Line data Source code
1 : //! Helper functions to download files from remote storage with a RemoteStorage
2 : //!
3 : //! The functions in this module retry failed operations automatically, according
4 : //! to the FAILED_DOWNLOAD_RETRIES constant.
5 :
6 : use std::collections::HashSet;
7 : use std::future::Future;
8 : use std::str::FromStr;
9 : use std::time::SystemTime;
10 :
11 : use anyhow::{anyhow, Context};
12 : use camino::{Utf8Path, Utf8PathBuf};
13 : use pageserver_api::shard::TenantShardId;
14 : use tokio::fs::{self, File, OpenOptions};
15 : use tokio::io::{AsyncSeekExt, AsyncWriteExt};
16 : use tokio_util::io::StreamReader;
17 : use tokio_util::sync::CancellationToken;
18 : use tracing::warn;
19 : use utils::backoff;
20 :
21 : use crate::config::PageServerConf;
22 : use crate::context::RequestContext;
23 : use crate::span::debug_assert_current_span_has_tenant_and_timeline_id;
24 : use crate::tenant::remote_timeline_client::{remote_layer_path, remote_timelines_path};
25 : use crate::tenant::storage_layer::LayerName;
26 : use crate::tenant::Generation;
27 : #[cfg_attr(target_os = "macos", allow(unused_imports))]
28 : use crate::virtual_file::owned_buffers_io::io_buf_ext::IoBufExt;
29 : use crate::virtual_file::{on_fatal_io_error, MaybeFatalIo, VirtualFile};
30 : use crate::TEMP_FILE_SUFFIX;
31 : use remote_storage::{DownloadError, DownloadOpts, GenericRemoteStorage, ListingMode, RemotePath};
32 : use utils::crashsafe::path_with_suffix_extension;
33 : use utils::id::{TenantId, TimelineId};
34 : use utils::pausable_failpoint;
35 :
36 : use super::index::{IndexPart, LayerFileMetadata};
37 : use super::{
38 : parse_remote_index_path, remote_index_path, remote_initdb_archive_path,
39 : remote_initdb_preserved_archive_path, remote_tenant_path, FAILED_DOWNLOAD_WARN_THRESHOLD,
40 : FAILED_REMOTE_OP_RETRIES, INITDB_PATH,
41 : };
42 :
43 : ///
44 : /// If 'metadata' is given, we will validate that the downloaded file's size matches that
45 : /// in the metadata. (In the future, we might do more cross-checks, like CRC validation)
46 : ///
47 : /// Returns the size of the downloaded file.
48 : #[allow(clippy::too_many_arguments)]
49 6 : pub async fn download_layer_file<'a>(
50 6 : conf: &'static PageServerConf,
51 6 : storage: &'a GenericRemoteStorage,
52 6 : tenant_shard_id: TenantShardId,
53 6 : timeline_id: TimelineId,
54 6 : layer_file_name: &'a LayerName,
55 6 : layer_metadata: &'a LayerFileMetadata,
56 6 : local_path: &Utf8Path,
57 6 : cancel: &CancellationToken,
58 6 : ctx: &RequestContext,
59 6 : ) -> Result<u64, DownloadError> {
60 6 : debug_assert_current_span_has_tenant_and_timeline_id();
61 6 :
62 6 : let timeline_path = conf.timeline_path(&tenant_shard_id, &timeline_id);
63 6 :
64 6 : let remote_path = remote_layer_path(
65 6 : &tenant_shard_id.tenant_id,
66 6 : &timeline_id,
67 6 : layer_metadata.shard,
68 6 : layer_file_name,
69 6 : layer_metadata.generation,
70 6 : );
71 6 :
72 6 : // Perform a rename inspired by durable_rename from file_utils.c.
73 6 : // The sequence:
74 6 : // write(tmp)
75 6 : // fsync(tmp)
76 6 : // rename(tmp, new)
77 6 : // fsync(new)
78 6 : // fsync(parent)
79 6 : // For more context about durable_rename check this email from postgres mailing list:
80 6 : // https://www.postgresql.org/message-id/56583BDD.9060302@2ndquadrant.com
81 6 : // If pageserver crashes the temp file will be deleted on startup and re-downloaded.
