/build/source/nativelink-store/src/s3_store.rs
Line | Count | Source |
1 | | // Copyright 2024 The NativeLink Authors. All rights reserved. |
2 | | // |
3 | | // Licensed under the Functional Source License, Version 1.1, Apache 2.0 Future License (the "License"); |
4 | | // you may not use this file except in compliance with the License. |
5 | | // You may obtain a copy of the License at |
6 | | // |
7 | | // See LICENSE file for details |
8 | | // |
9 | | // Unless required by applicable law or agreed to in writing, software |
10 | | // distributed under the License is distributed on an "AS IS" BASIS, |
11 | | // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
12 | | // See the License for the specific language governing permissions and |
13 | | // limitations under the License. |
14 | | |
15 | | use core::cmp; |
16 | | use core::pin::Pin; |
17 | | use core::time::Duration; |
18 | | use std::borrow::Cow; |
19 | | use std::sync::Arc; |
20 | | |
21 | | use async_trait::async_trait; |
22 | | use aws_config::default_provider::credentials; |
23 | | use aws_config::provider_config::ProviderConfig; |
24 | | use aws_config::{AppName, BehaviorVersion}; |
25 | | use aws_sdk_s3::Client; |
26 | | use aws_sdk_s3::config::{Credentials, Region}; |
27 | | use aws_sdk_s3::operation::create_multipart_upload::CreateMultipartUploadOutput; |
28 | | use aws_sdk_s3::operation::get_object::GetObjectError; |
29 | | use aws_sdk_s3::operation::head_object::HeadObjectError; |
30 | | use aws_sdk_s3::primitives::ByteStream; // SdkBody |
31 | | use aws_sdk_s3::types::builders::{CompletedMultipartUploadBuilder, CompletedPartBuilder}; |
32 | | use aws_smithy_runtime_api::client::http::HttpClient as SmithyHttpClient; |
33 | | use aws_smithy_types::body::SdkBody; |
34 | | use futures::future::FusedFuture; |
35 | | use futures::stream::{FuturesUnordered, unfold}; |
36 | | use futures::{FutureExt, StreamExt, TryFutureExt, TryStreamExt}; |
37 | | use nativelink_config::stores::ExperimentalAwsSpec; |
38 | | // Note: S3 store should be very careful about the error codes it returns |
39 | | // when in a retryable wrapper. Always prefer Code::Aborted or another |
40 | | // retryable code over Code::InvalidArgument or make_input_err!(). |
41 | | // ie: Don't import make_input_err!() to help prevent this. |
42 | | use nativelink_error::{Code, Error, ResultExt, make_err}; |
43 | | use nativelink_metric::MetricsComponent; |
44 | | use nativelink_util::buf_channel::{ |
45 | | DropCloserReadHalf, DropCloserWriteHalf, make_buf_channel_pair, |
46 | | }; |
47 | | use nativelink_util::health_utils::{HealthRegistryBuilder, HealthStatus, HealthStatusIndicator}; |
48 | | use nativelink_util::instant_wrapper::InstantWrapper; |
49 | | use nativelink_util::retry::{Retrier, RetryResult}; |
50 | | use nativelink_util::store_trait::{ |
51 | | RemoveCallback, StoreDriver, StoreKey, StoreOptimizations, UploadSizeInfo, |
52 | | }; |
53 | | use parking_lot::Mutex; |
54 | | use tokio::sync::mpsc; |
55 | | use tokio::time::sleep; |
56 | | use tracing::{error, info}; |
57 | | |
58 | | use crate::cas_utils::is_zero_digest; |
59 | | use crate::common_s3_utils::{BodyWrapper, TlsClient}; |
60 | | |
61 | | // S3 object cannot be larger than this number. See: |
62 | | // https://docs.aws.amazon.com/AmazonS3/latest/userguide/qfacts.html |
63 | | const MAX_UPLOAD_SIZE: u64 = 48 * 1024 * 1024 * 1024 * 1024; // 48TiB (technically should be 48.8 TiB, but close enough) |
64 | | |
65 | | // S3 parts cannot be smaller than this number. See: |
66 | | // https://docs.aws.amazon.com/AmazonS3/latest/userguide/qfacts.html |
67 | | const MIN_MULTIPART_SIZE: u64 = 5 * 1024 * 1024; // 5MB. |
