Line data Source code
1 : use std::collections::{HashMap, HashSet};
2 : use std::pin::Pin;
3 : use std::str::FromStr;
4 : use std::sync::Arc;
5 : use std::time::{Duration, Instant, SystemTime};
6 :
7 : use camino::Utf8PathBuf;
8 : use chrono::format::{DelayedFormat, StrftimeItems};
9 : use futures::Future;
10 : use metrics::UIntGauge;
11 : use pageserver_api::models::SecondaryProgress;
12 : use pageserver_api::shard::TenantShardId;
13 : use remote_storage::{DownloadError, DownloadKind, DownloadOpts, Etag, GenericRemoteStorage};
14 : use tokio_util::sync::CancellationToken;
15 : use tracing::{Instrument, info_span, instrument, warn};
16 : use utils::completion::Barrier;
17 : use utils::crashsafe::path_with_suffix_extension;
18 : use utils::id::TimelineId;
19 : use utils::{backoff, failpoint_support, fs_ext, pausable_failpoint, serde_system_time};
20 :
21 : use super::heatmap::{HeatMapLayer, HeatMapTenant, HeatMapTimeline};
22 : use super::scheduler::{
23 : self, Completion, JobGenerator, SchedulingResult, TenantBackgroundJobs, period_jitter,
24 : period_warmup,
25 : };
26 : use super::{
27 : CommandRequest, DownloadCommand, GetTenantError, SecondaryTenant, SecondaryTenantError,
28 : };
29 : use crate::TEMP_FILE_SUFFIX;
30 : use crate::config::PageServerConf;
31 : use crate::context::RequestContext;
32 : use crate::disk_usage_eviction_task::{
33 : DiskUsageEvictionInfo, EvictionCandidate, EvictionLayer, EvictionSecondaryLayer, finite_f32,
34 : };
35 : use crate::metrics::SECONDARY_MODE;
36 : use crate::tenant::config::SecondaryLocationConfig;
37 : use crate::tenant::debug_assert_current_span_has_tenant_and_timeline_id;
38 : use crate::tenant::ephemeral_file::is_ephemeral_file;
39 : use crate::tenant::mgr::TenantManager;
40 : use crate::tenant::remote_timeline_client::download::download_layer_file;
41 : use crate::tenant::remote_timeline_client::index::LayerFileMetadata;
42 : use crate::tenant::remote_timeline_client::{
43 : FAILED_DOWNLOAD_WARN_THRESHOLD, FAILED_REMOTE_OP_RETRIES, is_temp_download_file,
44 : remote_heatmap_path,
45 : };
46 : use crate::tenant::span::debug_assert_current_span_has_tenant_id;
47 : use crate::tenant::storage_layer::layer::local_layer_path;
48 : use crate::tenant::storage_layer::{LayerName, LayerVisibilityHint};
49 : use crate::tenant::tasks::{BackgroundLoopKind, warn_when_period_overrun};
50 : use crate::virtual_file::{MaybeFatalIo, VirtualFile, on_fatal_io_error};
51 :
52 : /// For each tenant, default period for how long must have passed since the last download_tenant call before
53 : /// calling it again. This default is replaced with the value of [`HeatMapTenant::upload_period_ms`] after first
54 : /// download, if the uploader populated it.
55 : const DEFAULT_DOWNLOAD_INTERVAL: Duration = Duration::from_millis(60000);
56 :
57 0 : pub(super) async fn downloader_task(
58 0 : tenant_manager: Arc<TenantManager>,
59 0 : remote_storage: GenericRemoteStorage,
60 0 : command_queue: tokio::sync::mpsc::Receiver<CommandRequest<DownloadCommand>>,
61 0 : background_jobs_can_start: Barrier,
62 0 : cancel: CancellationToken,
63 0 : root_ctx: RequestContext,
64 0 : ) {
65 0 : let concurrency = tenant_manager.get_conf().secondary_download_concurrency;
66 0 :
67 0 : let generator = SecondaryDownloader {
68 0 : tenant_manager,
69 0 : remote_storage,
70 0 : root_ctx,
71 0 : };
72 0 : let mut scheduler = Scheduler::new(generator, concurrency);
73 0 :
74 0 : scheduler
75 0 : .run(command_queue, background_jobs_can_start, cancel)
76 0 : .instrument(info_span!("secondary_download_scheduler"))
77 0 : .await
78 0 : }
79 :
80 : struct SecondaryDownloader {
81 : tenant_manager: Arc<TenantManager>,
82 : remote_storage: GenericRemoteStorage,
83 : root_ctx: RequestContext,
84 : }
85 :
86 : #[derive(Debug, Clone)]
87 : pub(super) struct OnDiskState {
88 : metadata: LayerFileMetadata,
89 : access_time: SystemTime,
90 : local_path: Utf8PathBuf,
91 : }
92 :
93 : impl OnDiskState {
94 0 : fn new(
95 0 : _conf: &'static PageServerConf,
96 0 : _tenant_shard_id: &TenantShardId,
97 0 : _imeline_id: &TimelineId,
98 0 : _ame: LayerName,
99 0 : metadata: LayerFileMetadata,
100 0 : access_time: SystemTime,
101 0 : local_path: Utf8PathBuf,
102 0 : ) -> Self {
103 0 : Self {
104 0 : metadata,
105 0 : access_time,
106 0 : local_path,
107 0 : }
108 0 : }
109 :
110 : // This is infallible, because all errors are either acceptable (ENOENT), or totally
111 : // unexpected (fatal).
112 0 : pub(super) fn remove_blocking(&self) {
113 0 : // We tolerate ENOENT, because between planning eviction and executing
114 0 : // it, the secondary downloader could have seen an updated heatmap that
115 0 : // resulted in a layer being deleted.
116 0 : // Other local I/O errors are process-fatal: these should never happen.
117 0 : std::fs::remove_file(&self.local_path)
118 0 : .or_else(fs_ext::ignore_not_found)
119 0 : .fatal_err("Deleting secondary layer")
120 0 : }
121 :
122 0 : pub(crate) fn file_size(&self) -> u64 {
123 0 : self.metadata.file_size
124 0 : }
125 : }
126 :
127 : #[derive(Debug, Clone, Default)]
128 : pub(super) struct SecondaryDetailTimeline {
129 : on_disk_layers: HashMap<LayerName, OnDiskState>,
130 :
131 : /// We remember when layers were evicted, to prevent re-downloading them.
