Line data Source code
1 : //! Azure Blob Storage wrapper
2 :
3 : use std::borrow::Cow;
4 : use std::collections::HashMap;
5 : use std::fmt::Display;
6 : use std::num::NonZeroU32;
7 : use std::pin::Pin;
8 : use std::str::FromStr;
9 : use std::sync::Arc;
10 : use std::time::{Duration, SystemTime};
11 : use std::{env, io};
12 :
13 : use anyhow::{Context, Result};
14 : use azure_core::request_options::{IfMatchCondition, MaxResults, Metadata, Range};
15 : use azure_core::{Continuable, HttpClient, RetryOptions, TransportOptions};
16 : use azure_storage::StorageCredentials;
17 : use azure_storage_blobs::blob::CopyStatus;
18 : use azure_storage_blobs::blob::operations::GetBlobBuilder;
19 : use azure_storage_blobs::prelude::{ClientBuilder, ContainerClient};
20 : use bytes::Bytes;
21 : use futures::FutureExt;
22 : use futures::future::Either;
23 : use futures::stream::Stream;
24 : use futures_util::{StreamExt, TryStreamExt};
25 : use http_types::{StatusCode, Url};
26 : use scopeguard::ScopeGuard;
27 : use tokio_util::sync::CancellationToken;
28 : use tracing::debug;
29 : use utils::backoff;
30 : use utils::backoff::exponential_backoff_duration_seconds;
31 :
32 : use super::REMOTE_STORAGE_PREFIX_SEPARATOR;
33 : use crate::config::AzureConfig;
34 : use crate::error::Cancelled;
35 : use crate::metrics::{AttemptOutcome, RequestKind, start_measuring_requests};
36 : use crate::{
37 : ConcurrencyLimiter, Download, DownloadError, DownloadKind, DownloadOpts, Listing, ListingMode,
38 : ListingObject, RemotePath, RemoteStorage, StorageMetadata, TimeTravelError, TimeoutOrCancel,
39 : };
40 :
41 : pub struct AzureBlobStorage {
42 : client: ContainerClient,
43 : container_name: String,
44 : prefix_in_container: Option<String>,
45 : max_keys_per_list_response: Option<NonZeroU32>,
46 : concurrency_limiter: ConcurrencyLimiter,
47 : // Per-request timeout. Accessible for tests.
48 : pub timeout: Duration,
49 :
50 : // Alternative timeout used for metadata objects which are expected to be small
51 : pub small_timeout: Duration,
52 : }
53 :
54 : impl AzureBlobStorage {
55 10 : pub fn new(
56 10 : azure_config: &AzureConfig,
57 10 : timeout: Duration,
58 10 : small_timeout: Duration,
59 10 : ) -> Result<Self> {
60 10 : debug!(
61 0 : "Creating azure remote storage for azure container {}",
62 : azure_config.container_name
63 : );
64 :
65 : // Use the storage account from the config by default, fall back to env var if not present.
66 10 : let account = azure_config.storage_account.clone().unwrap_or_else(|| {
67 10 : env::var("AZURE_STORAGE_ACCOUNT").expect("missing AZURE_STORAGE_ACCOUNT")
68 10 : });
69 :
70 : // If the `AZURE_STORAGE_ACCESS_KEY` env var has an access key, use that,
71 : // otherwise try the token based credentials.
