/build/source/nativelink-store/src/gcs_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::fmt::Debug; |
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 bytes::Bytes; |
23 | | use futures::stream::{FuturesUnordered, unfold}; |
24 | | use futures::{StreamExt, TryStreamExt}; |
25 | | use nativelink_config::stores::ExperimentalGcsSpec; |
26 | | use nativelink_error::{Code, Error, ResultExt, make_err}; |
27 | | use nativelink_metric::MetricsComponent; |
28 | | use nativelink_util::buf_channel::{DropCloserReadHalf, DropCloserWriteHalf}; |
29 | | use nativelink_util::health_utils::{HealthRegistryBuilder, HealthStatus, HealthStatusIndicator}; |
30 | | use nativelink_util::instant_wrapper::InstantWrapper; |
31 | | use nativelink_util::retry::{Retrier, RetryResult}; |
32 | | use nativelink_util::store_trait::{ |
33 | | RemoveCallback, StoreDriver, StoreKey, StoreOptimizations, UploadSizeInfo, |
34 | | }; |
35 | | use rand::Rng; |
36 | | use tokio::time::{sleep, timeout}; |
37 | | use tracing::warn; |
38 | | |
39 | | use crate::cas_utils::is_zero_digest; |
40 | | use crate::gcs_client::client::{GcsClient, GcsOperations}; |
41 | | use crate::gcs_client::types::{ |
42 | | CHUNK_SIZE, DEFAULT_CONCURRENT_UPLOADS, DEFAULT_MAX_RETRY_BUFFER_PER_REQUEST, |
43 | | MIN_MULTIPART_SIZE, ObjectPath, |
44 | | }; |
45 | | |
46 | | #[derive(MetricsComponent, Debug)] |
47 | | pub struct GcsStore<Client: GcsOperations, NowFn> { |
48 | | client: Arc<Client>, |
49 | | now_fn: NowFn, |
50 | | #[metric(help = "The bucket name for the GCS store")] |
51 | | bucket: String, |
52 | | #[metric(help = "The key prefix for the GCS store")] |
53 | | key_prefix: String, |
54 | | retrier: Retrier, |
55 | | #[metric(help = "The number of seconds to consider an object expired")] |
56 | | consider_expired_after_s: i64, |
57 | | #[metric(help = "The number of bytes to buffer for retrying requests")] |
58 | | max_retry_buffer_size: usize, |
59 | | #[metric(help = "The size of chunks for resumable uploads")] |
60 | | max_chunk_size: usize, |
61 | | #[metric(help = "The number of concurrent uploads allowed")] |
62 | | max_concurrent_uploads: usize, |
63 | | } |
64 | | |
65 | | impl<I, NowFn> GcsStore<GcsClient, NowFn> |
66 | | where |
67 | | I: InstantWrapper, |
68 | | NowFn: Fn() -> I + Send + Sync + Unpin + 'static, |
69 | | { |
70 | 0 | pub async fn new(spec: &ExperimentalGcsSpec, now_fn: NowFn) -> Result<Arc<Self>, Error> { |
71 | 0 | let client = Arc::new(GcsClient::new(spec).await?); |
72 | 0 | Self::new_with_ops(spec, client, now_fn) |
73 | 0 | } |
74 | | } |
75 | | |
76 | | impl<I, Client, NowFn> GcsStore<Client, NowFn> |
77 | | where |
78 | | I: InstantWrapper, |
79 | | Client: GcsOperations + Send + Sync, |
80 | | NowFn: Fn() -> I + Send + Sync + Unpin + 'static, |
81 | | { |
82 | | // Primarily used for injecting a mock or real operations implementation |
83 | 20 | pub fn new_with_ops( |
84 | 20 | spec: &ExperimentalGcsSpec, |
85 | 20 | client: Arc<Client>, |
86 | 20 | now_fn: NowFn, |
87 | 20 | ) -> Result<Arc<Self>, Error> { |