132 : pub(super) evicted_at: HashMap<LayerName, SystemTime>,
133 : }
134 :
135 : impl SecondaryDetailTimeline {
136 0 : pub(super) fn remove_layer(
137 0 : &mut self,
138 0 : name: &LayerName,
139 0 : resident_metric: &UIntGauge,
140 0 : ) -> Option<OnDiskState> {
141 0 : let removed = self.on_disk_layers.remove(name);
142 0 : if let Some(removed) = &removed {
143 0 : resident_metric.sub(removed.file_size());
144 0 : }
145 0 : removed
146 0 : }
147 :
148 : /// `local_path`
149 0 : fn touch_layer<F>(
150 0 : &mut self,
151 0 : conf: &'static PageServerConf,
152 0 : tenant_shard_id: &TenantShardId,
153 0 : timeline_id: &TimelineId,
154 0 : touched: &HeatMapLayer,
155 0 : resident_metric: &UIntGauge,
156 0 : local_path: F,
157 0 : ) where
158 0 : F: FnOnce() -> Utf8PathBuf,
159 0 : {
160 : use std::collections::hash_map::Entry;
161 0 : match self.on_disk_layers.entry(touched.name.clone()) {
162 0 : Entry::Occupied(mut v) => {
163 0 : v.get_mut().access_time = touched.access_time;
164 0 : }
165 0 : Entry::Vacant(e) => {
166 0 : e.insert(OnDiskState::new(
167 0 : conf,
168 0 : tenant_shard_id,
169 0 : timeline_id,
170 0 : touched.name.clone(),
171 0 : touched.metadata.clone(),
172 0 : touched.access_time,
173 0 : local_path(),
174 0 : ));
175 0 : resident_metric.add(touched.metadata.file_size);
176 0 : }
177 : }
178 0 : }
179 : }
180 :
181 : // Aspects of a heatmap that we remember after downloading it
182 : #[derive(Clone, Debug)]
183 : struct DownloadSummary {
184 : etag: Etag,
185 : #[allow(unused)]
186 : mtime: SystemTime,
187 : upload_period: Duration,
188 : }
189 :
190 : /// This state is written by the secondary downloader, it is opaque
191 : /// to TenantManager
192 : #[derive(Debug)]
193 : pub(super) struct SecondaryDetail {
194 : pub(super) config: SecondaryLocationConfig,
195 :
196 : last_download: Option<DownloadSummary>,
197 : next_download: Option<Instant>,
198 : timelines: HashMap<TimelineId, SecondaryDetailTimeline>,
199 : }
200 :
201 : /// Helper for logging SystemTime
202 0 : fn strftime(t: &'_ SystemTime) -> DelayedFormat<StrftimeItems<'_>> {
203 0 : let datetime: chrono::DateTime<chrono::Utc> = (*t).into();
204 0 : datetime.format("%d/%m/%Y %T")
205 0 : }
206 :
207 : /// Information returned from download function when it detects the heatmap has changed
208 : struct HeatMapModified {
209 : etag: Etag,
210 : last_modified: SystemTime,
211 : bytes: Vec<u8>,
212 : }
213 :
214 : enum HeatMapDownload {
215 : // The heatmap's etag has changed: return the new etag, mtime and the body bytes
216 : Modified(HeatMapModified),
217 : // The heatmap's etag is unchanged
218 : Unmodified,
219 : }
220 :
221 : impl SecondaryDetail {
222 0 : pub(super) fn new(config: SecondaryLocationConfig) -> Self {
223 0 : Self {
224 0 : config,
225 0 : last_download: None,
226 0 : next_download: None,
227 0 : timelines: HashMap::new(),
228 0 : }
229 0 : }
230 :
231 : #[cfg(feature = "testing")]
232 0 : pub(crate) fn total_resident_size(&self) -> u64 {
233 0 : self.timelines
234 0 : .values()
235 0 : .map(|tl| {
236 0 : tl.on_disk_layers
237 0 : .values()
238 0 : .map(|v| v.metadata.file_size)
239 0 : .sum::<u64>()
240 0 : })
241 0 : .sum::<u64>()
242 0 : }
243 :
244 0 : pub(super) fn evict_layer(
245 0 : &mut self,
246 0 : name: LayerName,
247 0 : timeline_id: &TimelineId,
248 0 : now: SystemTime,
249 0 : resident_metric: &UIntGauge,
250 0 : ) -> Option<OnDiskState> {
251 0 : let timeline = self.timelines.get_mut(timeline_id)?;
252 0 : let removed = timeline.remove_layer(&name, resident_metric);
253 0 : if removed.is_some() {
254 0 : timeline.evicted_at.insert(name, now);
255 0 : }
256 0 : removed
257 0 : }
258 :
259 0 : pub(super) fn remove_timeline(
260 0 : &mut self,
261 0 : timeline_id: &TimelineId,
262 0 : resident_metric: &UIntGauge,
263 0 : ) {
264 0 : let removed = self.timelines.remove(timeline_id);
265 0 : if let Some(removed) = removed {
266 0 : resident_metric.sub(
267 0 : removed
268 0 : .on_disk_layers
269 0 : .values()
270 0 : .map(|l| l.metadata.file_size)
271 0 : .sum(),
272 0 : );
273 0 : }
274 0 : }
275 :
276 : /// Additionally returns the total number of layers, used for more stable relative access time
277 : /// based eviction.
278 0 : pub(super) fn get_layers_for_eviction(
279 0 : &self,
280 0 : parent: &Arc<SecondaryTenant>,
281 0 : ) -> (DiskUsageEvictionInfo, usize) {
282 0 : let mut result = DiskUsageEvictionInfo::default();
283 0 : let mut total_layers = 0;
284 :
285 0 : for (timeline_id, timeline_detail) in &self.timelines {
286 0 : result
287 0 : .resident_layers
288 0 : .extend(timeline_detail.on_disk_layers.iter().map(|(name, ods)| {
289 0 : EvictionCandidate {
290 0 : layer: EvictionLayer::Secondary(EvictionSecondaryLayer {
291 0 : secondary_tenant: parent.clone(),
292 0 : timeline_id: *timeline_id,
293 0 : name: name.clone(),
294 0 : metadata: ods.metadata.clone(),
295 0 : }),
296 0 : last_activity_ts: ods.access_time,
297 0 : relative_last_activity: finite_f32::FiniteF32::ZERO,
298 0 : // Secondary location layers are presumed visible, because Covered layers
299 0 : // are excluded from the heatmap
300 0 : visibility: LayerVisibilityHint::Visible,
301 0 : }
302 0 : }));
303 0 :
304 0 : // total might be missing currently downloading layers, but as a lower than actual
305 0 : // value it is good enough approximation.