72 10 : let credentials = if let Ok(access_key) = env::var("AZURE_STORAGE_ACCESS_KEY") {
73 10 : StorageCredentials::access_key(account.clone(), access_key)
74 : } else {
75 0 : let token_credential = azure_identity::create_default_credential()
76 0 : .context("trying to obtain Azure default credentials")?;
77 0 : StorageCredentials::token_credential(token_credential)
78 : };
79 :
80 10 : let builder = ClientBuilder::new(account, credentials)
81 10 : // we have an outer retry
82 10 : .retry(RetryOptions::none())
83 10 : // Customize transport to configure conneciton pooling
84 10 : .transport(TransportOptions::new(Self::reqwest_client(
85 10 : azure_config.conn_pool_size,
86 10 : )));
87 10 :
88 10 : let client = builder.container_client(azure_config.container_name.to_owned());
89 :
90 10 : let max_keys_per_list_response =
91 10 : if let Some(limit) = azure_config.max_keys_per_list_response {
92 : Some(
93 4 : NonZeroU32::new(limit as u32)
94 4 : .ok_or_else(|| anyhow::anyhow!("max_keys_per_list_response can't be 0"))?,
95 : )
96 : } else {
97 6 : None
98 : };
99 :
100 10 : Ok(AzureBlobStorage {
101 10 : client,
102 10 : container_name: azure_config.container_name.to_owned(),
103 10 : prefix_in_container: azure_config.prefix_in_container.to_owned(),
104 10 : max_keys_per_list_response,
105 10 : concurrency_limiter: ConcurrencyLimiter::new(azure_config.concurrency_limit.get()),
106 10 : timeout,
107 10 : small_timeout,
108 10 : })
109 10 : }
110 :
111 10 : fn reqwest_client(conn_pool_size: usize) -> Arc<dyn HttpClient> {
112 10 : let client = reqwest::ClientBuilder::new()
113 10 : .pool_max_idle_per_host(conn_pool_size)
114 10 : .build()
115 10 : .expect("failed to build `reqwest` client");
116 10 : Arc::new(client)
117 10 : }
118 :
119 237 : pub fn relative_path_to_name(&self, path: &RemotePath) -> String {
120 237 : assert_eq!(std::path::MAIN_SEPARATOR, REMOTE_STORAGE_PREFIX_SEPARATOR);
121 237 : let path_string = path.get_path().as_str();
122 237 : match &self.prefix_in_container {
123 237 : Some(prefix) => {
124 237 : if prefix.ends_with(REMOTE_STORAGE_PREFIX_SEPARATOR) {
125 237 : prefix.clone() + path_string
126 : } else {
127 0 : format!("{prefix}{REMOTE_STORAGE_PREFIX_SEPARATOR}{path_string}")
128 : }
129 : }
130 0 : None => path_string.to_string(),
131 : }
132 237 : }
133 :
134 249 : fn name_to_relative_path(&self, key: &str) -> RemotePath {
135 249 : let relative_path =
136 249 : match key.strip_prefix(self.prefix_in_container.as_deref().unwrap_or_default()) {
137 249 : Some(stripped) => stripped,
138 : // we rely on Azure to return properly prefixed paths
139 : // for requests with a certain prefix
140 0 : None => panic!(
141 0 : "Key {key} does not start with container prefix {:?}",
142 0 : self.prefix_in_container
143 0 : ),
144 : };
145 249 : RemotePath(
146 249 : relative_path
147 249 : .split(REMOTE_STORAGE_PREFIX_SEPARATOR)
148 249 : .collect(),
149 249 : )
150 249 : }
151 :
152 11 : async fn download_for_builder(
153 11 : &self,
154 11 : builder: GetBlobBuilder,
155 11 : timeout: Duration,
156 11 : cancel: &CancellationToken,
157 11 : ) -> Result<Download, DownloadError> {
158 11 : let kind = RequestKind::Get;
159 :
160 11 : let _permit = self.permit(kind, cancel).await?;
161 11 : let cancel_or_timeout = crate::support::cancel_or_timeout(self.timeout, cancel.clone());
162 11 : let cancel_or_timeout_ = crate::support::cancel_or_timeout(self.timeout, cancel.clone());
163 11 :
164 11 : let mut etag = None;
165 11 : let mut last_modified = None;
166 11 : let mut metadata = HashMap::new();
167 11 :
168 11 : let started_at = start_measuring_requests(kind);
169 11 :
170 11 : let download = async {
171 11 : let response = builder
172 11 : // convert to concrete Pageable
173 11 : .into_stream()
174 11 : // convert to TryStream
175 11 : .into_stream()
176 11 : .map_err(to_download_error);
177 11 :
178 11 : // apply per request timeout
179 11 : let response = tokio_stream::StreamExt::timeout(response, timeout);
180 11 :
181 11 : // flatten
182 11 : let response = response.map(|res| match res {
183 11 : Ok(res) => res,
184 0 : Err(_elapsed) => Err(DownloadError::Timeout),
185 11 : });
186 11 :
187 11 : let mut response = Box::pin(response);
188 :
189 11 : let Some(part) = response.next().await else {
190 0 : return Err(DownloadError::Other(anyhow::anyhow!(
191 0 : "Azure GET response contained no response body"
192 0 : )));
193 : };
194 11 : let part = part?;
195 9 : if etag.is_none() {
196 9 : etag = Some(part.blob.properties.etag);
197 9 : }
198 9 : if last_modified.is_none() {
199 9 : last_modified = Some(part.blob.properties.last_modified.into());
200 9 : }
201 9 : if let Some(blob_meta) = part.blob.metadata {
202 0 : metadata.extend(blob_meta.iter().map(|(k, v)| (k.to_owned(), v.to_owned())));
203 0 : }
204 :
205 : // unwrap safety: if these were None, bufs would be empty and we would have returned an error already
206 9 : let etag = etag.unwrap();
207 9 : let last_modified = last_modified.unwrap();