68 | | |
69 | | // S3 parts cannot be larger than this number. See: |
70 | | // https://docs.aws.amazon.com/AmazonS3/latest/userguide/qfacts.html |
71 | | const MAX_MULTIPART_SIZE: u64 = 5 * 1024 * 1024 * 1024; // 5GB. |
72 | | |
73 | | // S3 parts cannot be more than this number. See: |
74 | | // https://docs.aws.amazon.com/AmazonS3/latest/userguide/qfacts.html |
75 | | // Note: Type 'u64' chosen to simplify calculations |
76 | | const MAX_UPLOAD_PARTS: u64 = 10_000; |
77 | | |
78 | | // Default max buffer size for retrying upload requests. |
79 | | // Note: If you change this, adjust the docs in the config. |
80 | | const DEFAULT_MAX_RETRY_BUFFER_PER_REQUEST: usize = 5 * 1024 * 1024; // 5MB. |
81 | | |
82 | | // Default limit for concurrent part uploads per multipart upload. |
83 | | // Note: If you change this, adjust the docs in the config. |
84 | | const DEFAULT_MULTIPART_MAX_CONCURRENT_UPLOADS: usize = 10; |
85 | | |
86 | | #[derive(Debug, MetricsComponent)] |
87 | | pub struct S3Store<NowFn> { |
88 | | s3_client: Arc<Client>, |
89 | | now_fn: NowFn, |
90 | | #[metric(help = "The bucket name for the S3 store")] |
91 | | bucket: String, |
92 | | #[metric(help = "The key prefix for the S3 store")] |
93 | | key_prefix: String, |
94 | | retrier: Retrier, |
95 | | #[metric(help = "The number of seconds to consider an object expired")] |
96 | | consider_expired_after_s: i64, |
97 | | #[metric(help = "The number of bytes to buffer for retrying requests")] |
98 | | max_retry_buffer_per_request: usize, |
99 | | #[metric(help = "The number of concurrent uploads allowed for multipart uploads")] |
100 | | multipart_max_concurrent_uploads: usize, |
101 | | |
102 | | remove_callbacks: Mutex<Vec<RemoveCallback>>, |
103 | | } |
104 | | |
105 | | impl<I, NowFn> S3Store<NowFn> |
106 | | where |
107 | | I: InstantWrapper, |
108 | | NowFn: Fn() -> I + Send + Sync + Unpin + 'static, |
109 | | { |
110 | 0 | pub async fn new(spec: &ExperimentalAwsSpec, now_fn: NowFn) -> Result<Arc<Self>, Error> { |
111 | 0 | Self::new_with_http_clients( |
112 | 0 | spec, |
113 | 0 | TlsClient::new(&spec.common.clone())?, |
114 | 0 | TlsClient::new_for_credentials(&spec.common)?, |
115 | 0 | now_fn, |
116 | | ) |
117 | 0 | .await |
118 | 0 | } |
119 | | |
120 | | /// Builds the store with caller-supplied HTTP clients. Production uses |
121 | | /// [`Self::new`] (which injects a [`TlsClient`]); tests inject a mock (e.g. |
122 | | /// `StaticReplayClient`) to exercise wire behavior such as endpoint |
123 | | /// overrides and path-style addressing without a live bucket. |
124 | 3 | pub async fn new_with_http_clients<C>( |
125 | 3 | spec: &ExperimentalAwsSpec, |
126 | 3 | http_client: C, |
127 | 3 | credential_http_client: C, |
128 | 3 | now_fn: NowFn, |
129 | 3 | ) -> Result<Arc<Self>, Error> |
130 | 3 | where |
131 | 3 | C: SmithyHttpClient + Clone + 'static, |
132 | 3 | { |
133 | 3 | let jitter_fn = spec.common.retry.make_jitter_fn(); |
134 | 3 | let s3_client = { |
135 | 3 | let region = Region::new(Cow::Owned(spec.region.clone())); |
136 | | |
137 | 3 | let loader = aws_config::defaults(BehaviorVersion::latest()) |
138 | 3 | .app_name(AppName::new("nativelink").expect("valid app name")) |
139 | 3 | .timeout_config( |
140 | 3 | aws_config::timeout::TimeoutConfig::builder() |
141 | 3 | .connect_timeout(Duration::from_secs(15)) |
142 | 3 | .build(), |
143 | | ) |
144 | 3 | .region(region.clone()) |
145 | 3 | .http_client(http_client); |
146 | | |
147 | 3 | let loader = if let Some(key_id) = &spec.access_key_id |
148 | 3 | && let Some(secret) = &spec.secret_access_key |
149 | | { |
150 | 3 | loader.credentials_provider(Credentials::new( |
