }
787 : // NonSeekableStream::Initial does not support reading because it is just much easier
788 : // to have the mutex in place where one does not poll the contents, or that's how it
789 : // seemed originally. If there is a version upgrade which changes the cloning, then
790 : // that support needs to be hacked in.
791 : //
792 : // including {self:?} into the message would be useful, but unsure how to unproject.
793 0 : _ => std::task::Poll::Ready(Err(std::io::Error::new(
794 0 : std::io::ErrorKind::Other,
795 0 : "cloned or initial values cannot be read",
796 0 : ))),
797 : }
798 93 : }
799 : }
800 :
801 : impl<S> Clone for NonSeekableStream<S> {
802 : /// Weird clone implementation exists to support the sdk doing cloning before issuing the first
803 : /// request, see type documentation.
804 93 : fn clone(&self) -> Self {
805 : use NonSeekableStream::*;
806 :
807 93 : match self {
808 93 : Initial { inner, len } => {
809 93 : if let Some(inner) = inner.lock().unwrap().take() {
810 93 : Actual {
811 93 : inner,
812 93 : len: *len,
813 93 : read_any: false,
814 93 : }
815 : } else {
816 0 : Self::Cloned { len_was: *len }
817 : }
818 : }
819 0 : Actual { len, .. } => Cloned { len_was: *len },
820 0 : Cloned { len_was } => Cloned { len_was: *len_was },
821 : }
822 93 : }
823 : }
824 :
825 : #[async_trait::async_trait]
826 : impl<S> azure_core::SeekableStream for NonSeekableStream<S>
827 : where
828 : S: Stream<Item = std::io::Result<Bytes>> + Unpin + Send + Sync + 'static,
829 : {
830 0 : async fn reset(&mut self) -> azure_core::error::Result<()> {
831 : use NonSeekableStream::*;
832 :
833 0 : let msg = match self {
834 0 : Initial { inner, .. } => {
835 0 : if inner.get_mut().unwrap().is_some() {
836 0 : return Ok(());
837 : } else {
838 0 : "reset after first clone is not supported"
839 : }
840 : }
841 0 : Actual { read_any, .. } if !*read_any => return Ok(()),
842 0 : Actual { .. } => "reset after reading is not supported",
843 0 : Cloned { .. } => "reset after second clone is not supported",
844 : };
845 0 : Err(azure_core::error::Error::new(
846 0 : azure_core::error::ErrorKind::Io,
847 0 : std::io::Error::new(std::io::ErrorKind::Other, msg),
848 0 : ))
849 0 : }
850 :
851 : // Note: it is not documented if this should be the total or remaining length, total passes the
852 : // tests.
853 93 : fn len(&self) -> usize {
854 : use NonSeekableStream::*;
855 93 : match self {
856 93 : Initial { len, .. } => *len,
857 0 : Actual { len, .. } => *len,
858 0 : Cloned { len_was, .. } => *len_was,
859 : }
860 93 : }
861 : }
|