208 9 :
209 9 : let tail_stream = response
210 9 : .map(|part| match part {
211 0 : Ok(part) => Either::Left(part.data.map(|r| r.map_err(io::Error::other))),
212 0 : Err(e) => {
213 0 : Either::Right(futures::stream::once(async { Err(io::Error::other(e)) }))
214 : }
215 9 : })
216 9 : .flatten();
217 9 : let stream = part
218 9 : .data
219 9 : .map(|r| r.map_err(io::Error::other))
220 9 : .chain(sync_wrapper::SyncStream::new(tail_stream));
221 9 : //.chain(SyncStream::from_pin(Box::pin(tail_stream)));
222 9 :
223 9 : let download_stream = crate::support::DownloadStream::new(cancel_or_timeout_, stream);
224 9 :
225 9 : Ok(Download {
226 9 : download_stream: Box::pin(download_stream),
227 9 : etag,
228 9 : last_modified,
229 9 : metadata: Some(StorageMetadata(metadata)),
230 9 : })
231 0 : };
232 :
233 11 : let download = tokio::select! {
234 11 : bufs = download => bufs,
235 11 : cancel_or_timeout = cancel_or_timeout => match cancel_or_timeout {
236 0 : TimeoutOrCancel::Timeout => return Err(DownloadError::Timeout),
237 0 : TimeoutOrCancel::Cancel => return Err(DownloadError::Cancelled),
238 : },
239 : };
240 11 : let started_at = ScopeGuard::into_inner(started_at);
241 11 : let outcome = match &download {
242 9 : Ok(_) => AttemptOutcome::Ok,
243 : // At this level in the stack 404 and 304 responses do not indicate an error.
244 : // There's expected cases when a blob may not exist or hasn't been modified since
245 : // the last get (e.g. probing for timeline indices and heatmap downloads).
246 : // Callers should handle errors if they are unexpected.
247 2 : Err(DownloadError::NotFound | DownloadError::Unmodified) => AttemptOutcome::Ok,
248 0 : Err(_) => AttemptOutcome::Err,
249 : };
250 11 : crate::metrics::BUCKET_METRICS
251 11 : .req_seconds
252 11 : .observe_elapsed(kind, outcome, started_at);
253 11 : download
254 0 : }
255 :
256 227 : async fn permit(
257 227 : &self,
258 227 : kind: RequestKind,
259 227 : cancel: &CancellationToken,
260 227 : ) -> Result<tokio::sync::SemaphorePermit<'_>, Cancelled> {
261 227 : let acquire = self.concurrency_limiter.acquire(kind);
262 227 :
263 227 : tokio::select! {
264 227 : permit = acquire => Ok(permit.expect("never closed")),
265 227 : _ = cancel.cancelled() => Err(Cancelled),
266 : }
267 0 : }
268 :
269 0 : pub fn container_name(&self) -> &str {
270 0 : &self.container_name
271 0 : }
272 : }
273 :
274 0 : fn to_azure_metadata(metadata: StorageMetadata) -> Metadata {
275 0 : let mut res = Metadata::new();
276 0 : for (k, v) in metadata.0.into_iter() {
277 0 : res.insert(k, v);
278 0 : }
279 0 : res
280 0 : }
281 :
282 3 : fn to_download_error(error: azure_core::Error) -> DownloadError {
283 3 : if let Some(http_err) = error.as_http_error() {
284 3 : match http_err.status() {
285 1 : StatusCode::NotFound => DownloadError::NotFound,
286 2 : StatusCode::NotModified => DownloadError::Unmodified,
287 0 : StatusCode::BadRequest => DownloadError::BadInput(anyhow::Error::new(error)),
288 0 : _ => DownloadError::Other(anyhow::Error::new(error)),
289 : }
290 : } else {
291 0 : DownloadError::Other(error.into())
292 : }
293 3 : }
294 :
295 : impl RemoteStorage for AzureBlobStorage {
296 26 : fn list_streaming(
297 26 : &self,
298 26 : prefix: Option<&RemotePath>,
299 26 : mode: ListingMode,
300 26 : max_keys: Option<NonZeroU32>,
301 26 : cancel: &CancellationToken,
302 26 : ) -> impl Stream<Item = Result<Listing, DownloadError>> {
303 26 : // get the passed prefix or if it is not set use prefix_in_bucket value
304 26 : let list_prefix = prefix.map(|p| self.relative_path_to_name(p)).or_else(|| {
305 10 : self.prefix_in_container.clone().map(|mut s| {
306 10 : if !s.ends_with(REMOTE_STORAGE_PREFIX_SEPARATOR) {
307 0 : s.push(REMOTE_STORAGE_PREFIX_SEPARATOR);
308 10 : }
309 10 : s
310 10 : })
311 26 : });
312 26 :
313 26 : async_stream::stream! {
314 26 : let _permit = self.permit(RequestKind::List, cancel).await?;
315 26 :
316 26 : let mut builder = self.client.list_blobs();
317 26 :
318 26 : if let ListingMode::WithDelimiter = mode {
319 26 : builder = builder.delimiter(REMOTE_STORAGE_PREFIX_SEPARATOR.to_string());
320 26 : }
321 26 :
322 26 : if let Some(prefix) = list_prefix {
323 26 : builder = builder.prefix(Cow::from(prefix.to_owned()));
324 26 : }
325 26 :
326 26 : if let Some(limit) = self.max_keys_per_list_response {
327 26 : builder = builder.max_results(MaxResults::new(limit));
328 26 : }
329 26 :
330 26 : let mut next_marker = None;
331 26 :
332 26 : let mut timeout_try_cnt = 1;
333 26 :
334 26 : 'outer: loop {
335 26 : let mut builder = builder.clone();
336 26 : if let Some(marker) = next_marker.clone() {
337 26 : builder = builder.marker(marker);
338 26 : }
339 26 : // Azure Blob Rust SDK does not expose the list blob API directly. Users have to use
340 26 : // their pageable iterator wrapper that returns all keys as a stream. We want to have
341 26 : // full control of paging, and therefore we only take the first item from the stream.