151 | 3 | key_id, |
152 | 3 | secret, |
153 | 3 | None, |
154 | 3 | None, |
155 | | "s3-explicit", |
156 | | )) |
157 | | } else { |
158 | 0 | loader.credentials_provider( |
159 | 0 | credentials::DefaultCredentialsChain::builder() |
160 | 0 | .configure( |
161 | 0 | ProviderConfig::without_region() |
162 | 0 | .with_region(Some(region)) |
163 | 0 | .with_http_client(credential_http_client), |
164 | 0 | ) |
165 | 0 | .build() |
166 | 0 | .await, |
167 | | ) |
168 | | }; |
169 | | |
170 | 3 | let config = loader.load().await; |
171 | | |
172 | 3 | let mut config_builder = |
173 | 3 | aws_sdk_s3::config::Builder::from(&config).force_path_style(spec.force_path_style); |
174 | | |
175 | 3 | if let Some(endpoint) = &spec.endpoint { |
176 | 3 | config_builder = config_builder.endpoint_url(endpoint); |
177 | 3 | }0 |
178 | | |
179 | 3 | Client::from_conf(config_builder.build()) |
180 | | }; |
181 | 3 | Self::new_with_client_and_jitter(spec, s3_client, jitter_fn, now_fn) |
182 | 3 | } |
183 | | |
184 | 22 | pub fn new_with_client_and_jitter( |
185 | 22 | spec: &ExperimentalAwsSpec, |
186 | 22 | s3_client: Client, |
187 | 22 | jitter_fn: Arc<dyn Fn(Duration) -> Duration + Send + Sync>, |
188 | 22 | now_fn: NowFn, |
189 | 22 | ) -> Result<Arc<Self>, Error> { |
190 | 22 | Ok(Arc::new(Self { |
191 | 22 | s3_client: Arc::new(s3_client), |
192 | 22 | now_fn, |
193 | 22 | bucket: spec.bucket.clone(), |
194 | 22 | key_prefix: spec |
195 | 22 | .common |
196 | 22 | .key_prefix |
197 | 22 | .as_ref() |
198 | 22 | .unwrap_or(&String::new()) |
199 | 22 | .clone(), |
200 | 22 | retrier: Retrier::new( |
201 | 22 | Arc::new(|duration| Box::pin2 (sleep2 (duration2 ))), |
202 | 22 | jitter_fn, |
203 | 22 | spec.common.retry.clone(), |
204 | | ), |
205 | 22 | consider_expired_after_s: i64::from(spec.common.consider_expired_after_s), |
206 | 22 | max_retry_buffer_per_request: spec |
207 | 22 | .common |
208 | 22 | .max_retry_buffer_per_request |
209 | 22 | .unwrap_or(DEFAULT_MAX_RETRY_BUFFER_PER_REQUEST), |
210 | 22 | multipart_max_concurrent_uploads: spec |
211 | 22 | .common |
212 | 22 | .multipart_max_concurrent_uploads |
213 | 22 | .unwrap_or(DEFAULT_MULTIPART_MAX_CONCURRENT_UPLOADS), |
214 | 22 | remove_callbacks: Mutex::new(Vec::new()), |
215 | | })) |
216 | 22 | } |
217 | | |
218 | 22 | fn make_s3_path(&self, key: &StoreKey<'_>) -> String { |
219 | 22 | format!("{}{}", self.key_prefix, key.as_str()) |
220 | 22 | } |
221 | | |
222 | 7 | async fn has(self: Pin<&Self>, digest: StoreKey<'_>) -> Result<Option<u64>, Error> { |
223 | 7 | let digest_clone = digest.into_owned(); |
224 | 7 | self.retrier |
225 | 8 | .retry7 (unfold7 (()7 , move |state| { |
226 | 8 | let local_digest = digest_clone.clone(); |
227 | 8 | async move { |
228 | 8 | let result = self |
229 | 8 | .s3_client |
230 | 8 | .head_object() |
231 | 8 | .bucket(&self.bucket) |
232 | 8 | .key(self.make_s3_path(&local_digest)) |
233 | 8 | .send() |
234 | 8 | .await; |
235 | | |
236 | 8 | match result { |
237 | 5 | Ok(head_object_output) => { |
238 | 5 | if self.consider_expired_after_s != 0 |
239 | 2 | && let Some(last_modified) = head_object_output.last_modified |
240 | | { |
241 | 2 | let now_s = (self.now_fn)() |
242 | 2 | .unix_timestamp() |
243 | 2 | .try_into() |
244 | 2 | .unwrap_or(i64::MAX); |
245 | 2 | if last_modified.secs() + self.consider_expired_after_s <= now_s { |
246 | 1 | let remove_callbacks = self.remove_callbacks.lock().clone(); |
247 | 1 | let mut callbacks: FuturesUnordered<_> = remove_callbacks |
248 | 1 | .iter() |
249 | 1 | .map(|callback| callback0 .callback0 (local_digest0 .borrow0 ())) |
250 | 1 | .collect(); |
