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