342 26 : let mut response_stream = builder.into_stream();
343 26 : let response = response_stream.next();
344 26 : // Timeout mechanism: Azure client will sometimes stuck on a request, but retrying that request
345 26 : // would immediately succeed. Therefore, we use exponential backoff timeout to retry the request.
346 26 : // (Usually, exponential backoff is used to determine the sleep time between two retries.) We
347 26 : // start with 10.0 second timeout, and double the timeout for each failure, up to 5 failures.
348 26 : // timeout = min(5 * (1.0+1.0)^n, self.timeout).
349 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());
350 26 : let response = tokio::time::timeout(Duration::from_secs_f64(this_timeout), response);
351 26 : let response = response.map(|res| {
352 45 : match res {
353 45 : Ok(Some(Ok(res))) => Ok(Some(res)),
354 26 : Ok(Some(Err(e))) => Err(to_download_error(e)),
355 26 : Ok(None) => Ok(None),
356 26 : Err(_elasped) => Err(DownloadError::Timeout),
357 26 : }
358 26 : });
359 26 : let mut max_keys = max_keys.map(|mk| mk.get());
360 26 : let next_item = tokio::select! {
361 26 : op = response => op,
362 26 : _ = cancel.cancelled() => Err(DownloadError::Cancelled),
363 26 : };
364 26 :
365 26 : if let Err(DownloadError::Timeout) = &next_item {
366 26 : timeout_try_cnt += 1;
367 26 : if timeout_try_cnt <= 5 {
368 26 : continue;
369 26 : }
370 26 : }
371 26 :
372 26 : let next_item = next_item?;
373 26 :
374 26 : // Log a warning if we saw two timeouts in a row before a successful request
375 26 : if timeout_try_cnt > 2 {
376 26 : tracing::warn!("Azure Blob Storage list timed out and succeeded after {} tries", timeout_try_cnt);
377 26 : }
378 26 : timeout_try_cnt = 1;
379 26 :
380 26 : let Some(entry) = next_item else {
381 26 : // The list is complete, so yield it.