251 | 1 | while callbacks.next().await.is_some() {}0 |
252 | 1 | return Some((RetryResult::Ok(None), state)); |
253 | 1 | } |
254 | 3 | } |
255 | 4 | let Some(length) = head_object_output.content_length else { |
256 | 0 | return Some((RetryResult::Ok(None), state)); |
257 | | }; |
258 | 4 | if length >= 0 { |
259 | 4 | return Some((RetryResult::Ok(Some(length as u64)), state)); |
260 | 0 | } |
261 | 0 | Some(( |
262 | 0 | RetryResult::Err(make_err!( |
263 | 0 | Code::InvalidArgument, |
264 | 0 | "Negative content length in S3: {length:?}", |
265 | 0 | )), |
266 | 0 | state, |
267 | 0 | )) |
268 | | } |
269 | 3 | Err(sdk_error) => match sdk_error.into_service_error() { |
270 | 2 | HeadObjectError::NotFound(_) => Some((RetryResult::Ok(None), state)), |
271 | 1 | other => Some(( |
272 | 1 | RetryResult::Retry( |
273 | 1 | Error::from_std_err(Code::Unavailable, &other) |
274 | 1 | .append("Unhandled HeadObjectError in S3"), |
275 | 1 | ), |
276 | 1 | state, |
277 | 1 | )), |
278 | | }, |
279 | | } |
280 | 8 | } |
281 | 8 | })) |
282 | 7 | .await |
283 | 7 | } |
284 | | } |
285 | | |
286 | | #[async_trait] |
287 | | impl<I, NowFn> StoreDriver for S3Store<NowFn> |
288 | | where |
289 | | I: InstantWrapper, |
290 | | NowFn: Fn() -> I + Send + Sync + Unpin + 'static, |
291 | | { |
292 | 0 | async fn post_init(self: Arc<Self>) -> Result<(), Error> { |
293 | | Ok(()) |
294 | 0 | } |
295 | | |
296 | | async fn has_with_results( |
297 | | self: Pin<&Self>, |
298 | | keys: &[StoreKey<'_>], |
299 | | results: &mut [Option<u64>], |
300 | 8 | ) -> Result<(), Error> { |
301 | | keys.iter() |
302 | | .zip(results.iter_mut()) |
303 | 8 | .map(|(key, result)| async move { |
304 | | // We need to do a special pass to ensure our zero key exist. |
305 | 8 | if is_zero_digest(key.borrow()) { |
306 | 1 | *result = Some(0); |
307 | 1 | return Ok::<_, Error>(()); |
308 | 7 | } |
309 | 7 | *result = self.has(key.borrow()).await?0 ; |
310 | 7 | Ok::<_, Error>(()) |
311 | 16 | }) |
312 | | .collect::<FuturesUnordered<_>>() |
313 | | .try_collect() |
314 | | .await |
315 | 8 | } |
316 | | |
317 | 0 | fn optimized_for(&self, optimization: StoreOptimizations) -> bool { |
318 | 0 | matches!(optimization, StoreOptimizations::LazyExistenceOnSync) |
319 | 0 | } |
320 | | |
321 | | async fn update( |
322 | | self: Pin<&Self>, |
323 | | digest: StoreKey<'_>, |
324 | | mut reader: DropCloserReadHalf, |
325 | | upload_size: UploadSizeInfo, |
326 | 7 | ) -> Result<u64, Error> { |
327 | | let s3_path = &self.make_s3_path(&digest); |
328 | | |
329 | | let max_size = match upload_size { |
330 | | UploadSizeInfo::ExactSize(sz) | UploadSizeInfo::MaxSize(sz) => sz, |
331 | | }; |
332 | | |
333 | | // Sanity check S3 maximum upload size. |
334 | | if max_size > MAX_UPLOAD_SIZE { |
335 | | return Err(make_err!( |
336 | | Code::FailedPrecondition, |
337 | | "File size exceeds max of {MAX_UPLOAD_SIZE}" |
338 | | )); |
339 | | } |
340 | | |
341 | | // Note(aaronmondal) It might be more optimal to use a different |
342 | | // heuristic here, but for simplicity we use a hard coded value. |
343 | | // Anything going down this if-statement will have the advantage of only |
344 | | // 1 network request for the upload instead of minimum of 3 required for |
345 | | // multipart upload requests. |
346 | | // |
347 | | // Note(aaronmondal) If the upload size is not known, we go down the multipart upload path. |
348 | | // This is not very efficient, but it greatly reduces the complexity of the code. |
349 | | if max_size < MIN_MULTIPART_SIZE && matches!(upload_size, UploadSizeInfo::ExactSize(_)) { |
350 | | let UploadSizeInfo::ExactSize(sz) = upload_size else { |