88 | | // Chunks must be a multiple of 256kb according to the documentation. |
89 | | const CHUNK_MULTIPLE: usize = 256 * 1024; |
90 | | |
91 | 20 | let max_connections = spec |
92 | 20 | .common |
93 | 20 | .multipart_max_concurrent_uploads |
94 | 20 | .unwrap_or(DEFAULT_CONCURRENT_UPLOADS); |
95 | | |
96 | 20 | let jitter_amt = spec.common.retry.jitter; |
97 | 20 | let jitter_fn = Arc::new(move |delay: Duration| {2 |
98 | 2 | if jitter_amt == 0.0 { |
99 | 2 | return delay; |
100 | 0 | } |
101 | 0 | delay.mul_f32(jitter_amt.mul_add(rand::rng().random::<f32>() - 0.5, 1.)) |
102 | 2 | }); |
103 | | |
104 | 20 | let max_chunk_size = |
105 | 20 | core::cmp::min(spec.resumable_chunk_size.unwrap_or(CHUNK_SIZE), CHUNK_SIZE); |
106 | | |
107 | 20 | let max_chunk_size = if max_chunk_size.is_multiple_of(CHUNK_MULTIPLE) { |
108 | 0 | max_chunk_size |
109 | | } else { |
110 | 20 | ((max_chunk_size + CHUNK_MULTIPLE / 2) / CHUNK_MULTIPLE) * CHUNK_MULTIPLE |
111 | | }; |
112 | | |
113 | 20 | let max_retry_buffer_size = spec |
114 | 20 | .common |
115 | 20 | .max_retry_buffer_per_request |
116 | 20 | .unwrap_or(DEFAULT_MAX_RETRY_BUFFER_PER_REQUEST); |
117 | | |
118 | | // The retry buffer should be at least as big as the chunk size. |
119 | 20 | let max_retry_buffer_size = if max_retry_buffer_size < max_chunk_size { |
120 | 0 | max_chunk_size |
121 | | } else { |
122 | 20 | max_retry_buffer_size |
123 | | }; |
124 | | |
125 | 20 | Ok(Arc::new(Self { |
126 | 20 | client, |
127 | 20 | now_fn, |
128 | 20 | bucket: spec.bucket.clone(), |
129 | 20 | key_prefix: spec |
130 | 20 | .common |
131 | 20 | .key_prefix |
132 | 20 | .as_ref() |
133 | 20 | .unwrap_or(&String::new()) |
134 | 20 | .clone(), |
135 | 20 | retrier: Retrier::new( |
136 | 20 | Arc::new(|duration| Box::pin2 (sleep2 (duration2 ))), |
137 | 20 | jitter_fn, |
138 | 20 | spec.common.retry.clone(), |
139 | | ), |
140 | 20 | consider_expired_after_s: i64::from(spec.common.consider_expired_after_s), |
141 | 20 | max_retry_buffer_size, |
142 | 20 | max_chunk_size, |
143 | 20 | max_concurrent_uploads: max_connections, |
144 | | })) |
145 | 20 | } |
146 | | |
147 | 9 | async fn has(self: Pin<&Self>, key: &StoreKey<'_>) -> Result<Option<u64>, Error> { |
148 | 9 | let object_path = self.make_object_path(key); |
149 | 9 | let client = &self.client; |
150 | 9 | let consider_expired_after_s = self.consider_expired_after_s; |
151 | 9 | let now_fn = &self.now_fn; |
152 | | |
153 | 9 | self.retrier |
154 | 9 | .retry(unfold(object_path, move |object_path| async move { |
155 | 9 | match client.read_object_metadata(&object_path).await.err_tip(|| {2 |
156 | 2 | format!( |
157 | | "Error while trying to read - bucket: {} path: {}", |
158 | | object_path.bucket, object_path.path |
159 | | ) |
160 | 2 | }) { |
161 | 4 | Ok(Some(metadata)) => { |
162 | 4 | if consider_expired_after_s != 0 |
163 | 2 | && let Some(update_time) = &metadata.update_time |
164 | | { |
165 | 2 | let now_s = now_fn().unix_timestamp() as i64; |
166 | 2 | if update_time.seconds + consider_expired_after_s <= now_s { |