382 26 : break;
383 26 : };
384 26 :
385 26 : let mut res = Listing::default();
386 26 : next_marker = entry.continuation();
387 26 : let prefix_iter = entry
388 26 : .blobs
389 26 : .prefixes()
390 52 : .map(|prefix| self.name_to_relative_path(&prefix.name));
391 26 : res.prefixes.extend(prefix_iter);
392 26 :
393 26 : let blob_iter = entry
394 26 : .blobs
395 26 : .blobs()
396 197 : .map(|k| ListingObject{
397 197 : key: self.name_to_relative_path(&k.name),
398 197 : last_modified: k.properties.last_modified.into(),
399 197 : size: k.properties.content_length,
400 197 : }
401 26 : );
402 26 :
403 26 : for key in blob_iter {
404 26 : res.keys.push(key);
405 26 :
406 26 : if let Some(mut mk) = max_keys {
407 26 : assert!(mk > 0);
408 26 : mk -= 1;
409 26 : if mk == 0 {
410 26 : yield Ok(res); // limit reached
411 26 : break 'outer;
412 26 : }
413 26 : max_keys = Some(mk);
414 26 : }
415 26 : }
416 26 : yield Ok(res);
417 26 :
418 26 : // We are done here
419 26 : if next_marker.is_none() {
420 26 : break;
421 26 : }
422 26 : }
423 26 : }
424 26 : }
425 :
426 3 : async fn head_object(
427 3 : &self,
428 3 : key: &RemotePath,
429 3 : cancel: &CancellationToken,
430 3 : ) -> Result<ListingObject, DownloadError> {
431 3 : let kind = RequestKind::Head;
432 3 : let _permit = self.permit(kind, cancel).await?;
433 :
434 3 : let started_at = start_measuring_requests(kind);
435 3 :
436 3 : let blob_client = self.client.blob_client(self.relative_path_to_name(key));
437 3 : let properties_future = blob_client.get_properties().into_future();
438 3 :
439 3 : let properties_future = tokio::time::timeout(self.small_timeout, properties_future);
440 :
441 3 : let res = tokio::select! {
442 3 : res = properties_future => res,
443 3 : _ = cancel.cancelled() => return Err(TimeoutOrCancel::Cancel.into()),
444 : };
445 :
446 3 : if let Ok(inner) = &res {
447 3 : // do not incl. timeouts as errors in metrics but cancellations
448 3 : let started_at = ScopeGuard::into_inner(started_at);
449 3 : crate::metrics::BUCKET_METRICS
450 3 : .req_seconds
451 3 : .observe_elapsed(kind, inner, started_at);
452 3 : }
453 :
454 3 : let data = match res {
455 2 : Ok(Ok(data)) => Ok(data),
456 1 : Ok(Err(sdk)) => Err(to_download_error(sdk)),
457 0 : Err(_timeout) => Err(DownloadError::Timeout),
458 1 : }?;
459 :
460 2 : let properties = data.blob.properties;
461 2 : Ok(ListingObject {
462 2 : key: key.to_owned(),
463 2 : last_modified: SystemTime::from(properties.last_modified),
464 2 : size: properties.content_length,
465 2 : })
466 0 : }
467 :
468 93 : async fn upload(
469 93 : &self,
470 93 : from: impl Stream<Item = std::io::Result<Bytes>> + Send + Sync + 'static,
471 93 : data_size_bytes: usize,
472 93 : to: &RemotePath,
473 93 : metadata: Option<StorageMetadata>,
474 93 : cancel: &CancellationToken,
475 93 : ) -> anyhow::Result<()> {
476 93 : let kind = RequestKind::Put;
477 93 : let _permit = self.permit(kind, cancel).await?;
478 :
479 93 : let started_at = start_measuring_requests(kind);
480 93 :
481 93 : let op = async {
482 93 : let blob_client = self.client.blob_client(self.relative_path_to_name(to));
483 93 :
484 93 : let from: Pin<Box<dyn Stream<Item = std::io::Result<Bytes>> + Send + Sync + 'static>> =
485 93 : Box::pin(from);
486 93 :
487 93 : let from = NonSeekableStream::new(from, data_size_bytes);
488 93 :
489 93 : let body = azure_core::Body::SeekableStream(Box::new(from));
490 93 :
491 93 : let mut builder = blob_client.put_block_blob(body);
492 :
493 93 : if let Some(metadata) = metadata {
494 0 : builder = builder.metadata(to_azure_metadata(metadata));
495 0 : }
496 :
497 93 : let fut = builder.into_future();
498 93 : let fut = tokio::time::timeout(self.timeout, fut);
499 93 :
500 93 : match fut.await {
501 93 : Ok(Ok(_response)) => Ok(()),
502 0 : Ok(Err(azure)) => Err(azure.into()),
503 0 : Err(_timeout) => Err(TimeoutOrCancel::Timeout.into()),
504 : }
505 0 : };
506 :
507 93 : let res = tokio::select! {
508 93 : res = op => res,
509 93 : _ = cancel.cancelled() => return Err(TimeoutOrCancel::Cancel.into()),
510 : };
511 :
512 93 : let outcome = match res {
513 93 : Ok(_) => AttemptOutcome::Ok,
514 0 : Err(_) => AttemptOutcome::Err,
515 : };
516 93 : let started_at = ScopeGuard::into_inner(started_at);
517 93 : crate::metrics::BUCKET_METRICS
518 93 : .req_seconds
519 93 : .observe_elapsed(kind, outcome, started_at);
520 93 :
521 93 : res
522 0 : }
523 :
524 11 : async fn download(
525 11 : &self,
526 11 : from: &RemotePath,
527 11 : opts: &DownloadOpts,
528 11 : cancel: &CancellationToken,
529 11 : ) -> Result<Download, DownloadError> {
530 11 : let blob_client = self.client.blob_client(self.relative_path_to_name(from));
531 11 :
532 11 : let mut builder = blob_client.get();
533 :