351 | | unreachable!("upload_size must be UploadSizeInfo::ExactSize here"); |
352 | | }; |
353 | | reader.set_max_recent_data_size( |
354 | | u64::try_from(self.max_retry_buffer_per_request) |
355 | | .err_tip(|| "Could not convert max_retry_buffer_per_request to u64")?, |
356 | | ); |
357 | | return self |
358 | | .retrier |
359 | 5 | .retry(unfold(reader, move |mut reader| async move { |
360 | | // We need to make a new pair here because the aws sdk does not give us |
361 | | // back the body after we send it in order to retry. |
362 | 5 | let (mut tx, rx) = make_buf_channel_pair(); |
363 | | |
364 | | // Upload the data to the S3 backend. |
365 | 5 | let result = { |
366 | 5 | let reader_ref = &mut reader; |
367 | 5 | let (upload_res, bind_res) = tokio::join!( |
368 | 5 | self.s3_client |
369 | 5 | .put_object() |
370 | 5 | .bucket(&self.bucket) |
371 | 5 | .key(s3_path.clone()) |
372 | 5 | .content_length(sz.try_into().unwrap_or(i64::MAX)) |
373 | 5 | .body(ByteStream::from_body_1_x(BodyWrapper { |
374 | 5 | reader: rx, |
375 | 5 | size: sz, |
376 | 5 | })) |
377 | 5 | .send() |
378 | 5 | .map_ok_or_else(|e| Err(Error::from_std_err0 (Code::Aborted0 , &e0 )), |_| Ok(sz)), |
379 | | // Stream all data from the reader channel to the writer channel. |
380 | 5 | tx.bind_buffered(reader_ref) |
381 | | ); |
382 | 5 | match (upload_res, bind_res) { |
383 | 5 | (Ok(size), Ok(())) => Ok(size), |
384 | 0 | (Err(e), _) | (_, Err(e)) => Err(e), |
385 | | } |
386 | 5 | .err_tip(|| "Failed to upload file to s3 in single chunk") |
387 | | }; |
388 | | |
389 | | // If we failed to upload the file, check to see if we can retry. |
390 | 5 | let retry_result = result.map_or_else(|mut err| {0 |
391 | | // Ensure our code is Code::Aborted, so the client can retry if possible. |
392 | 0 | err.code = Code::Aborted; |
393 | 0 | let bytes_received = reader.get_bytes_received(); |
394 | 0 | if let Err(try_reset_err) = reader.try_reset_stream() { |
395 | 0 | error!( |
396 | | ?bytes_received, |
397 | | err = ?try_reset_err, |
398 | | "Unable to reset stream after failed upload in S3Store::update" |
399 | | ); |
400 | 0 | return RetryResult::Err(err |
401 | 0 | .merge(try_reset_err) |
402 | 0 | .append(format!("Failed to retry upload with {bytes_received} bytes received in S3Store::update"))); |
403 | 0 | } |
404 | 0 | let err = err.append(format!("Retry on upload happened with {bytes_received} bytes received in S3Store::update")); |
405 | 0 | info!( |
406 | | ?err, |
407 | | ?bytes_received, |
408 | | "Retryable S3 error" |
409 | | ); |
410 | 0 | RetryResult::Retry(err) |
411 | 0 | }, RetryResult::Ok); |
412 | 5 | Some((retry_result, reader)) |
413 | 10 | })) |
414 | | .await; |
415 | | } |
416 | | |
417 | | let upload_id = &self |
418 | | .retrier |
419 | 2 | .retry(unfold((), move |()| async move { |
420 | 2 | let retry_result = self |
421 | 2 | .s3_client |
422 | 2 | .create_multipart_upload() |
423 | 2 | .bucket(&self.bucket) |
424 | 2 | .key(s3_path) |
425 | 2 | .send() |
426 | 2 | .await |
427 | 2 | .map_or_else( |
428 | 0 | |e| { |
429 | 0 | RetryResult::Retry( |
430 | 0 | Error::from_std_err(Code::Aborted, &e) |
431 | 0 | .append("Failed to create multipart upload to s3"), |
432 | 0 | ) |
433 | 0 | }, |
434 | 2 | |CreateMultipartUploadOutput { upload_id, .. }| { |
435 | 2 | upload_id.map_or_else( |
436 | 0 | || { |
437 | 0 | RetryResult::Err(make_err!( |
438 | 0 | Code::Internal, |
439 | 0 | "Expected upload_id to be set by s3 response" |
440 | 0 | )) |
441 | 0 | }, |
442 | | RetryResult::Ok, |
443 | | ) |
444 | 2 | }, |
445 | | ); |
446 | 2 | Some((retry_result, ())) |
447 | 4 | })) |
448 | | .await?; |