167 | 1 | return Some((RetryResult::Ok(None), object_path)); |
168 | 1 | } |
169 | 2 | } |
170 | | |
171 | 3 | if metadata.size >= 0 { |
172 | 3 | Some((RetryResult::Ok(Some(metadata.size as u64)), object_path)) |
173 | | } else { |
174 | 0 | Some(( |
175 | 0 | RetryResult::Err(make_err!( |
176 | 0 | Code::InvalidArgument, |
177 | 0 | "Invalid metadata size in GCS: {}", |
178 | 0 | metadata.size |
179 | 0 | )), |
180 | 0 | object_path, |
181 | 0 | )) |
182 | | } |
183 | | } |
184 | 3 | Ok(None) => Some((RetryResult::Ok(None), object_path)), |
185 | 2 | Err(e1 ) if e.code == Code::NotFound1 => { |
186 | 1 | Some((RetryResult::Ok(None), object_path)) |
187 | | } |
188 | 1 | Err(e) => Some((RetryResult::Retry(e), object_path)), |
189 | | } |
190 | 18 | })) |
191 | 9 | .await |
192 | 9 | } |
193 | | |
194 | 16 | fn make_object_path(&self, key: &StoreKey) -> ObjectPath { |
195 | 16 | ObjectPath::new( |
196 | 16 | self.bucket.clone(), |
197 | 16 | &format!("{}{}", self.key_prefix, key.as_str()), |
198 | | ) |
199 | 16 | } |
200 | | } |
201 | | |
202 | | #[async_trait] |
203 | | impl<I, Client, NowFn> StoreDriver for GcsStore<Client, NowFn> |
204 | | where |
205 | | I: InstantWrapper, |
206 | | Client: GcsOperations + 'static, |
207 | | NowFn: Fn() -> I + Send + Sync + Unpin + 'static, |
208 | | { |
209 | 0 | async fn post_init(self: Arc<Self>) -> Result<(), Error> { |
210 | | Ok(()) |
211 | 0 | } |
212 | | |
213 | | async fn has_with_results( |
214 | | self: Pin<&Self>, |
215 | | keys: &[StoreKey<'_>], |
216 | | results: &mut [Option<u64>], |
217 | 8 | ) -> Result<(), Error> { |
218 | | keys.iter() |
219 | | .zip(results.iter_mut()) |
220 | 10 | .map(|(key, result)| async move { |
221 | 10 | if is_zero_digest(key.borrow()) { |
222 | 1 | *result = Some(0); |
223 | 1 | return Ok(()); |
224 | 9 | } |
225 | 9 | *result = self.has(key).await?1 ; |
226 | 8 | Ok(()) |
227 | 20 | }) |
228 | | .collect::<FuturesUnordered<_>>() |
229 | | .try_collect() |
230 | | .await |
231 | 8 | } |
232 | | |
233 | 0 | fn optimized_for(&self, optimization: StoreOptimizations) -> bool { |
234 | 0 | matches!(optimization, StoreOptimizations::LazyExistenceOnSync) |
235 | 0 | } |
236 | | |
237 | | async fn update( |
238 | | self: Pin<&Self>, |
239 | | digest: StoreKey<'_>, |
240 | | mut reader: DropCloserReadHalf, |
241 | | upload_size: UploadSizeInfo, |
242 | 4 | ) -> Result<u64, Error> { |
243 | | if is_zero_digest(digest.borrow()) { |
244 | 2 | return reader.recv().await.and_then(|should_be_empty| { |
245 | 2 | if should_be_empty.is_empty() { |
246 | 1 | Ok(0) |
247 | | } else { |
248 | 1 | Err(make_err!(Code::Internal, "Zero byte hash not empty")) |
249 | | } |
250 | 2 | }); |
251 | | } |
252 | | |
253 | | let object_path = self.make_object_path(&digest); |
254 | | |
255 | | reader.set_max_recent_data_size( |
256 | | u64::try_from(self.max_retry_buffer_size) |
257 | | .err_tip(|| "Could not convert max_retry_buffer_size to u64")?, |
258 | | ); |
259 | | |