306 0 : total_layers += timeline_detail.on_disk_layers.len() + timeline_detail.evicted_at.len();
307 0 : }
308 0 : result.max_layer_size = result
309 0 : .resident_layers
310 0 : .iter()
311 0 : .map(|l| l.layer.get_file_size())
312 0 : .max();
313 0 :
314 0 : tracing::debug!(
315 0 : "eviction: secondary tenant {} found {} timelines, {} layers",
316 0 : parent.get_tenant_shard_id(),
317 0 : self.timelines.len(),
318 0 : result.resident_layers.len()
319 : );
320 :
321 0 : (result, total_layers)
322 0 : }
323 : }
324 :
325 : struct PendingDownload {
326 : secondary_state: Arc<SecondaryTenant>,
327 : last_download: Option<DownloadSummary>,
328 : target_time: Option<Instant>,
329 : }
330 :
331 : impl scheduler::PendingJob for PendingDownload {
332 0 : fn get_tenant_shard_id(&self) -> &TenantShardId {
333 0 : self.secondary_state.get_tenant_shard_id()
334 0 : }
335 : }
336 :
337 : struct RunningDownload {
338 : barrier: Barrier,
339 : }
340 :
341 : impl scheduler::RunningJob for RunningDownload {
342 0 : fn get_barrier(&self) -> Barrier {
343 0 : self.barrier.clone()
344 0 : }
345 : }
346 :
347 : struct CompleteDownload {
348 : secondary_state: Arc<SecondaryTenant>,
349 : completed_at: Instant,
350 : result: Result<(), UpdateError>,
351 : }
352 :
353 : impl scheduler::Completion for CompleteDownload {
354 0 : fn get_tenant_shard_id(&self) -> &TenantShardId {
355 0 : self.secondary_state.get_tenant_shard_id()
356 0 : }
357 : }
358 :
359 : type Scheduler = TenantBackgroundJobs<
360 : SecondaryDownloader,
361 : PendingDownload,
362 : RunningDownload,
363 : CompleteDownload,
364 : DownloadCommand,
365 : >;
366 :
367 : impl JobGenerator<PendingDownload, RunningDownload, CompleteDownload, DownloadCommand>
368 : for SecondaryDownloader
369 : {
370 : #[instrument(skip_all, fields(tenant_id=%completion.get_tenant_shard_id().tenant_id, shard_id=%completion.get_tenant_shard_id().shard_slug()))]
371 : fn on_completion(&mut self, completion: CompleteDownload) {
372 : let CompleteDownload {
373 : secondary_state,
374 : completed_at: _completed_at,
375 : result,
376 : } = completion;
377 :
378 : tracing::debug!("Secondary tenant download completed");
379 :
380 : let mut detail = secondary_state.detail.lock().unwrap();
381 :
382 : match result {
383 : Err(UpdateError::Restart) => {
384 : // Start downloading again as soon as we can. This will involve waiting for the scheduler's
385 : // scheduling interval. This slightly reduces the peak download speed of tenants that hit their
386 : // deadline and keep restarting, but that also helps give other tenants a chance to execute rather
387 : // that letting one big tenant dominate for a long time.
388 : detail.next_download = Some(Instant::now());
389 : }
390 : _ => {
391 : let period = detail
392 : .last_download
393 : .as_ref()
394 0 : .map(|d| d.upload_period)
395 : .unwrap_or(DEFAULT_DOWNLOAD_INTERVAL);
396 :
397 : // We advance next_download irrespective of errors: we don't want error cases to result in
398 : // expensive busy-polling.
399 : detail.next_download = Some(Instant::now() + period_jitter(period, 5));
400 : }
401 : }
402 : }
403 :
404 0 : async fn schedule(&mut self) -> SchedulingResult<PendingDownload> {
405 0 : let mut result = SchedulingResult {
406 0 : jobs: Vec::new(),
407 0 : want_interval: None,
408 0 : };
409 0 :
410 0 : // Step 1: identify some tenants that we may work on
411 0 : let mut tenants: Vec<Arc<SecondaryTenant>> = Vec::new();
412 0 : self.tenant_manager
413 0 : .foreach_secondary_tenants(|_id, secondary_state| {
414 0 : tenants.push(secondary_state.clone());
415 0 : });
416 0 :
417 0 : // Step 2: filter out tenants which are not yet elegible to run
418 0 : let now = Instant::now();
419 0 : result.jobs = tenants
420 0 : .into_iter()
421 0 : .filter_map(|secondary_tenant| {
422 0 : let (last_download, next_download) = {
423 0 : let mut detail = secondary_tenant.detail.lock().unwrap();
424 0 :
425 0 : if !detail.config.warm {
426 : // Downloads are disabled for this tenant
427 0 : detail.next_download = None;
428 0 : return None;
429 0 : }
430 0 :
431 0 : if detail.next_download.is_none() {
432 0 : // Initialize randomly in the range from 0 to our interval: this uniformly spreads the start times. Subsequent
433 0 : // rounds will use a smaller jitter to avoid accidentally synchronizing later.
434 0 : detail.next_download = Some(now.checked_add(period_warmup(DEFAULT_DOWNLOAD_INTERVAL)).expect(
435 0 : "Using our constant, which is known to be small compared with clock range",
436 0 : ));
437 0 : }
438 0 : (detail.last_download.clone(), detail.next_download.unwrap())
439 0 : };
440 0 :
441 0 : if now > next_download {
442 0 : Some(PendingDownload {
443 0 : secondary_state: secondary_tenant,
444 0 : last_download,
445 0 : target_time: Some(next_download),
446 0 : })
447 : } else {
448 0 : None
449 : }
450 0 : })
451 0 : .collect();
452 0 :
453 0 : // Step 3: sort by target execution time to run most urgent first.
454 0 : result.jobs.sort_by_key(|j| j.target_time);
455 0 :
456 0 : result
457 0 : }
458 :
459 0 : fn on_command(
460 0 : &mut self,
461 0 : command: DownloadCommand,
462 0 : ) -> Result<PendingDownload, SecondaryTenantError> {
463 0 : let tenant_shard_id = command.get_tenant_shard_id();
464 :
465 0 : let tenant = self
466 0 : .tenant_manager
467 0 : .get_secondary_tenant_shard(*tenant_shard_id)
468 0 : .ok_or(GetTenantError::ShardNotFound(*tenant_shard_id))?;
469 :
470 0 : Ok(PendingDownload {
471 0 : target_time: None,
472 0 : last_download: None,
473 0 : secondary_state: tenant,
474 0 : })
475 0 : }
476 :
477 0 : fn spawn(
478 0 : &mut self,
479 0 : job: PendingDownload,
480 0 : ) -> (
481 0 : RunningDownload,
482 0 : Pin<Box<dyn Future<Output = CompleteDownload> + Send>>,
483 0 : ) {
484 0 : let PendingDownload {
485 0 : secondary_state,
486 0 : last_download,
487 0 : target_time,
488 0 : } = job;
489 0 :
490 0 : let (completion, barrier) = utils::completion::channel();
491 0 : let remote_storage = self.remote_storage.clone();
492 0 : let conf = self.tenant_manager.get_conf();
493 0 : let tenant_shard_id = *secondary_state.get_tenant_shard_id();
494 0 : let download_ctx = self
495 0 : .root_ctx
496 0 : .attached_child()
497 0 : .with_scope_secondary_tenant(&tenant_shard_id);
498 0 : (RunningDownload { barrier }, Box::pin(async move {
499 0 : let _completion = completion;
500 :
501 0 : let result = TenantDownloader::new(conf, &remote_storage, &secondary_state)
502 0 : .download(&download_ctx)
503 0 : .await;
504 0 : match &result
505 : {
506 : Err(UpdateError::NoData) => {
507 0 : tracing::info!("No heatmap found for tenant. This is fine if it is new.");
508 : },
509 : Err(UpdateError::NoSpace) => {
510 0 : tracing::warn!("Insufficient space while downloading. Will retry later.");
511 : }
512 : Err(UpdateError::Cancelled) => {
513 0 : tracing::info!("Shut down while downloading");
514 : },
515 0 : Err(UpdateError::Deserialize(e)) => {
516 0 : tracing::error!("Corrupt content while downloading tenant: {e}");
517 : },
518 0 : Err(e @ (UpdateError::DownloadError(_) | UpdateError::Other(_))) => {
519 0 : tracing::error!("Error while downloading tenant: {e}");
520 : },
521 : Err(UpdateError::Restart) => {
522 0 : tracing::info!("Download reached deadline & will restart to update heatmap")
523 : }
524 0 : Ok(()) => {}
525 : };
526 :
527 : // Irrespective of the result, we will reschedule ourselves to run after our usual period.