449 | | |
450 | | // S3 requires us to upload in parts if the size is greater than 5GB. The part size must be at least |
451 | | // 5MB (except last part) and can have up to 10,000 parts. |
452 | | |
453 | | // Calculate of number of chunks if we upload in 5MB chucks (min chunk size), clamping to |
454 | | // 10,000 parts and correcting for lossy integer division. This provides the |
455 | | let chunk_count = (max_size / MIN_MULTIPART_SIZE).clamp(0, MAX_UPLOAD_PARTS - 1) + 1; |
456 | | |
457 | | // Using clamped first approximation of number of chunks, calculate byte count of each |
458 | | // chunk, excluding last chunk, clamping to min/max upload size 5MB, 5GB. |
459 | | let bytes_per_upload_part = |
460 | | (max_size / chunk_count).clamp(MIN_MULTIPART_SIZE, MAX_MULTIPART_SIZE); |
461 | | |
462 | | // Sanity check before continuing. |
463 | | if !(MIN_MULTIPART_SIZE..MAX_MULTIPART_SIZE).contains(&bytes_per_upload_part) { |
464 | | return Err(make_err!( |
465 | | Code::FailedPrecondition, |
466 | | "Failed to calculate file chuck size (min, max, calc): {MIN_MULTIPART_SIZE}, {MAX_MULTIPART_SIZE}, {bytes_per_upload_part}", |
467 | | )); |
468 | | } |
469 | | |
470 | 2 | let upload_parts = move || async move { |
471 | | // This will ensure we only have `multipart_max_concurrent_uploads` * `bytes_per_upload_part` |
472 | | // bytes in memory at any given time waiting to be uploaded. |
473 | 2 | let (tx, mut rx) = mpsc::channel(self.multipart_max_concurrent_uploads); |
474 | | |
475 | 2 | let read_stream_fut = async move { |
476 | 2 | let retrier = &Pin::get_ref(self).retrier; |
477 | 2 | let mut total_uploaded = 0; |
478 | | // Note: Our break condition is when we reach EOF. |
479 | 9 | for part_number in 1..i32::MAX2 { |
480 | 9 | let write_buf = reader |
481 | 9 | .consume(Some(usize::try_from(bytes_per_upload_part).err_tip( |
482 | | || "Could not convert bytes_per_upload_part to usize", |
483 | 0 | )?)) |
484 | 9 | .await |
485 | 9 | .err_tip(|| "Failed to read chunk in s3_store")?0 ; |
486 | 9 | if write_buf.is_empty() { |
487 | 2 | break; // Reached EOF. |
488 | 7 | } |
489 | | |
490 | 7 | total_uploaded += write_buf.len() as u64; |
491 | | |
492 | 7 | tx.send(retrier.retry(unfold(write_buf, move |write_buf| { |
493 | 7 | async move { |
494 | 7 | let retry_result = self |
495 | 7 | .s3_client |
496 | 7 | .upload_part() |
497 | 7 | .bucket(&self.bucket) |
498 | 7 | .key(s3_path) |
499 | 7 | .upload_id(upload_id) |
500 | 7 | .body(ByteStream::new(SdkBody::from(write_buf.clone()))) |
501 | 7 | .part_number(part_number) |
502 | 7 | .send() |
503 | 7 | .await |
504 | 7 | .map_or_else( |
505 | 0 | |e| { |
506 | 0 | RetryResult::Retry( |
507 | 0 | Error::from_std_err(Code::Aborted, &e).append(format!( |
508 | 0 | "Failed to upload part {part_number} in S3 store" |
509 | 0 | )), |
510 | 0 | ) |
511 | 0 | }, |
512 | 7 | |mut response| { |
513 | 7 | RetryResult::Ok( |
514 | 7 | CompletedPartBuilder::default() |
515 | 7 | // Only set an entity tag if it exists. This saves |
516 | 7 | // 13 bytes per part on the final request if it can |
517 | 7 | // omit the `<ETAG><ETAG/>` string. |
518 | 7 | .set_e_tag(response.e_tag.take()) |
519 | 7 | .part_number(part_number) |
520 | 7 | .build(), |
521 | 7 | ) |
522 | 7 | }, |
523 | | ); |
524 | 7 | Some((retry_result, write_buf)) |
525 | 7 | } |
526 | 7 | }))) |
527 | 7 | .await |
528 | 7 | .map_err(|err| {0 |
529 | 0 | Error::from_std_err(Code::Internal, &err) |
530 | 0 | .append("Failed to send part to channel in s3_store") |
531 | 0 | })?; |
532 | | } |
533 | 2 | Result::<_, Error>::Ok(total_uploaded) |
534 | 2 | } |
535 | 2 | .fuse(); |
536 | | |