260 | | // For small files with exact size, we'll use simple upload |
261 | | if let UploadSizeInfo::ExactSize(size) = upload_size |
262 | | && size < MIN_MULTIPART_SIZE |
263 | | { |
264 | | let content = reader.consume(Some(usize::try_from(size)?)).await?; |
265 | | let content_len = content.len() as u64; |
266 | | let client = &self.client; |
267 | | |
268 | | return self |
269 | | .retrier |
270 | 1 | .retry(unfold(content, |content| async { |
271 | 1 | match client.write_object(&object_path, content.to_vec()).await { |
272 | 1 | Ok(()) => Some((RetryResult::Ok(content_len), content)), |
273 | 0 | Err(e) => Some((RetryResult::Retry(e), content)), |
274 | | } |
275 | 2 | })) |
276 | | .await; |
277 | | } |
278 | | |
279 | | // For larger files, we'll use resumable upload |
280 | | // Stream and upload data in chunks |
281 | | let mut offset = 0u64; |
282 | | let mut total_size = if let UploadSizeInfo::ExactSize(size) = upload_size { |
283 | | Some(size) |
284 | | } else { |
285 | | None |
286 | | }; |
287 | | let mut upload_id: Option<String> = None; |
288 | | let client = &self.client; |
289 | | |
290 | | loop { |
291 | | let chunk = reader.consume(Some(self.max_chunk_size)).await?; |
292 | | if chunk.is_empty() { |
293 | | break; |
294 | | } |
295 | | // If a full chunk wasn't read, then this is the full length. |
296 | | if chunk.len() < self.max_chunk_size { |
297 | | total_size = Some(offset + chunk.len() as u64); |
298 | | } |
299 | | |
300 | | let upload_id_ref = if let Some(upload_id_ref) = &upload_id { |
301 | | upload_id_ref |
302 | | } else { |
303 | | // Initiate the upload session on the first non-empty chunk. |
304 | | upload_id = Some( |
305 | | self.retrier |
306 | 1 | .retry(unfold((), |()| async { |
307 | 1 | match client.start_resumable_write(&object_path).await { |
308 | 1 | Ok(id) => Some((RetryResult::Ok(id), ())), |
309 | 0 | Err(e) => Some(( |
310 | 0 | RetryResult::Retry(make_err!( |
311 | 0 | Code::Aborted, |
312 | 0 | "Failed to start resumable upload: {:?}", |
313 | 0 | e |
314 | 0 | )), |
315 | 0 | (), |
316 | 0 | )), |
317 | | } |
318 | 2 | })) |
319 | | .await?, |
320 | | ); |
321 | | upload_id.as_deref().unwrap() |
322 | | }; |
323 | | |
324 | | let current_offset = offset; |
325 | | offset += chunk.len() as u64; |
326 | | |
327 | | // Uploading the chunk with a retry |
328 | | let object_path_ref = &object_path; |
329 | | self.retrier |
330 | 6 | .retry(unfold(chunk, |chunk| async move { |
331 | 6 | match client |
332 | 6 | .upload_chunk( |
333 | 6 | upload_id_ref, |
334 | 6 | object_path_ref, |
335 | 6 | chunk.clone(), |
336 | 6 | current_offset, |
337 | 6 | offset, |
338 | 6 | total_size, |
339 | 6 | ) |
340 | 6 | .await |
341 | | { |
342 | 6 | Ok(()) => Some((RetryResult::Ok(()), chunk)), |
343 | 0 | Err(e) => Some((RetryResult::Retry(e), chunk)), |
344 | | } |
345 | 12 | })) |
346 | | .await?; |
347 | | } |
348 | | |
349 | | // Handle the case that the stream was of unknown length and |
350 | | // happened to be an exact multiple of chunk size. |
351 | | if let Some(upload_id_ref) = &upload_id { |