528 :
529 : // If the job had a target execution time, we may check our final execution
530 : // time against that for observability purposes.
531 0 : if let (Some(target_time), Some(last_download)) = (target_time, last_download) {
532 0 : // Elapsed time includes any scheduling lag as well as the execution of the job
533 0 : let elapsed = Instant::now().duration_since(target_time);
534 0 :
535 0 : warn_when_period_overrun(
536 0 : elapsed,
537 0 : last_download.upload_period,
538 0 : BackgroundLoopKind::SecondaryDownload,
539 0 : );
540 0 : }
541 :
542 0 : CompleteDownload {
543 0 : secondary_state,
544 0 : completed_at: Instant::now(),
545 0 : result
546 0 : }
547 0 : }.instrument(info_span!(parent: None, "secondary_download", tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()))))
548 0 : }
549 : }
550 :
551 : enum LayerAction {
552 : Download,
553 : NoAction,
554 : Skip,
555 : Touch,
556 : }
557 :
558 : /// This type is a convenience to group together the various functions involved in
559 : /// freshening a secondary tenant.
560 : struct TenantDownloader<'a> {
561 : conf: &'static PageServerConf,
562 : remote_storage: &'a GenericRemoteStorage,
563 : secondary_state: &'a SecondaryTenant,
564 : }
565 :
566 : /// Errors that may be encountered while updating a tenant
567 : #[derive(thiserror::Error, Debug)]
568 : enum UpdateError {
569 : /// This is not a true failure, but it's how a download indicates that it would like to be restarted by
570 : /// the scheduler, to pick up the latest heatmap
571 : #[error("Reached deadline, restarting downloads")]
572 : Restart,
573 :
574 : #[error("No remote data found")]
575 : NoData,
576 : #[error("Insufficient local storage space")]
577 : NoSpace,
578 : #[error("Failed to download")]
579 : DownloadError(DownloadError),
580 : #[error(transparent)]
581 : Deserialize(#[from] serde_json::Error),
582 : #[error("Cancelled")]
583 : Cancelled,
584 : #[error(transparent)]
585 : Other(#[from] anyhow::Error),
586 : }
587 :
588 : impl From<DownloadError> for UpdateError {
589 0 : fn from(value: DownloadError) -> Self {
590 0 : match &value {
591 0 : DownloadError::Cancelled => Self::Cancelled,
592 0 : DownloadError::NotFound => Self::NoData,
593 0 : _ => Self::DownloadError(value),
594 : }
595 0 : }
596 : }
597 :
598 : impl From<std::io::Error> for UpdateError {
599 0 : fn from(value: std::io::Error) -> Self {
600 0 : if let Some(nix::errno::Errno::ENOSPC) = value.raw_os_error().map(nix::errno::from_i32) {
601 0 : UpdateError::NoSpace
602 0 : } else if value
603 0 : .get_ref()
604 0 : .and_then(|x| x.downcast_ref::<DownloadError>())
605 0 : .is_some()
606 : {
607 0 : UpdateError::from(DownloadError::from(value))
608 : } else {
609 : // An I/O error from e.g. tokio::io::copy_buf is most likely a remote storage issue
610 0 : UpdateError::Other(anyhow::anyhow!(value))
611 : }
612 0 : }
613 : }
614 :
615 : impl<'a> TenantDownloader<'a> {
616 0 : fn new(
617 0 : conf: &'static PageServerConf,
618 0 : remote_storage: &'a GenericRemoteStorage,
619 0 : secondary_state: &'a SecondaryTenant,
620 0 : ) -> Self {
621 0 : Self {
622 0 : conf,
623 0 : remote_storage,
624 0 : secondary_state,
625 0 : }
626 0 : }
627 :
628 0 : async fn download(&self, ctx: &RequestContext) -> Result<(), UpdateError> {
629 0 : debug_assert_current_span_has_tenant_id();
630 :
631 : // For the duration of a download, we must hold the SecondaryTenant::gate, to ensure
632 : // cover our access to local storage.
633 0 : let Ok(_guard) = self.secondary_state.gate.enter() else {
634 : // Shutting down
635 0 : return Err(UpdateError::Cancelled);
636 : };
637 :
638 0 : let tenant_shard_id = self.secondary_state.get_tenant_shard_id();
639 0 :
640 0 : // We will use the etag from last successful download to make the download conditional on changes
641 0 : let last_download = self
642 0 : .secondary_state
643 0 : .detail
644 0 : .lock()
645 0 : .unwrap()
646 0 : .last_download
647 0 : .clone();
648 :
649 : // Download the tenant's heatmap
650 : let HeatMapModified {
651 0 : last_modified: heatmap_mtime,
652 0 : etag: heatmap_etag,
653 0 : bytes: heatmap_bytes,
654 0 : } = match tokio::select!(
655 0 : bytes = self.download_heatmap(last_download.as_ref().map(|d| &d.etag)) => {bytes?},
656 0 : _ = self.secondary_state.cancel.cancelled() => return Ok(())
657 : ) {
658 : HeatMapDownload::Unmodified => {
659 0 : tracing::info!("Heatmap unchanged since last successful download");
660 0 : return Ok(());
661 : }
662 0 : HeatMapDownload::Modified(m) => m,
663 0 : };
664 0 :
665 0 : // Heatmap storage location
666 0 : let heatmap_path = self.conf.tenant_heatmap_path(tenant_shard_id);
667 :
668 0 : let last_heatmap = if last_download.is_none() {
669 0 : match load_heatmap(&heatmap_path, ctx).await {
670 0 : Ok(htm) => htm,
671 0 : Err(e) => {
672 0 : tracing::warn!("Couldn't load heatmap from {heatmap_path}: {e:?}");
673 0 : None
674 : }
675 : }
676 : } else {
677 0 : None
678 : };
679 :
680 0 : let last_heatmap_timelines = last_heatmap.as_ref().map(|htm| {
681 0 : htm.timelines
682 0 : .iter()
683 0 : .map(|tl| (tl.timeline_id, tl))
684 0 : .collect::<HashMap<_, _>>()
685 0 : });
686 :
687 0 : let heatmap = serde_json::from_slice::<HeatMapTenant>(&heatmap_bytes)?;
688 :
689 0 : let temp_path = path_with_suffix_extension(&heatmap_path, TEMP_FILE_SUFFIX);
690 0 : let context_msg = format!("write tenant {tenant_shard_id} heatmap to {heatmap_path}");
691 0 : let heatmap_path_bg = heatmap_path.clone();
692 0 : VirtualFile::crashsafe_overwrite(heatmap_path_bg, temp_path, heatmap_bytes)
693 0 : .await
694 0 : .maybe_fatal_err(&context_msg)?;
695 :
696 0 : tracing::debug!(
697 0 : "Wrote local heatmap to {}, with {} timelines",
698 0 : heatmap_path,
699 0 : heatmap.timelines.len()
700 : );
701 :
702 : // Get or initialize the local disk state for the timelines we will update
703 0 : let mut timeline_states = HashMap::new();
704 0 : for timeline in &heatmap.timelines {
705 0 : let timeline_state = self
706 0 : .secondary_state
707 0 : .detail
708 0 : .lock()
709 0 : .unwrap()
710 0 : .timelines
711 0 : .get(&timeline.timeline_id)
712 0 : .cloned();
713 :
714 0 : let timeline_state = match timeline_state {
715 0 : Some(t) => t,
716 : None => {
717 0 : let last_heatmap =
718 0 : last_heatmap_timelines
719 0 : .as_ref()
720 0 : .and_then(|last_heatmap_timelines| {
721 0 : last_heatmap_timelines.get(&timeline.timeline_id).copied()
722 0 : });
723 : // We have no existing state: need to scan local disk for layers first.