537 | 2 | let mut upload_futures = FuturesUnordered::new(); |
538 | 2 | let mut total_uploaded = 0; |
539 | | |
540 | 2 | let mut completed_parts = Vec::with_capacity( |
541 | 2 | usize::try_from(cmp::min(MAX_UPLOAD_PARTS, chunk_count)) |
542 | 2 | .err_tip(|| "Could not convert u64 to usize")?0 , |
543 | | ); |
544 | 2 | tokio::pin!(read_stream_fut); |
545 | | loop { |
546 | 18 | if read_stream_fut.is_terminated() && rx16 .is_empty16 () && upload_futures6 .is_empty6 () { |
547 | 2 | break; // No more data to process. |
548 | 16 | } |
549 | 16 | tokio::select! { |
550 | 16 | result2 = &mut read_stream_fut => { |
551 | 2 | total_uploaded = result?0 ; |
552 | | }, // Return error or wait for other futures. |
553 | 16 | Some(upload_result7 ) = upload_futures.next() => completed_parts7 .push7 (upload_result7 ?0 ), |
554 | 16 | Some(fut7 ) = rx.recv() => upload_futures7 .push7 (fut7 ), |
555 | | } |
556 | | } |
557 | | |
558 | | // Even though the spec does not require parts to be sorted by number, we do it just in case |
559 | | // there's an S3 implementation that requires it. |
560 | 2 | completed_parts.sort_unstable_by_key(|part| part.part_number); |
561 | | |
562 | 2 | self.retrier |
563 | 2 | .retry(unfold(completed_parts, move |completed_parts| async move { |
564 | | Some(( |
565 | 2 | self.s3_client |
566 | 2 | .complete_multipart_upload() |
567 | 2 | .bucket(&self.bucket) |
568 | 2 | .key(s3_path) |
569 | 2 | .multipart_upload( |
570 | 2 | CompletedMultipartUploadBuilder::default() |
571 | 2 | .set_parts(Some(completed_parts.clone())) |
572 | 2 | .build(), |
573 | 2 | ) |
574 | 2 | .upload_id(upload_id) |
575 | 2 | .send() |
576 | 2 | .await |
577 | 2 | .map_or_else( |
578 | 0 | |e| { |
579 | 0 | RetryResult::Retry( |
580 | 0 | Error::from_std_err(Code::Aborted, &e).append( |
581 | 0 | "Failed to complete multipart upload in S3 store", |
582 | 0 | ), |
583 | 0 | ) |
584 | 0 | }, |
585 | 2 | |_| RetryResult::Ok(total_uploaded), |
586 | | ), |
587 | 2 | completed_parts, |
588 | | )) |
589 | 4 | })) |
590 | 2 | .await |
591 | 4 | }; |
592 | | // Upload our parts and complete the multipart upload. |
593 | | // If we fail attempt to abort the multipart upload (cleanup). |
594 | | upload_parts() |
595 | 0 | .or_else(move |mut e| async move { |
596 | 0 | let abort_res = self |
597 | 0 | .s3_client |
598 | 0 | .abort_multipart_upload() |
599 | 0 | .bucket(&self.bucket) |
600 | 0 | .key(s3_path) |
601 | 0 | .upload_id(upload_id) |
602 | 0 | .send() |
603 | 0 | .await; |
604 | 0 | if let Err(abort_err) = abort_res { |
605 | 0 | let err = Error::from_std_err(Code::Aborted, &abort_err) |
606 | 0 | .append("Failed to abort multipart upload in S3 store"); |
607 | 0 | info!(?err, "Multipart upload error"); |
608 | 0 | e = e.merge(err); |
609 | 0 | } |
610 | 0 | Err(e) |
611 | 0 | }) |
612 | | .await |
613 | 7 | } |
614 | | |
615 | | async fn get_part( |
616 | | self: Pin<&Self>, |
617 | | key: StoreKey<'_>, |
618 | | writer: &mut DropCloserWriteHalf, |
619 | | offset: u64, |
620 | | length: Option<u64>, |
621 | 8 | ) -> Result<(), Error> { |
622 | | if is_zero_digest(key.borrow()) { |
623 | | writer |
624 | | .send_eof() |
625 | | .err_tip(|| "Failed to send zero EOF in filesystem store get_part")?; |
626 | | return Ok(()); |
627 | | } |
628 | | |
629 | | let s3_path = &self.make_s3_path(&key); |
630 | | let end_read_byte = length |
631 | 3 | .map_or(Some(None), |length| Some(offset.checked_add(length))) |
632 | | .err_tip(|| "Integer overflow protection triggered")?; |
633 | | |
634 | | self.retrier |
635 | 8 | .retry(unfold(writer, move |writer| async move { |
636 | 8 | let result = self |
637 | 8 | .s3_client |