82 6 : let temp_file_path = path_with_suffix_extension(local_path, TEMP_DOWNLOAD_EXTENSION);
83 :
84 6 : let bytes_amount = download_retry(
85 59 : || async { download_object(storage, &remote_path, &temp_file_path, cancel, ctx).await },
86 6 : &format!("download {remote_path:?}"),
87 6 : cancel,
88 6 : )
89 59 : .await?;
90 :
91 6 : let expected = layer_metadata.file_size;
92 6 : if expected != bytes_amount {
93 0 : return Err(DownloadError::Other(anyhow!(
94 0 : "According to layer file metadata should have downloaded {expected} bytes but downloaded {bytes_amount} bytes into file {temp_file_path:?}",
95 0 : )));
96 6 : }
97 6 :
98 6 : fail::fail_point!("remote-storage-download-pre-rename", |_| {
99 0 : Err(DownloadError::Other(anyhow!(
100 0 : "remote-storage-download-pre-rename failpoint triggered"
101 0 : )))
102 6 : });
103 :
104 6 : fs::rename(&temp_file_path, &local_path)
105 4 : .await
106 6 : .with_context(|| format!("rename download layer file to {local_path}"))
107 6 : .map_err(DownloadError::Other)?;
108 :
109 : // We use fatal_err() below because the after the rename above,
110 : // the in-memory state of the filesystem already has the layer file in its final place,
111 : // and subsequent pageserver code could think it's durable while it really isn't.
112 6 : let work = {
113 6 : let ctx = ctx.detached_child(ctx.task_kind(), ctx.download_behavior());
114 6 : async move {
115 6 : let timeline_dir = VirtualFile::open(&timeline_path, &ctx)
116 3 : .await
117 6 : .fatal_err("VirtualFile::open for timeline dir fsync");
118 6 : timeline_dir
119 6 : .sync_all()
120 3 : .await
121 6 : .fatal_err("VirtualFile::sync_all timeline dir");
122 6 : }
123 : };
124 6 : crate::virtual_file::io_engine::get()
125 6 : .spawn_blocking_and_block_on_if_std(work)
126 9 : .await;
127 :
128 6 : tracing::debug!("download complete: {local_path}");
129 :
130 6 : Ok(bytes_amount)
131 6 : }
132 :
133 : /// Download the object `src_path` in the remote `storage` to local path `dst_path`.
134 : ///
135 : /// If Ok() is returned, the download succeeded and the inode & data have been made durable.
136 : /// (Note that the directory entry for the inode is not made durable.)
137 : /// The file size in bytes is returned.
138 : ///
139 : /// If Err() is returned, there was some error. The file at `dst_path` has been unlinked.
140 : /// The unlinking has _not_ been made durable.
141 6 : async fn download_object<'a>(
142 6 : storage: &'a GenericRemoteStorage,
143 6 : src_path: &RemotePath,
144 6 : dst_path: &Utf8PathBuf,
145 6 : cancel: &CancellationToken,
146 6 : #[cfg_attr(target_os = "macos", allow(unused_variables))] ctx: &RequestContext,
147 6 : ) -> Result<u64, DownloadError> {
148 6 : let res = match crate::virtual_file::io_engine::get() {
149 0 : crate::virtual_file::io_engine::IoEngine::NotSet => panic!("unset"),
150 : crate::virtual_file::io_engine::IoEngine::StdFs => {
151 3 : async {
152 3 : let destination_file = tokio::fs::File::create(dst_path)
153 2 : .await
154 3 : .with_context(|| format!("create a destination file for layer '{dst_path}'"))
155 3 : .map_err(DownloadError::Other)?;
156 :
157 3 : let download = storage
158 3 : .download(src_path, &DownloadOpts::default(), cancel)
159 4 : .await?;
160 :
161 3 : pausable_failpoint!("before-downloading-layer-stream-pausable");
162 :
163 3 : let mut buf_writer =
164 3 : tokio::io::BufWriter::with_capacity(super::BUFFER_SIZE, destination_file);
165 3 :
166 3 : let mut reader = tokio_util::io::StreamReader::new(download.download_stream);
167 :
168 14 : let bytes_amount = tokio::io::copy_buf(&mut reader, &mut buf_writer).await?;
169 3 : buf_writer.flush().await?;
170 :
171 3 : let mut destination_file = buf_writer.into_inner();
172 3 :
173 3 : // Tokio doc here: https://docs.rs/tokio/1.17.0/tokio/fs/struct.File.html states that:
174 3 : // A file will not be closed immediately when it goes out of scope if there are any IO operations
175 3 : // that have not yet completed. To ensure that a file is closed immediately when it is dropped,
176 3 : // you should call flush before dropping it.