724 0 : let timeline_state = init_timeline_state(
725 0 : self.conf,
726 0 : tenant_shard_id,
727 0 : last_heatmap,
728 0 : timeline,
729 0 : &self.secondary_state.resident_size_metric,
730 0 : )
731 0 : .await;
732 :
733 : // Re-acquire detail lock now that we're done with async load from local FS
734 0 : self.secondary_state
735 0 : .detail
736 0 : .lock()
737 0 : .unwrap()
738 0 : .timelines
739 0 : .insert(timeline.timeline_id, timeline_state.clone());
740 0 : timeline_state
741 : }
742 : };
743 :
744 0 : timeline_states.insert(timeline.timeline_id, timeline_state);
745 : }
746 :
747 : // Clean up any local layers that aren't in the heatmap. We do this first for all timelines, on the general
748 : // principle that deletions should be done before writes wherever possible, and so that we can use this
749 : // phase to initialize our SecondaryProgress.
750 : {
751 0 : *self.secondary_state.progress.lock().unwrap() =
752 0 : self.prepare_timelines(&heatmap, heatmap_mtime).await?;
753 : }
754 :
755 : // Calculate a deadline for downloads: if downloading takes longer than this, it is useful to drop out and start again,
756 : // so that we are always using reasonably a fresh heatmap. Otherwise, if we had really huge content to download, we might
757 : // spend 10s of minutes downloading layers we don't need.
758 : // (see https://github.com/neondatabase/neon/issues/8182)
759 0 : let deadline = {
760 0 : let period = self
761 0 : .secondary_state
762 0 : .detail
763 0 : .lock()
764 0 : .unwrap()
765 0 : .last_download
766 0 : .as_ref()
767 0 : .map(|d| d.upload_period)
768 0 : .unwrap_or(DEFAULT_DOWNLOAD_INTERVAL);
769 0 :
770 0 : // Use double the period: we are not promising to complete within the period, this is just a heuristic
771 0 : // to keep using a "reasonably fresh" heatmap.
772 0 : Instant::now() + period * 2
773 : };
774 :
775 : // Download the layers in the heatmap
776 0 : for timeline in heatmap.timelines {
777 0 : let ctx = &ctx.with_scope_secondary_timeline(tenant_shard_id, &timeline.timeline_id);
778 0 : let timeline_state = timeline_states
779 0 : .remove(&timeline.timeline_id)
780 0 : .expect("Just populated above");
781 0 :
782 0 : if self.secondary_state.cancel.is_cancelled() {
783 0 : tracing::debug!(
784 0 : "Cancelled before downloading timeline {}",
785 : timeline.timeline_id
786 : );
787 0 : return Ok(());
788 0 : }
789 0 :
790 0 : let timeline_id = timeline.timeline_id;
791 0 : self.download_timeline(timeline, timeline_state, deadline, ctx)
792 0 : .instrument(tracing::info_span!(
793 : "secondary_download_timeline",
794 : tenant_id=%tenant_shard_id.tenant_id,
795 0 : shard_id=%tenant_shard_id.shard_slug(),
796 : %timeline_id
797 : ))
798 0 : .await?;
799 : }
800 :
801 : // Metrics consistency check in testing builds
802 0 : self.secondary_state.validate_metrics();
803 0 : // Only update last_etag after a full successful download: this way will not skip
804 0 : // the next download, even if the heatmap's actual etag is unchanged.
805 0 : self.secondary_state.detail.lock().unwrap().last_download = Some(DownloadSummary {
806 0 : etag: heatmap_etag,
807 0 : mtime: heatmap_mtime,
808 0 : upload_period: heatmap
809 0 : .upload_period_ms
810 0 : .map(|ms| Duration::from_millis(ms as u64))
811 0 : .unwrap_or(DEFAULT_DOWNLOAD_INTERVAL),
812 0 : });
813 0 :
814 0 : // Robustness: we should have updated progress properly, but in case we didn't, make sure
815 0 : // we don't leave the tenant in a state where we claim to have successfully downloaded
816 0 : // everything, but our progress is incomplete. The invariant here should be that if
817 0 : // we have set `last_download` to this heatmap's etag, then the next time we see that
818 0 : // etag we can safely do no work (i.e. we must be complete).
819 0 : let mut progress = self.secondary_state.progress.lock().unwrap();
820 0 : debug_assert!(progress.layers_downloaded == progress.layers_total);
821 0 : debug_assert!(progress.bytes_downloaded == progress.bytes_total);
822 0 : if progress.layers_downloaded != progress.layers_total
823 0 : || progress.bytes_downloaded != progress.bytes_total
824 : {
825 0 : tracing::warn!("Correcting drift in progress stats ({progress:?})");
826 0 : progress.layers_downloaded = progress.layers_total;
827 0 : progress.bytes_downloaded = progress.bytes_total;
828 0 : }
829 :
830 0 : Ok(())
831 0 : }
832 :
833 : /// Do any fast local cleanup that comes before the much slower process of downloading
834 : /// layers from remote storage. In the process, initialize the SecondaryProgress object
835 : /// that will later be updated incrementally as we download layers.