352 | | if total_size.is_none() { |
353 | | let object_path_ref = &object_path; |
354 | | self.retrier |
355 | 0 | .retry(unfold((), |()| async move { |
356 | 0 | match client |
357 | 0 | .upload_chunk( |
358 | 0 | upload_id_ref, |
359 | 0 | object_path_ref, |
360 | 0 | Bytes::new(), |
361 | 0 | offset, |
362 | 0 | offset, |
363 | 0 | Some(offset), |
364 | 0 | ) |
365 | 0 | .await |
366 | | { |
367 | 0 | Ok(()) => Some((RetryResult::Ok(offset), ())), |
368 | 0 | Err(e) => Some((RetryResult::Retry(e), ())), |
369 | | } |
370 | 0 | })) |
371 | | .await?; |
372 | | } |
373 | | } else { |
374 | | // Handle streamed empty file. |
375 | | return self |
376 | | .retrier |
377 | 0 | .retry(unfold((), |()| async { |
378 | 0 | match client.write_object(&object_path, Vec::new()).await { |
379 | 0 | Ok(()) => Some((RetryResult::Ok(0), ())), |
380 | 0 | Err(e) => Some((RetryResult::Retry(e), ())), |
381 | | } |
382 | 0 | })) |
383 | | .await; |
384 | | } |
385 | | |
386 | | // Verifying if the upload was successful |
387 | | self.retrier |
388 | 1 | .retry(unfold((), |()| async { |
389 | 1 | match client.object_exists(&object_path).await { |
390 | 1 | Ok(true) => Some((RetryResult::Ok(()), ())), |
391 | 0 | Ok(false) => Some(( |
392 | 0 | RetryResult::Retry(make_err!( |
393 | 0 | Code::Internal, |
394 | 0 | "Object not found after upload completion" |
395 | 0 | )), |
396 | 0 | (), |
397 | 0 | )), |
398 | 0 | Err(e) => Some((RetryResult::Retry(e), ())), |
399 | | } |
400 | 2 | })) |
401 | | .await?; |
402 | | |
403 | | Ok(offset) |
404 | 4 | } |
405 | | |
406 | | async fn get_part( |
407 | | self: Pin<&Self>, |
408 | | key: StoreKey<'_>, |
409 | | writer: &mut DropCloserWriteHalf, |
410 | | offset: u64, |
411 | | length: Option<u64>, |
412 | 6 | ) -> Result<(), Error> { |
413 | | if is_zero_digest(key.borrow()) { |
414 | | writer.send_eof()?; |
415 | | return Ok(()); |
416 | | } |
417 | | |
418 | | let object_path = self.make_object_path(&key); |
419 | 1 | let end_offset = length.map(|len| offset + len); |
420 | | let client = &self.client; |
421 | | |
422 | | let object_path_ref = &object_path; |
423 | | self.retrier |
424 | | .retry(unfold( |
425 | | (offset, writer), |
426 | 7 | |(mut offset, writer)| async move { |
427 | 7 | let mut stream2 = match client |
428 | 7 | .read_object_content(object_path_ref, offset, end_offset) |
429 | 7 | .await |
430 | | { |
431 | 2 | Ok(stream) => stream, |
432 | | // NotFound is intentionally not special-cased here: |
433 | | // the retrier doesn't retry NotFound by default, but |
434 | | // emitting `Retry` lets `retry.retry_on_errors` opt |
435 | | // reads into retrying read-after-write races where |
436 | | // an object is still finalizing or being repopulated. |
437 | 5 | Err(e) => return Some((RetryResult::Retry(e), (offset, writer))), |
438 | | }; |
439 | | |
440 | 4 | while let Some(next_chunk2 ) = stream.next().await { |
441 | 2 | match next_chunk { |
442 | 2 | Ok(bytes) => { |
443 | 2 | offset += bytes.len() as u64; |