177 3 : //
178 3 : // From the tokio code I see that it waits for pending operations to complete. There shouldt be any because
179 3 : // we assume that `destination_file` file is fully written. I e there is no pending .write(...).await operations.
180 3 : // But for additional safety lets check/wait for any pending operations.
181 3 : destination_file
182 3 : .flush()
183 0 : .await
184 3 : .maybe_fatal_err("download_object sync_all")
185 3 : .with_context(|| format!("flush source file at {dst_path}"))
186 3 : .map_err(DownloadError::Other)?;
187 :
188 : // not using sync_data because it can lose file size update
189 3 : destination_file
190 3 : .sync_all()
191 3 : .await
192 3 : .maybe_fatal_err("download_object sync_all")
193 3 : .with_context(|| format!("failed to fsync source file at {dst_path}"))
194 3 : .map_err(DownloadError::Other)?;
195 :
196 3 : Ok(bytes_amount)
197 3 : }
198 25 : .await
199 : }
200 : #[cfg(target_os = "linux")]
201 : crate::virtual_file::io_engine::IoEngine::TokioEpollUring => {
202 : use crate::virtual_file::owned_buffers_io::{self, util::size_tracking_writer};
203 : use bytes::BytesMut;
204 3 : async {
205 3 : let destination_file = VirtualFile::create(dst_path, ctx)
206 3 : .await
207 3 : .with_context(|| format!("create a destination file for layer '{dst_path}'"))
208 3 : .map_err(DownloadError::Other)?;
209 :
210 3 : let mut download = storage
211 3 : .download(src_path, &DownloadOpts::default(), cancel)
212 5 : .await?;
213 :
214 3 : pausable_failpoint!("before-downloading-layer-stream-pausable");
215 :
216 : // TODO: use vectored write (writev) once supported by tokio-epoll-uring.
217 : // There's chunks_vectored() on the stream.