836 0 : async fn prepare_timelines(
837 0 : &self,
838 0 : heatmap: &HeatMapTenant,
839 0 : heatmap_mtime: SystemTime,
840 0 : ) -> Result<SecondaryProgress, UpdateError> {
841 0 : let heatmap_stats = heatmap.get_stats();
842 0 : // We will construct a progress object, and then populate its initial "downloaded" numbers
843 0 : // while iterating through local layer state in [`Self::prepare_timelines`]
844 0 : let mut progress = SecondaryProgress {
845 0 : layers_total: heatmap_stats.layers,
846 0 : bytes_total: heatmap_stats.bytes,
847 0 : heatmap_mtime: Some(serde_system_time::SystemTime(heatmap_mtime)),
848 0 : layers_downloaded: 0,
849 0 : bytes_downloaded: 0,
850 0 : };
851 0 :
852 0 : // Also expose heatmap bytes_total as a metric
853 0 : self.secondary_state
854 0 : .heatmap_total_size_metric
855 0 : .set(heatmap_stats.bytes);
856 0 :
857 0 : // Accumulate list of things to delete while holding the detail lock, for execution after dropping the lock
858 0 : let mut delete_layers = Vec::new();
859 0 : let mut delete_timelines = Vec::new();
860 0 : {
861 0 : let mut detail = self.secondary_state.detail.lock().unwrap();
862 0 : for (timeline_id, timeline_state) in &mut detail.timelines {
863 0 : let Some(heatmap_timeline_index) = heatmap
864 0 : .timelines
865 0 : .iter()
866 0 : .position(|t| t.timeline_id == *timeline_id)
867 : else {
868 : // This timeline is no longer referenced in the heatmap: delete it locally
869 0 : delete_timelines.push(*timeline_id);
870 0 : continue;
871 : };
872 :
873 0 : let heatmap_timeline = heatmap.timelines.get(heatmap_timeline_index).unwrap();
874 0 :
875 0 : let layers_in_heatmap = heatmap_timeline
876 0 : .hot_layers()
877 0 : .map(|l| (&l.name, l.metadata.generation))
878 0 : .collect::<HashSet<_>>();
879 0 : let layers_on_disk = timeline_state
880 0 : .on_disk_layers
881 0 : .iter()
882 0 : .map(|l| (l.0, l.1.metadata.generation))
883 0 : .collect::<HashSet<_>>();
884 0 :
885 0 : let mut layer_count = layers_on_disk.len();
886 0 : let mut layer_byte_count: u64 = timeline_state
887 0 : .on_disk_layers
888 0 : .values()
889 0 : .map(|l| l.metadata.file_size)
890 0 : .sum();
891 :
892 : // Remove on-disk layers that are no longer present in heatmap
893 0 : for (layer_file_name, generation) in layers_on_disk.difference(&layers_in_heatmap) {
894 0 : layer_count -= 1;
895 0 : layer_byte_count -= timeline_state
896 0 : .on_disk_layers
897 0 : .get(layer_file_name)
898 0 : .unwrap()
899 0 : .metadata
900 0 : .file_size;
901 0 :
902 0 : let local_path = local_layer_path(
903 0 : self.conf,
904 0 : self.secondary_state.get_tenant_shard_id(),
905 0 : timeline_id,
906 0 : layer_file_name,
907 0 : generation,
908 0 : );
909 0 :
910 0 : delete_layers.push((*timeline_id, (*layer_file_name).clone(), local_path));
911 0 : }
912 :
913 0 : progress.bytes_downloaded += layer_byte_count;
914 0 : progress.layers_downloaded += layer_count;
915 : }
916 :
917 0 : for delete_timeline in &delete_timelines {
918 0 : // We haven't removed from disk yet, but optimistically remove from in-memory state: if removal
919 0 : // from disk fails that will be a fatal error.
920 0 : detail.remove_timeline(delete_timeline, &self.secondary_state.resident_size_metric);
921 0 : }
922 : }
923 :
924 : // Execute accumulated deletions
925 0 : for (timeline_id, layer_name, local_path) in delete_layers {
926 0 : tracing::info!(timeline_id=%timeline_id, "Removing secondary local layer {layer_name} because it's absent in heatmap",);
927 :
928 0 : tokio::fs::remove_file(&local_path)
929 0 : .await
930 0 : .or_else(fs_ext::ignore_not_found)
931 0 : .maybe_fatal_err("Removing secondary layer")?;
932 :
933 : // Update in-memory housekeeping to reflect the absence of the deleted layer
934 0 : let mut detail = self.secondary_state.detail.lock().unwrap();
935 0 : let Some(timeline_state) = detail.timelines.get_mut(&timeline_id) else {
936 0 : continue;
937 : };
938 0 : timeline_state.remove_layer(&layer_name, &self.secondary_state.resident_size_metric);
939 : }
940 :
941 0 : for timeline_id in delete_timelines {
942 0 : let timeline_path = self
943 0 : .conf
944 0 : .timeline_path(self.secondary_state.get_tenant_shard_id(), &timeline_id);
945 0 : tracing::info!(timeline_id=%timeline_id,
946 0 : "Timeline no longer in heatmap, removing from secondary location"
947 : );
948 0 : tokio::fs::remove_dir_all(&timeline_path)
949 0 : .await
950 0 : .or_else(fs_ext::ignore_not_found)
951 0 : .maybe_fatal_err("Removing secondary timeline")?;
952 : }
953 :
954 0 : Ok(progress)
955 0 : }
956 :
957 : /// Returns downloaded bytes if the etag differs from `prev_etag`, or None if the object
958 : /// still matches `prev_etag`.
959 0 : async fn download_heatmap(
960 0 : &self,
961 0 : prev_etag: Option<&Etag>,
962 0 : ) -> Result<HeatMapDownload, UpdateError> {
963 0 : debug_assert_current_span_has_tenant_id();
964 0 : let tenant_shard_id = self.secondary_state.get_tenant_shard_id();
965 0 : tracing::debug!("Downloading heatmap for secondary tenant",);
966 :
967 0 : let heatmap_path = remote_heatmap_path(tenant_shard_id);
968 0 : let cancel = &self.secondary_state.cancel;
969 0 : let opts = DownloadOpts {
970 0 : etag: prev_etag.cloned(),
971 0 : kind: DownloadKind::Small,
972 0 : ..Default::default()
973 0 : };
974 0 :
975 0 : backoff::retry(
976 0 : || async {
977 0 : let download = match self
978 0 : .remote_storage
979 0 : .download(&heatmap_path, &opts, cancel)
980 0 : .await
981 : {
982 0 : Ok(download) => download,
983 0 : Err(DownloadError::Unmodified) => return Ok(HeatMapDownload::Unmodified),
984 0 : Err(err) => return Err(err.into()),
985 : };
986 :
987 0 : let mut heatmap_bytes = Vec::new();
988 0 : let mut body = tokio_util::io::StreamReader::new(download.download_stream);
989 0 : let _size = tokio::io::copy_buf(&mut body, &mut heatmap_bytes).await?;
990 0 : Ok(HeatMapDownload::Modified(HeatMapModified {
991 0 : etag: download.etag,
992 0 : last_modified: download.last_modified,
993 0 : bytes: heatmap_bytes,
994 0 : }))
995 0 : },
996 0 : |e| matches!(e, UpdateError::NoData | UpdateError::Cancelled),
997 0 : FAILED_DOWNLOAD_WARN_THRESHOLD,
998 0 : FAILED_REMOTE_OP_RETRIES,
999 0 : "download heatmap",
1000 0 : cancel,
1001 0 : )
1002 0 : .await
1003 0 : .ok_or_else(|| UpdateError::Cancelled)
1004 0 : .and_then(|x| x)
1005 0 : .inspect(|_| SECONDARY_MODE.download_heatmap.inc())
1006 0 : }
1007 :
1008 : /// Download heatmap layers that are not present on local disk, or update their
1009 : /// access time if they are already present.
1010 0 : async fn download_timeline_layers(
1011 0 : &self,
1012 0 : tenant_shard_id: &TenantShardId,
1013 0 : timeline: HeatMapTimeline,
1014 0 : timeline_state: SecondaryDetailTimeline,
1015 0 : deadline: Instant,
1016 0 : ctx: &RequestContext,
1017 0 : ) -> (Result<(), UpdateError>, Vec<HeatMapLayer>) {
1018 0 : // Accumulate updates to the state
1019 0 : let mut touched = Vec::new();
1020 0 :
1021 0 : let timeline_id = timeline.timeline_id;
1022 0 : for layer in timeline.into_hot_layers() {
1023 0 : if self.secondary_state.cancel.is_cancelled() {
1024 0 : tracing::debug!("Cancelled -- dropping out of layer loop");
1025 0 : return (Err(UpdateError::Cancelled), touched);
1026 0 : }
1027 0 :
1028 0 : if Instant::now() > deadline {
1029 : // We've been running downloads for a while, restart to download latest heatmap.