444 | 2 | if let Err(err0 ) = writer.send(bytes).await { |
445 | 0 | return Some((RetryResult::Err(err), (offset, writer))); |
446 | 2 | } |
447 | | } |
448 | 0 | Err(err) => return Some((RetryResult::Retry(err), (offset, writer))), |
449 | | } |
450 | | } |
451 | | |
452 | 2 | if let Err(err0 ) = writer.send_eof() { |
453 | 0 | return Some((RetryResult::Err(err), (offset, writer))); |
454 | 2 | } |
455 | | |
456 | 2 | Some((RetryResult::Ok(()), (offset, writer))) |
457 | 14 | }, |
458 | | )) |
459 | | .await |
460 | 6 | } |
461 | | |
462 | 0 | fn inner_store(&self, _digest: Option<StoreKey>) -> &'_ dyn StoreDriver { |
463 | 0 | self |
464 | 0 | } |
465 | | |
466 | 0 | fn as_any<'a>(&'a self) -> &'a (dyn core::any::Any + Sync + Send + 'static) { |
467 | 0 | self |
468 | 0 | } |
469 | | |
470 | 0 | fn as_any_arc(self: Arc<Self>) -> Arc<dyn core::any::Any + Sync + Send + 'static> { |
471 | 0 | self |
472 | 0 | } |
473 | | |
474 | 0 | fn register_health(self: Arc<Self>, registry: &mut HealthRegistryBuilder) { |
475 | 0 | registry.register_indicator(self); |
476 | 0 | } |
477 | | |
478 | 0 | fn register_remove_callback(self: Arc<Self>, _callback: RemoveCallback) -> Result<(), Error> { |
479 | | // As we're backed by GCS, this store doesn't actually drop stuff |
480 | | // so we can actually just ignore this |
481 | 0 | Ok(()) |
482 | 0 | } |
483 | | } |
484 | | |
485 | | #[async_trait] |
486 | | impl<I, Client, NowFn> HealthStatusIndicator for GcsStore<Client, NowFn> |
487 | | where |
488 | | I: InstantWrapper, |
489 | | Client: GcsOperations + 'static, |
490 | | NowFn: Fn() -> I + Send + Sync + Unpin + 'static, |
491 | | { |
492 | 0 | fn get_name(&self) -> &'static str { |
493 | 0 | "GcsStore" |
494 | 0 | } |
495 | | |
496 | | /// Lightweight probe: a single `object_exists` against a fixed |
497 | | /// never-existing path. Shares no resources with production traffic |
498 | | /// and stays well under the `HealthServer` per-indicator budget. |
499 | 2 | async fn check_health(&self, _namespace: Cow<'static, str>) -> HealthStatus { |
500 | | const HEALTH_PROBE_TIMEOUT: Duration = Duration::from_secs(2); |
501 | | |
502 | | let probe_path = ObjectPath::new( |
503 | | self.bucket.clone(), |
504 | | "__nativelink_health_probe__/does-not-exist", |
505 | | ); |
506 | | |
507 | | let probe = self.client.object_exists(&probe_path); |
508 | | match timeout(HEALTH_PROBE_TIMEOUT, probe).await { |
509 | | Ok(Ok(_)) => HealthStatus::new_ok(self, "GcsStore::check_health: ok".into()), |
510 | | Ok(Err(e)) => { |
511 | | warn!(?e, "GcsStore::check_health: object_exists errored"); |
512 | | HealthStatus::new_failed( |
513 | | self, |
514 | | format!("GcsStore::check_health: object_exists errored: {e}").into(), |
515 | | ) |
516 | | } |
517 | | Err(_) => { |
518 | | warn!( |
519 | | timeout_secs = HEALTH_PROBE_TIMEOUT.as_secs(), |
520 | | "GcsStore::check_health: probe timed out", |
521 | | ); |
522 | | HealthStatus::Timeout { |
523 | | struct_name: self.struct_name(), |
524 | | } |
525 | | } |
526 | | } |
527 | 2 | } |
528 | | } |