218 3 : let (bytes_amount, destination_file) = async {
219 3 : let size_tracking = size_tracking_writer::Writer::new(destination_file);
220 3 : let mut buffered = owned_buffers_io::write::BufferedWriter::<BytesMut, _>::new(
221 3 : size_tracking,
222 3 : BytesMut::with_capacity(super::BUFFER_SIZE),
223 3 : );
224 18 : while let Some(res) =
225 21 : futures::StreamExt::next(&mut download.download_stream).await
226 : {
227 18 : let chunk = match res {
228 18 : Ok(chunk) => chunk,
229 0 : Err(e) => return Err(e),
230 : };
231 18 : buffered.write_buffered(chunk.slice_len(), ctx).await?;
232 : }
233 3 : let size_tracking = buffered.flush_and_into_inner(ctx).await?;
234 3 : Ok(size_tracking.into_inner())
235 3 : }
236 20 : .await?;
237 :
238 : // not using sync_data because it can lose file size update
239 3 : destination_file
240 3 : .sync_all()
241 3 : .await
242 3 : .maybe_fatal_err("download_object sync_all")
243 3 : .with_context(|| format!("failed to fsync source file at {dst_path}"))
244 3 : .map_err(DownloadError::Other)?;
245 :
246 3 : Ok(bytes_amount)
247 3 : }
248 34 : .await
249 : }
250 : };
251 :
252 : // in case the download failed, clean up
253 6 : match res {
254 6 : Ok(bytes_amount) => Ok(bytes_amount),
255 0 : Err(e) => {
256 0 : if let Err(e) = tokio::fs::remove_file(dst_path).await {
257 0 : if e.kind() != std::io::ErrorKind::NotFound {
258 0 : on_fatal_io_error(&e, &format!("Removing temporary file {dst_path}"));
259 0 : }
260 0 : }
261 0 : Err(e)
262 : }
263 : }
264 6 : }
265 :
266 : const TEMP_DOWNLOAD_EXTENSION: &str = "temp_download";
267 :
268 0 : pub(crate) fn is_temp_download_file(path: &Utf8Path) -> bool {
269 0 : let extension = path.extension();
270 0 : match extension {
271 0 : Some(TEMP_DOWNLOAD_EXTENSION) => true,
272 0 : Some(_) => false,
273 0 : None => false,
274 : }
275 0 : }
276 :
277 186 : async fn list_identifiers<T>(
278 186 : storage: &GenericRemoteStorage,
279 186 : prefix: RemotePath,
280 186 : cancel: CancellationToken,
281 186 : ) -> anyhow::Result<(HashSet<T>, HashSet<String>)>
282 186 : where
283 186 : T: FromStr + Eq + std::hash::Hash,
284 186 : {
285 186 : let listing = download_retry_forever(
286 186 : || storage.list(Some(&prefix), ListingMode::WithDelimiter, None, &cancel),
287 186 : &format!("list identifiers in prefix {prefix}"),
288 186 : &cancel,
289 186 : )
290 716 : .await?;
291 :
292 186 : let mut parsed_ids = HashSet::new();
293 186 : let mut other_prefixes = HashSet::new();
294 :
295 192 : for id_remote_storage_key in listing.prefixes {
296 6 : let object_name = id_remote_storage_key.object_name().ok_or_else(|| {
297 0 : anyhow::anyhow!("failed to get object name for key {id_remote_storage_key}")
298 6 : })?;
299 :
300 6 : match object_name.parse::<T>() {
301 6 : Ok(t) => parsed_ids.insert(t),
302 0 : Err(_) => other_prefixes.insert(object_name.to_string()),
303 : };
304 : }
305 :
306 186 : for object in listing.keys {
307 0 : let object_name = object
308 0 : .key
309 0 : .object_name()
310 0 : .ok_or_else(|| anyhow::anyhow!("object name for key {}", object.key))?;
311 0 : other_prefixes.insert(object_name.to_string());
312 : }
313 :
314 186 : Ok((parsed_ids, other_prefixes))
315 186 : }
316 :
317 : /// List shards of given tenant in remote storage
318 0 : pub(crate) async fn list_remote_tenant_shards(
319 0 : storage: &GenericRemoteStorage,
320 0 : tenant_id: TenantId,
321 0 : cancel: CancellationToken,
322 0 : ) -> anyhow::Result<(HashSet<TenantShardId>, HashSet<String>)> {
323 0 : let remote_path = remote_tenant_path(&TenantShardId::unsharded(tenant_id));
324 0 : list_identifiers::<TenantShardId>(storage, remote_path, cancel).await
325 0 : }
326 :
327 : /// List timelines of given tenant shard in remote storage
328 186 : pub async fn list_remote_timelines(
329 186 : storage: &GenericRemoteStorage,
330 186 : tenant_shard_id: TenantShardId,
331 186 : cancel: CancellationToken,