1030 0 : return (Err(UpdateError::Restart), touched);
1031 0 : }
1032 0 :
1033 0 : match self.layer_action(&timeline_state, &layer).await {
1034 0 : LayerAction::Download => (),
1035 0 : LayerAction::NoAction => continue,
1036 : LayerAction::Skip => {
1037 0 : self.skip_layer(layer);
1038 0 : continue;
1039 : }
1040 : LayerAction::Touch => {
1041 0 : touched.push(layer);
1042 0 : continue;
1043 : }
1044 : }
1045 :
1046 0 : match self
1047 0 : .download_layer(tenant_shard_id, &timeline_id, layer, ctx)
1048 0 : .await
1049 : {
1050 0 : Ok(Some(layer)) => touched.push(layer),
1051 0 : Ok(None) => {
1052 0 : // Not an error but we didn't download it: remote layer is missing. Don't add it to the list of
1053 0 : // things to consider touched.
1054 0 : }
1055 0 : Err(e) => {
1056 0 : return (Err(e), touched);
1057 : }
1058 : }
1059 : }
1060 :
1061 0 : (Ok(()), touched)
1062 0 : }
1063 :
1064 0 : async fn layer_action(
1065 0 : &self,
1066 0 : timeline_state: &SecondaryDetailTimeline,
1067 0 : layer: &HeatMapLayer,
1068 0 : ) -> LayerAction {
1069 : // Existing on-disk layers: just update their access time.
1070 0 : if let Some(on_disk) = timeline_state.on_disk_layers.get(&layer.name) {
1071 0 : tracing::debug!("Layer {} is already on disk", layer.name);
1072 :
1073 0 : if cfg!(debug_assertions) {
1074 : // Debug for https://github.com/neondatabase/neon/issues/6966: check that the files we think
1075 : // are already present on disk are really there.
1076 0 : match tokio::fs::metadata(&on_disk.local_path).await {
1077 0 : Ok(meta) => {
1078 0 : tracing::debug!(
1079 0 : "Layer {} present at {}, size {}",
1080 0 : layer.name,
1081 0 : on_disk.local_path,
1082 0 : meta.len(),
1083 : );
1084 : }
1085 0 : Err(e) => {
1086 0 : tracing::warn!(
1087 0 : "Layer {} not found at {} ({})",
1088 : layer.name,
1089 : on_disk.local_path,
1090 : e
1091 : );
1092 0 : debug_assert!(false);
1093 : }
1094 : }
1095 0 : }
1096 :
1097 0 : if on_disk.metadata.generation_file_size() != layer.metadata.generation_file_size() {
1098 0 : tracing::info!(
1099 0 : "Re-downloading layer {} with changed size or generation: {:?}->{:?}",
1100 0 : layer.name,
1101 0 : on_disk.metadata.generation_file_size(),
1102 0 : layer.metadata.generation_file_size()
1103 : );
1104 0 : return LayerAction::Download;
1105 0 : }
1106 0 : if on_disk.metadata != layer.metadata || on_disk.access_time != layer.access_time {
1107 : // We already have this layer on disk. Update its access time.
1108 0 : tracing::debug!(
1109 0 : "Access time updated for layer {}: {} -> {}",
1110 0 : layer.name,
1111 0 : strftime(&on_disk.access_time),
1112 0 : strftime(&layer.access_time)
1113 : );
1114 0 : return LayerAction::Touch;
1115 0 : }
1116 0 : return LayerAction::NoAction;
1117 : } else {
1118 0 : tracing::debug!("Layer {} not present on disk yet", layer.name);
1119 : }
1120 :
1121 : // Eviction: if we evicted a layer, then do not re-download it unless it was accessed more
1122 : // recently than it was evicted.
1123 0 : if let Some(evicted_at) = timeline_state.evicted_at.get(&layer.name) {
1124 0 : if &layer.access_time > evicted_at {
1125 0 : tracing::info!(
1126 0 : "Re-downloading evicted layer {}, accessed at {}, evicted at {}",
1127 0 : layer.name,
1128 0 : strftime(&layer.access_time),
1129 0 : strftime(evicted_at)
1130 : );
1131 : } else {
1132 0 : tracing::trace!(
1133 0 : "Not re-downloading evicted layer {}, accessed at {}, evicted at {}",
1134 0 : layer.name,
1135 0 : strftime(&layer.access_time),
1136 0 : strftime(evicted_at)
1137 : );
1138 0 : return LayerAction::Skip;
1139 : }
1140 0 : }
1141 0 : LayerAction::Download
1142 0 : }
1143 :
1144 0 : async fn download_timeline(
1145 0 : &self,
1146 0 : timeline: HeatMapTimeline,
1147 0 : timeline_state: SecondaryDetailTimeline,
1148 0 : deadline: Instant,
1149 0 : ctx: &RequestContext,
1150 0 : ) -> Result<(), UpdateError> {
1151 0 : debug_assert_current_span_has_tenant_and_timeline_id();
1152 0 : let tenant_shard_id = self.secondary_state.get_tenant_shard_id();
1153 0 : let timeline_id = timeline.timeline_id;
1154 0 :
1155 0 : tracing::debug!(timeline_id=%timeline_id, "Downloading layers, {} in heatmap", timeline.hot_layers().count());
1156 :
1157 0 : let (result, touched) = self
1158 0 : .download_timeline_layers(tenant_shard_id, timeline, timeline_state, deadline, ctx)
1159 0 : .await;
1160 :
1161 : // Write updates to state to record layers we just downloaded or touched, irrespective of whether the overall result was successful
1162 : {
1163 0 : let mut detail = self.secondary_state.detail.lock().unwrap();
1164 0 : let timeline_detail = detail.timelines.entry(timeline_id).or_default();
1165 0 :
1166 0 : tracing::info!("Wrote timeline_detail for {} touched layers", touched.len());
1167 0 : touched.into_iter().for_each(|t| {
1168 0 : timeline_detail.touch_layer(
1169 0 : self.conf,
1170 0 : tenant_shard_id,
1171 0 : &timeline_id,
1172 0 : &t,
1173 0 : &self.secondary_state.resident_size_metric,
1174 0 : || {
1175 0 : local_layer_path(
1176 0 : self.conf,
1177 0 : tenant_shard_id,
1178 0 : &timeline_id,
1179 0 : &t.name,
1180 0 : &t.metadata.generation,
1181 0 : )
1182 0 : },
1183 0 : )
1184 0 : });
1185 0 : }
1186 0 :
1187 0 : result
1188 0 : }
1189 :
1190 : /// Call this during timeline download if a layer will _not_ be downloaded, to update progress statistics
1191 0 : fn skip_layer(&self, layer: HeatMapLayer) {
1192 0 : let mut progress = self.secondary_state.progress.lock().unwrap();
1193 0 : progress.layers_total = progress.layers_total.saturating_sub(1);
1194 0 : progress.bytes_total = progress
1195 0 : .bytes_total
1196 0 : .saturating_sub(layer.metadata.file_size);
1197 0 : }
1198 :
1199 0 : async fn download_layer(
1200 0 : &self,
1201 0 : tenant_shard_id: &TenantShardId,
1202 0 : timeline_id: &TimelineId,
1203 0 : layer: HeatMapLayer,
1204 0 : ctx: &RequestContext,
1205 0 : ) -> Result<Option<HeatMapLayer>, UpdateError> {
1206 0 : // Failpoints for simulating slow remote storage
1207 0 : failpoint_support::sleep_millis_async!(
1208 : "secondary-layer-download-sleep",
1209 0 : &self.secondary_state.cancel
1210 : );
1211 :
1212 0 : pausable_failpoint!("secondary-layer-download-pausable");
1213 :
1214 0 : let local_path = local_layer_path(
1215 0 : self.conf,
1216 0 : tenant_shard_id,
1217 0 : timeline_id,
1218 0 : &layer.name,
1219 0 : &layer.metadata.generation,
1220 0 : );
1221 0 :
1222 0 : // Note: no backoff::retry wrapper here because download_layer_file does its own retries internally
1223 0 : tracing::info!(
1224 0 : "Starting download of layer {}, size {}",
1225 : layer.name,
1226 : layer.metadata.file_size
1227 : );
1228 0 : let downloaded_bytes = download_layer_file(
1229 0 : self.conf,
1230 0 : self.remote_storage,
1231 0 : *tenant_shard_id,
1232 0 : *timeline_id,
1233 0 : &layer.name,
1234 0 : &layer.metadata,
1235 0 : &local_path,
1236 0 : &self.secondary_state.gate,
1237 0 : &self.secondary_state.cancel,
1238 0 : ctx,
1239 0 : )
1240 0 : .await;
1241 :
1242 0 : let downloaded_bytes = match downloaded_bytes {
1243 0 : Ok(bytes) => bytes,
1244 : Err(DownloadError::NotFound) => {
1245 : // A heatmap might be out of date and refer to a layer that doesn't exist any more.