534 11 : if let Some(ref etag) = opts.etag {
535 3 : builder = builder.if_match(IfMatchCondition::NotMatch(etag.to_string()))
536 0 : }
537 :
538 11 : if let Some((start, end)) = opts.byte_range() {
539 5 : builder = builder.range(match end {
540 3 : Some(end) => Range::Range(start..end),
541 2 : None => Range::RangeFrom(start..),
542 : });
543 0 : }
544 :
545 11 : let timeout = match opts.kind {
546 0 : DownloadKind::Small => self.small_timeout,
547 11 : DownloadKind::Large => self.timeout,
548 : };
549 :
550 11 : self.download_for_builder(builder, timeout, cancel).await
551 0 : }
552 :
553 86 : async fn delete(&self, path: &RemotePath, cancel: &CancellationToken) -> anyhow::Result<()> {
554 86 : self.delete_objects(std::array::from_ref(path), cancel)
555 86 : .await
556 0 : }
557 :
558 93 : async fn delete_objects(
559 93 : &self,
560 93 : paths: &[RemotePath],
561 93 : cancel: &CancellationToken,
562 93 : ) -> anyhow::Result<()> {
563 93 : let kind = RequestKind::Delete;
564 93 : let _permit = self.permit(kind, cancel).await?;
565 93 : let started_at = start_measuring_requests(kind);
566 93 :
567 93 : let op = async {
568 : // TODO batch requests are not supported by the SDK
569 : // https://github.com/Azure/azure-sdk-for-rust/issues/1068
570 205 : for path in paths {
571 112 : #[derive(Debug)]
572 112 : enum AzureOrTimeout {
573 : AzureError(azure_core::Error),
574 : Timeout,
575 : Cancel,
576 112 : }
577 112 : impl Display for AzureOrTimeout {
578 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
579 0 : write!(f, "{self:?}")
580 0 : }
581 : }
582 112 : let warn_threshold = 3;
583 112 : let max_retries = 5;
584 112 : backoff::retry(
585 112 : || async {
586 112 : let blob_client = self.client.blob_client(self.relative_path_to_name(path));
587 112 :
588 112 : let request = blob_client.delete().into_future();
589 :
590 112 : let res = tokio::time::timeout(self.timeout, request).await;
591 :
592 112 : match res {
593 90 : Ok(Ok(_v)) => Ok(()),
594 22 : Ok(Err(azure_err)) => {
595 22 : if let Some(http_err) = azure_err.as_http_error() {
596 22 : if http_err.status() == StatusCode::NotFound {
597 22 : return Ok(());
598 0 : }
599 0 : }
600 0 : Err(AzureOrTimeout::AzureError(azure_err))
601 : }
602 0 : Err(_elapsed) => Err(AzureOrTimeout::Timeout),
603 : }
604 112 : },
605 112 : |err| match err {
606 0 : AzureOrTimeout::AzureError(_) | AzureOrTimeout::Timeout => false,
607 0 : AzureOrTimeout::Cancel => true,
608 112 : },
609 112 : warn_threshold,
610 112 : max_retries,
611 112 : "deleting remote object",
612 112 : cancel,
613 112 : )
614 112 : .await
615 112 : .ok_or_else(|| AzureOrTimeout::Cancel)
616 112 : .and_then(|x| x)
617 112 : .map_err(|e| match e {
618 0 : AzureOrTimeout::AzureError(err) => anyhow::Error::from(err),
619 0 : AzureOrTimeout::Timeout => TimeoutOrCancel::Timeout.into(),
620 0 : AzureOrTimeout::Cancel => TimeoutOrCancel::Cancel.into(),
621 112 : })?;
622 : }
623 93 : Ok(())
624 0 : };
625 :
626 93 : let res = tokio::select! {
627 93 : res = op => res,
628 93 : _ = cancel.cancelled() => return Err(TimeoutOrCancel::Cancel.into()),
629 : };
630 :
631 93 : let started_at = ScopeGuard::into_inner(started_at);
632 93 : crate::metrics::BUCKET_METRICS
633 93 : .req_seconds
634 93 : .observe_elapsed(kind, &res, started_at);
635 93 : res
636 0 : }
637 :
638 0 : fn max_keys_per_delete(&self) -> usize {
639 0 : super::MAX_KEYS_PER_DELETE_AZURE
640 0 : }
641 :
642 1 : async fn copy(
643 1 : &self,
644 1 : from: &RemotePath,
645 1 : to: &RemotePath,
646 1 : cancel: &CancellationToken,
647 1 : ) -> anyhow::Result<()> {
648 1 : let kind = RequestKind::Copy;
649 1 : let _permit = self.permit(kind, cancel).await?;
650 1 : let started_at = start_measuring_requests(kind);
651 1 :
652 1 : let timeout = tokio::time::sleep(self.timeout);
653 1 :
654 1 : let mut copy_status = None;
655 1 :
656 1 : let op = async {
657 1 : let blob_client = self.client.blob_client(self.relative_path_to_name(to));
658 :
659 1 : let source_url = format!(
660 1 : "{}/{}",
661 1 : self.client.url()?,
662 1 : self.relative_path_to_name(from)
663 : );
664 :
665 1 : let builder = blob_client.copy(Url::from_str(&source_url)?);
666 1 : let copy = builder.into_future();
667 :
668 1 : let result = copy.await?;
669 :
670 1 : copy_status = Some(result.copy_status);
671 : loop {
672 1 : match copy_status.as_ref().expect("we always set it to Some") {
673 : CopyStatus::Aborted => {
674 0 : anyhow::bail!("Received abort for copy from {from} to {to}.");
675 : }
676 : CopyStatus::Failed => {
677 0 : anyhow::bail!("Received failure response for copy from {from} to {to}.");
678 : }
679 1 : CopyStatus::Success => return Ok(()),
680 0 : CopyStatus::Pending => (),
681 0 : }
682 0 : // The copy is taking longer. Waiting a second and then re-trying.