638 | 8 | .get_object() |
639 | 8 | .bucket(&self.bucket) |
640 | 8 | .key(s3_path) |
641 | 8 | .range(format!( |
642 | | "bytes={}-{}", |
643 | 8 | offset + writer.get_bytes_written(), |
644 | 8 | end_read_byte.map_or_else(String::new, |v| v3 .to_string3 ()) |
645 | | )) |
646 | 8 | .send() |
647 | 8 | .await; |
648 | | |
649 | 8 | let mut s3_in_stream6 = match result { |
650 | 6 | Ok(head_object_output) => head_object_output.body, |
651 | 2 | Err(sdk_error) => match sdk_error.into_service_error() { |
652 | 1 | GetObjectError::NoSuchKey(e) => { |
653 | 1 | return Some(( |
654 | 1 | RetryResult::Err( |
655 | 1 | Error::from_std_err(Code::NotFound, &e) |
656 | 1 | .append("No such key in S3"), |
657 | 1 | ), |
658 | 1 | writer, |
659 | 1 | )); |
660 | | } |
661 | 1 | other => { |
662 | 1 | return Some(( |
663 | 1 | RetryResult::Retry( |
664 | 1 | Error::from_std_err(Code::Unavailable, &other) |
665 | 1 | .append("Unhandled GetObjectError in S3"), |
666 | 1 | ), |
667 | 1 | writer, |
668 | 1 | )); |
669 | | } |
670 | | }, |
671 | | }; |
672 | | |
673 | | // Copy data from s3 input stream to the writer stream. |
674 | 11 | while let Some(maybe_bytes5 ) = s3_in_stream.next().await { |
675 | 5 | match maybe_bytes { |
676 | 5 | Ok(bytes) => { |
677 | 5 | if bytes.is_empty() { |
678 | | // Ignore possible EOF. Different implementations of S3 may or may not |
679 | | // send EOF this way. |
680 | 1 | continue; |
681 | 4 | } |
682 | 4 | if let Err(e0 ) = writer.send(bytes).await { |
683 | 0 | return Some(( |
684 | 0 | RetryResult::Err( |
685 | 0 | Error::from_std_err(Code::Aborted, &e) |
686 | 0 | .append("Error sending bytes to consumer in S3"), |
687 | 0 | ), |
688 | 0 | writer, |
689 | 0 | )); |
690 | 4 | } |
691 | | } |
692 | 0 | Err(e) => { |
693 | 0 | return Some(( |
694 | 0 | RetryResult::Retry( |
695 | 0 | Error::from_std_err(Code::Aborted, &e) |
696 | 0 | .append("Bad bytestream element in S3"), |
697 | 0 | ), |
698 | 0 | writer, |
699 | 0 | )); |
700 | | } |
701 | | } |
702 | | } |
703 | 6 | if let Err(e0 ) = writer.send_eof() { |
704 | 0 | return Some(( |
705 | 0 | RetryResult::Err( |
706 | 0 | Error::from_std_err(Code::Aborted, &e) |
707 | 0 | .append("Failed to send EOF to consumer in S3"), |
708 | 0 | ), |
709 | 0 | writer, |
710 | 0 | )); |
711 | 6 | } |
712 | 6 | Some((RetryResult::Ok(()), writer)) |
713 | 16 | })) |
714 | | .await |
715 | 8 | } |
716 | | |
717 | 0 | fn inner_store(&self, _digest: Option<StoreKey>) -> &'_ dyn StoreDriver { |
718 | 0 | self |
719 | 0 | } |
720 | | |
721 | 0 | fn as_any<'a>(&'a self) -> &'a (dyn core::any::Any + Sync + Send + 'static) { |
722 | 0 | self |
723 | 0 | } |
724 | | |
725 | 0 | fn as_any_arc(self: Arc<Self>) -> Arc<dyn core::any::Any + Sync + Send + 'static> { |
726 | 0 | self |
727 | 0 | } |
728 | | |
729 | 0 | fn register_health(self: Arc<Self>, registry: &mut HealthRegistryBuilder) { |
730 | 0 | registry.register_indicator(self); |
731 | 0 | } |
732 | | |
733 | 0 | fn register_remove_callback(self: Arc<Self>, callback: RemoveCallback) -> Result<(), Error> { |
734 | 0 | self.remove_callbacks.lock().push(callback); |
735 | 0 | Ok(()) |
736 | 0 | } |
737 | | } |
738 | | |
739 | | #[async_trait] |
740 | | impl<I, NowFn> HealthStatusIndicator for S3Store<NowFn> |
741 | | where |
742 | | I: InstantWrapper, |
743 | | NowFn: Fn() -> I + Send + Sync + Unpin + 'static, |
744 | | { |
745 | 0 | fn get_name(&self) -> &'static str { |
746 | 0 | "S3Store" |
747 | 0 | } |
748 | | |
749 | 0 | async fn check_health(&self, namespace: Cow<'static, str>) -> HealthStatus { |
750 | | StoreDriver::check_health(Pin::new(self), namespace).await |
751 | 0 | } |
752 | | } |