332 186 : ) -> anyhow::Result<(HashSet<TimelineId>, HashSet<String>)> {
333 186 : fail::fail_point!("storage-sync-list-remote-timelines", |_| {
334 0 : anyhow::bail!("storage-sync-list-remote-timelines");
335 186 : });
336 :
337 186 : let remote_path = remote_timelines_path(&tenant_shard_id).add_trailing_slash();
338 716 : list_identifiers::<TimelineId>(storage, remote_path, cancel).await
339 186 : }
340 :
341 34 : async fn do_download_index_part(
342 34 : storage: &GenericRemoteStorage,
343 34 : tenant_shard_id: &TenantShardId,
344 34 : timeline_id: &TimelineId,
345 34 : index_generation: Generation,
346 34 : cancel: &CancellationToken,
347 34 : ) -> Result<(IndexPart, Generation, SystemTime), DownloadError> {
348 34 : let remote_path = remote_index_path(tenant_shard_id, timeline_id, index_generation);
349 :
350 34 : let (index_part_bytes, index_part_mtime) = download_retry_forever(
351 34 : || async {
352 34 : let download = storage
353 34 : .download(&remote_path, &DownloadOpts::default(), cancel)
354 50 : .await?;
355 :
356 20 : let mut bytes = Vec::new();
357 20 :
358 20 : let stream = download.download_stream;
359 20 : let mut stream = StreamReader::new(stream);
360 20 :
361 20 : tokio::io::copy_buf(&mut stream, &mut bytes).await?;
362 :
363 20 : Ok((bytes, download.last_modified))
364 68 : },
365 34 : &format!("download {remote_path:?}"),
366 34 : cancel,
367 34 : )
368 67 : .await?;
369 :
370 20 : let index_part: IndexPart = serde_json::from_slice(&index_part_bytes)
371 20 : .with_context(|| format!("deserialize index part file at {remote_path:?}"))
372 20 : .map_err(DownloadError::Other)?;
373 :
374 20 : Ok((index_part, index_generation, index_part_mtime))
375 34 : }
376 :
377 : /// index_part.json objects are suffixed with a generation number, so we cannot
378 : /// directly GET the latest index part without doing some probing.
379 : ///
380 : /// In this function we probe for the most recent index in a generation <= our current generation.
381 : /// See "Finding the remote indices for timelines" in docs/rfcs/025-generation-numbers.md
382 20 : #[tracing::instrument(skip_all, fields(generation=?my_generation))]
383 : pub(crate) async fn download_index_part(
384 : storage: &GenericRemoteStorage,
385 : tenant_shard_id: &TenantShardId,
386 : timeline_id: &TimelineId,
387 : my_generation: Generation,
388 : cancel: &CancellationToken,
389 : ) -> Result<(IndexPart, Generation, SystemTime), DownloadError> {
390 : debug_assert_current_span_has_tenant_and_timeline_id();
391 :
392 : if my_generation.is_none() {
393 : // Operating without generations: just fetch the generation-less path
394 : return do_download_index_part(
395 : storage,
396 : tenant_shard_id,
397 : timeline_id,
398 : my_generation,
399 : cancel,
400 : )
401 : .await;
402 : }
403 :
404 : // Stale case: If we were intentionally attached in a stale generation, there may already be a remote
405 : // index in our generation.
406 : //
407 : // This is an optimization to avoid doing the listing for the general case below.
408 : let res =
409 : do_download_index_part(storage, tenant_shard_id, timeline_id, my_generation, cancel).await;
410 : match res {
411 : Ok(index_part) => {
412 : tracing::debug!(
413 : "Found index_part from current generation (this is a stale attachment)"
414 : );
415 : return Ok(index_part);
416 : }
417 : Err(DownloadError::NotFound) => {}
418 : Err(e) => return Err(e),
419 : };
420 :
421 : // Typical case: the previous generation of this tenant was running healthily, and had uploaded
422 : // and index part. We may safely start from this index without doing a listing, because:
423 : // - We checked for current generation case above
424 : // - generations > my_generation are to be ignored
425 : // - any other indices that exist would have an older generation than `previous_gen`, and
426 : // we want to find the most recent index from a previous generation.