683 0 : // TODO estimate time based on copy_progress and adjust time based on that
684 0 : tokio::time::sleep(Duration::from_millis(1000)).await;
685 0 : let properties = blob_client.get_properties().into_future().await?;
686 0 : let Some(status) = properties.blob.properties.copy_status else {
687 0 : tracing::warn!("copy_status for copy is None!, from={from}, to={to}");
688 0 : return Ok(());
689 : };
690 0 : copy_status = Some(status);
691 : }
692 0 : };
693 :
694 1 : let res = tokio::select! {
695 1 : res = op => res,
696 1 : _ = cancel.cancelled() => return Err(anyhow::Error::new(TimeoutOrCancel::Cancel)),
697 1 : _ = timeout => {
698 0 : let e = anyhow::Error::new(TimeoutOrCancel::Timeout);
699 0 : let e = e.context(format!("Timeout, last status: {copy_status:?}"));
700 0 : Err(e)
701 : },
702 : };
703 :
704 1 : let started_at = ScopeGuard::into_inner(started_at);
705 1 : crate::metrics::BUCKET_METRICS
706 1 : .req_seconds
707 1 : .observe_elapsed(kind, &res, started_at);
708 1 : res
709 0 : }
710 :
711 0 : async fn time_travel_recover(
712 0 : &self,
713 0 : _prefix: Option<&RemotePath>,
714 0 : _timestamp: SystemTime,
715 0 : _done_if_after: SystemTime,
716 0 : _cancel: &CancellationToken,
717 0 : ) -> Result<(), TimeTravelError> {
718 0 : // TODO use Azure point in time recovery feature for this
719 0 : // https://learn.microsoft.com/en-us/azure/storage/blobs/point-in-time-restore-overview
720 0 : Err(TimeTravelError::Unimplemented)
721 0 : }
722 : }
723 :
724 : pin_project_lite::pin_project! {
725 : /// Hack to work around not being able to stream once with azure sdk.
726 : ///
727 : /// Azure sdk clones streams around with the assumption that they are like
728 : /// `Arc<tokio::fs::File>` (except not supporting tokio), however our streams are not like
729 : /// that. For example for an `index_part.json` we just have a single chunk of [`Bytes`]
730 : /// representing the whole serialized vec. It could be trivially cloneable and "semi-trivially"
731 : /// seekable, but we can also just re-try the request easier.
732 : #[project = NonSeekableStreamProj]
733 : enum NonSeekableStream<S> {
734 : /// A stream wrappers initial form.
735 : ///
736 : /// Mutex exists to allow moving when cloning. If the sdk changes to do less than 1
737 : /// clone before first request, then this must be changed.
738 : Initial {
739 : inner: std::sync::Mutex<Option<tokio_util::compat::Compat<tokio_util::io::StreamReader<S, Bytes>>>>,
740 : len: usize,
741 : },
742 : /// The actually readable variant, produced by cloning the Initial variant.
743 : ///
744 : /// The sdk currently always clones once, even without retry policy.
745 : Actual {
746 : #[pin]
747 : inner: tokio_util::compat::Compat<tokio_util::io::StreamReader<S, Bytes>>,
748 : len: usize,
749 : read_any: bool,
750 : },
751 : /// Most likely unneeded, but left to make life easier, in case more clones are added.
752 : Cloned {
753 : len_was: usize,
754 : }
755 : }
756 : }
757 :
758 : impl<S> NonSeekableStream<S>
759 : where
760 : S: Stream<Item = std::io::Result<Bytes>> + Send + Sync + 'static,
761 : {
762 93 : fn new(inner: S, len: usize) -> NonSeekableStream<S> {
763 : use tokio_util::compat::TokioAsyncReadCompatExt;
764 :
765 93 : let inner = tokio_util::io::StreamReader::new(inner).compat();
766 93 : let inner = Some(inner);
767 93 : let inner = std::sync::Mutex::new(inner);
768 93 : NonSeekableStream::Initial { inner, len }
769 93 : }
770 : }
771 :
772 : impl<S> std::fmt::Debug for NonSeekableStream<S> {
773 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
774 0 : match self {
775 0 : Self::Initial { len, .. } => f.debug_struct("Initial").field("len", len).finish(),
776 0 : Self::Actual { len, .. } => f.debug_struct("Actual").field("len", len).finish(),
777 0 : Self::Cloned { len_was, .. } => f.debug_struct("Cloned").field("len", len_was).finish(),
778 : }
779 0 : }
780 : }
781 :
782 : impl<S> futures::io::AsyncRead for NonSeekableStream<S>
783 : where
784 : S: Stream<Item = std::io::Result<Bytes>>,
785 : {
786 93 : fn poll_read(
787 93 : self: std::pin::Pin<&mut Self>,
788 93 : cx: &mut std::task::Context<'_>,
789 93 : buf: &mut [u8],
790 93 : ) -> std::task::Poll<std::io::Result<usize>> {
791 93 : match self.project() {
792 : NonSeekableStreamProj::Actual {
793 93 : inner, read_any, ..
794 93 : } => {
795 93 : *read_any = true;
796 93 : inner.poll_read(cx, buf)
797 : }
798 : // NonSeekableStream::Initial does not support reading because it is just much easier
799 : // to have the mutex in place where one does not poll the contents, or that's how it
800 : // seemed originally. If there is a version upgrade which changes the cloning, then
801 : // that support needs to be hacked in.
802 : //
803 : // including {self:?} into the message would be useful, but unsure how to unproject.
804 0 : _ => std::task::Poll::Ready(Err(std::io::Error::new(
805 0 : std::io::ErrorKind::Other,
806 0 : "cloned or initial values cannot be read",
807 0 : ))),
808 : }
809 0 : }
810 : }
811 :
812 : impl<S> Clone for NonSeekableStream<S> {
813 : /// Weird clone implementation exists to support the sdk doing cloning before issuing the first
814 : /// request, see type documentation.
815 93 : fn clone(&self) -> Self {
816 : use NonSeekableStream::*;
817 :
818 93 : match self {
819 93 : Initial { inner, len } => {
820 93 : if let Some(inner) = inner.lock().unwrap().take() {
821 93 : Actual {
822 93 : inner,
823 93 : len: *len,
824 93 : read_any: false,
825 93 : }
826 : } else {
827 0 : Self::Cloned { len_was: *len }
828 : }
829 : }
830 0 : Actual { len, .. } => Cloned { len_was: *len },
831 0 : Cloned { len_was } => Cloned { len_was: *len_was },
832 : }
833 0 : }
834 : }
835 :
836 : #[async_trait::async_trait]
837 : impl<S> azure_core::SeekableStream for NonSeekableStream<S>
838 : where
839 : S: Stream<Item = std::io::Result<Bytes>> + Unpin + Send + Sync + 'static,
840 : {
841 0 : async fn reset(&mut self) -> azure_core::error::Result<()> {
842 : use NonSeekableStream::*;
843 :
844 0 : let msg = match self {
845 0 : Initial { inner, .. } => {
846 0 : if inner.get_mut().unwrap().is_some() {
847 0 : return Ok(());
848 : } else {
849 0 : "reset after first clone is not supported"
850 : }
851 : }
852 0 : Actual { read_any, .. } if !*read_any => return Ok(()),
853 0 : Actual { .. } => "reset after reading is not supported",
854 0 : Cloned { .. } => "reset after second clone is not supported",
855 : };
856 0 : Err(azure_core::error::Error::new(
857 0 : azure_core::error::ErrorKind::Io,
858 0 : std::io::Error::new(std::io::ErrorKind::Other, msg),
859 0 : ))
860 0 : }
861 :
862 : // Note: it is not documented if this should be the total or remaining length, total passes the
863 : // tests.
864 93 : fn len(&self) -> usize {
865 : use NonSeekableStream::*;
866 93 : match self {
867 93 : Initial { len, .. } => *len,
868 0 : Actual { len, .. } => *len,
869 0 : Cloned { len_was, .. } => *len_was,
870 : }
871 0 : }
872 : }
|