427 : //
428 : // This is an optimization to avoid doing the listing for the general case below.
429 : let res = do_download_index_part(
430 : storage,
431 : tenant_shard_id,
432 : timeline_id,
433 : my_generation.previous(),
434 : cancel,
435 : )
436 : .await;
437 : match res {
438 : Ok(index_part) => {
439 : tracing::debug!("Found index_part from previous generation");
440 : return Ok(index_part);
441 : }
442 : Err(DownloadError::NotFound) => {
443 : tracing::debug!(
444 : "No index_part found from previous generation, falling back to listing"
445 : );
446 : }
447 : Err(e) => {
448 : return Err(e);
449 : }
450 : }
451 :
452 : // General case/fallback: if there is no index at my_generation or prev_generation, then list all index_part.json
453 : // objects, and select the highest one with a generation <= my_generation. Constructing the prefix is equivalent
454 : // to constructing a full index path with no generation, because the generation is a suffix.
455 : let index_prefix = remote_index_path(tenant_shard_id, timeline_id, Generation::none());
456 :
457 : let indices = download_retry(
458 6 : || async {
459 6 : storage
460 6 : .list(Some(&index_prefix), ListingMode::NoDelimiter, None, cancel)
461 6 : .await
462 12 : },
463 : "list index_part files",
464 : cancel,
465 : )
466 : .await?
467 : .keys;
468 :
469 : // General case logic for which index to use: the latest index whose generation
470 : // is <= our own. See "Finding the remote indices for timelines" in docs/rfcs/025-generation-numbers.md
471 : let max_previous_generation = indices
472 : .into_iter()
473 18 : .filter_map(|o| parse_remote_index_path(o.key))
474 12 : .filter(|g| g <= &my_generation)
475 : .max();
476 :
477 : match max_previous_generation {
478 : Some(g) => {
479 : tracing::debug!("Found index_part in generation {g:?}");
480 : do_download_index_part(storage, tenant_shard_id, timeline_id, g, cancel).await
481 : }
482 : None => {
483 : // Migration from legacy pre-generation state: we have a generation but no prior
484 : // attached pageservers did. Try to load from a no-generation path.
485 : tracing::debug!("No index_part.json* found");
486 : do_download_index_part(
487 : storage,
488 : tenant_shard_id,
489 : timeline_id,
490 : Generation::none(),
491 : cancel,
492 : )
493 : .await
494 : }
495 : }
496 : }
497 :
498 2 : pub(crate) async fn download_initdb_tar_zst(
499 2 : conf: &'static PageServerConf,
500 2 : storage: &GenericRemoteStorage,
501 2 : tenant_shard_id: &TenantShardId,
502 2 : timeline_id: &TimelineId,
503 2 : cancel: &CancellationToken,
504 2 : ) -> Result<(Utf8PathBuf, File), DownloadError> {
505 2 : debug_assert_current_span_has_tenant_and_timeline_id();
506 2 :
507 2 : let remote_path = remote_initdb_archive_path(&tenant_shard_id.tenant_id, timeline_id);
508 2 :
509 2 : let remote_preserved_path =
510 2 : remote_initdb_preserved_archive_path(&tenant_shard_id.tenant_id, timeline_id);
511 2 :
512 2 : let timeline_path = conf.timelines_path(tenant_shard_id);
513 2 :
514 2 : if !timeline_path.exists() {
515 0 : tokio::fs::create_dir_all(&timeline_path)
516 0 : .await
517 0 : .with_context(|| format!("timeline dir creation {timeline_path}"))
518 0 : .map_err(DownloadError::Other)?;
519 2 : }
520 2 : let temp_path = timeline_path.join(format!(
521 2 : "{INITDB_PATH}.download-{timeline_id}.{TEMP_FILE_SUFFIX}"
522 2 : ));
523 :
524 2 : let file = download_retry(
525 2 : || async {
526 2 : let file = OpenOptions::new()
527 2 : .create(true)
528 2 : .truncate(true)
529 2 : .read(true)
530 2 : .write(true)
531 2 : .open(&temp_path)
532 2 : .await
533 2 : .with_context(|| format!("tempfile creation {temp_path}"))
534 2 : .map_err(DownloadError::Other)?;
535 :
536 2 : let download = match storage
537 2 : .download(&remote_path, &DownloadOpts::default(), cancel)
538 4 : .await
539 : {
540 2 : Ok(dl) => dl,
541 : Err(DownloadError::NotFound) => {
542 0 : storage
543 0 : .download(&remote_preserved_path, &DownloadOpts::default(), cancel)
544 0 : .await?
545 : }
546 0 : Err(other) => Err(other)?,
547 : };
548 2 : let mut download = tokio_util::io::StreamReader::new(download.download_stream);
549 2 : let mut writer = tokio::io::BufWriter::with_capacity(super::BUFFER_SIZE, file);
550 2 :
551 357 : tokio::io::copy_buf(&mut download, &mut writer).await?;
552 :
553 2 : let mut file = writer.into_inner();
554 2 :
555 2 : file.seek(std::io::SeekFrom::Start(0))
556 0 : .await
557 2 : .with_context(|| format!("rewinding initdb.tar.zst at: {remote_path:?}"))
558 2 : .map_err(DownloadError::Other)?;
559 :
560 2 : Ok(file)
561 4 : },
562 2 : &format!("download {remote_path}"),
563 2 : cancel,
564 2 : )
565 363 : .await
566 2 : .inspect_err(|_e| {
567 : // Do a best-effort attempt at deleting the temporary file upon encountering an error.
568 : // We don't have async here nor do we want to pile on any extra errors.
569 0 : if let Err(e) = std::fs::remove_file(&temp_path) {
570 0 : if e.kind() != std::io::ErrorKind::NotFound {
571 0 : warn!("error deleting temporary file {temp_path}: {e}");
572 0 : }
573 0 : }
574 2 : })?;
575 :
576 2 : Ok((temp_path, file))
577 2 : }
578 :
579 : /// Helper function to handle retries for a download operation.
580 : ///
581 : /// Remote operations can fail due to rate limits (S3), spurious network
582 : /// problems, or other external reasons. Retry FAILED_DOWNLOAD_RETRIES times,
583 : /// with backoff.
584 : ///
585 : /// (See similar logic for uploads in `perform_upload_task`)
586 14 : pub(super) async fn download_retry<T, O, F>(
587 14 : op: O,
588 14 : description: &str,
589 14 : cancel: &CancellationToken,
590 14 : ) -> Result<T, DownloadError>
591 14 : where
592 14 : O: FnMut() -> F,
593 14 : F: Future<Output = Result<T, DownloadError>>,
594 14 : {
595 14 : backoff::retry(
596 14 : op,
597 14 : DownloadError::is_permanent,
598 14 : FAILED_DOWNLOAD_WARN_THRESHOLD,
599 14 : FAILED_REMOTE_OP_RETRIES,
600 14 : description,
601 14 : cancel,
602 14 : )
603 428 : .await
604 14 : .ok_or_else(|| DownloadError::Cancelled)
605 14 : .and_then(|x| x)
606 14 : }
607 :
608 220 : async fn download_retry_forever<T, O, F>(
609 220 : op: O,
610 220 : description: &str,
611 220 : cancel: &CancellationToken,
612 220 : ) -> Result<T, DownloadError>
613 220 : where
614 220 : O: FnMut() -> F,
615 220 : F: Future<Output = Result<T, DownloadError>>,
616 220 : {
617 220 : backoff::retry(
618 220 : op,
619 220 : DownloadError::is_permanent,
620 220 : FAILED_DOWNLOAD_WARN_THRESHOLD,
621 220 : u32::MAX,
622 220 : description,
623 220 : cancel,
624 220 : )
625 783 : .await
626 220 : .ok_or_else(|| DownloadError::Cancelled)
627 220 : .and_then(|x| x)
628 220 : }
|