1246 : // This is harmless: continue to download the next layer. It is expected during compaction
1247 : // GC.
1248 0 : tracing::debug!(
1249 0 : "Skipped downloading missing layer {}, raced with compaction/gc?",
1250 : layer.name
1251 : );
1252 0 : self.skip_layer(layer);
1253 0 :
1254 0 : return Ok(None);
1255 : }
1256 0 : Err(e) => return Err(e.into()),
1257 : };
1258 :
1259 0 : if downloaded_bytes != layer.metadata.file_size {
1260 0 : let local_path = local_layer_path(
1261 0 : self.conf,
1262 0 : tenant_shard_id,
1263 0 : timeline_id,
1264 0 : &layer.name,
1265 0 : &layer.metadata.generation,
1266 0 : );
1267 0 :
1268 0 : tracing::warn!(
1269 0 : "Downloaded layer {} with unexpected size {} != {}. Removing download.",
1270 : layer.name,
1271 : downloaded_bytes,
1272 : layer.metadata.file_size
1273 : );
1274 :
1275 0 : tokio::fs::remove_file(&local_path)
1276 0 : .await
1277 0 : .or_else(fs_ext::ignore_not_found)?;
1278 : } else {
1279 0 : tracing::info!("Downloaded layer {}, size {}", layer.name, downloaded_bytes);
1280 0 : let mut progress = self.secondary_state.progress.lock().unwrap();
1281 0 : progress.bytes_downloaded += downloaded_bytes;
1282 0 : progress.layers_downloaded += 1;
1283 : }
1284 :
1285 0 : SECONDARY_MODE.download_layer.inc();
1286 0 :
1287 0 : Ok(Some(layer))
1288 0 : }
1289 : }
1290 :
1291 : /// Scan local storage and build up Layer objects based on the metadata in a HeatMapTimeline
1292 0 : async fn init_timeline_state(
1293 0 : conf: &'static PageServerConf,
1294 0 : tenant_shard_id: &TenantShardId,
1295 0 : last_heatmap: Option<&HeatMapTimeline>,
1296 0 : heatmap: &HeatMapTimeline,
1297 0 : resident_metric: &UIntGauge,
1298 0 : ) -> SecondaryDetailTimeline {
1299 0 : let timeline_path = conf.timeline_path(tenant_shard_id, &heatmap.timeline_id);
1300 0 : let mut detail = SecondaryDetailTimeline::default();
1301 :
1302 0 : let mut dir = match tokio::fs::read_dir(&timeline_path).await {
1303 0 : Ok(d) => d,
1304 0 : Err(e) => {
1305 0 : if e.kind() == std::io::ErrorKind::NotFound {
1306 0 : let context = format!("Creating timeline directory {timeline_path}");
1307 0 : tracing::info!("{}", context);
1308 0 : tokio::fs::create_dir_all(&timeline_path)
1309 0 : .await
1310 0 : .fatal_err(&context);
1311 0 :
1312 0 : // No entries to report: drop out.
1313 0 : return detail;
1314 : } else {
1315 0 : on_fatal_io_error(&e, &format!("Reading timeline dir {timeline_path}"));
1316 : }
1317 : }
1318 : };
1319 :
1320 : // As we iterate through layers found on disk, we will look up their metadata from this map.
1321 : // Layers not present in metadata will be discarded.
1322 0 : let heatmap_metadata: HashMap<&LayerName, &HeatMapLayer> =
1323 0 : heatmap.hot_layers().map(|l| (&l.name, l)).collect();
1324 :
1325 0 : let last_heatmap_metadata: HashMap<&LayerName, &HeatMapLayer> =
1326 0 : if let Some(last_heatmap) = last_heatmap {
1327 0 : last_heatmap.hot_layers().map(|l| (&l.name, l)).collect()
1328 : } else {
1329 0 : HashMap::new()
1330 : };
1331 :
1332 0 : while let Some(dentry) = dir
1333 0 : .next_entry()
1334 0 : .await
1335 0 : .fatal_err(&format!("Listing {timeline_path}"))
1336 : {
1337 0 : let Ok(file_path) = Utf8PathBuf::from_path_buf(dentry.path()) else {
1338 0 : tracing::warn!("Malformed filename at {}", dentry.path().to_string_lossy());
1339 0 : continue;
1340 : };
1341 0 : let local_meta = dentry
1342 0 : .metadata()
1343 0 : .await
1344 0 : .fatal_err(&format!("Read metadata on {}", file_path));
1345 0 :
1346 0 : let file_name = file_path.file_name().expect("created it from the dentry");
1347 0 : if crate::is_temporary(&file_path)
1348 0 : || is_temp_download_file(&file_path)
1349 0 : || is_ephemeral_file(file_name)
1350 : {
1351 : // Temporary files are frequently left behind from restarting during downloads
1352 0 : tracing::info!("Cleaning up temporary file {file_path}");
1353 0 : if let Err(e) = tokio::fs::remove_file(&file_path)
1354 0 : .await
1355 0 : .or_else(fs_ext::ignore_not_found)
1356 : {
1357 0 : tracing::error!("Failed to remove temporary file {file_path}: {e}");
1358 0 : }
1359 0 : continue;
1360 0 : }
1361 0 :
1362 0 : match LayerName::from_str(file_name) {
1363 0 : Ok(name) => {
1364 0 : let remote_meta = heatmap_metadata.get(&name);
1365 0 : let last_meta = last_heatmap_metadata.get(&name);
1366 0 : let mut remove = false;
1367 0 : match remote_meta {
1368 0 : Some(remote_meta) => {
1369 0 : let last_meta_generation_file_size = last_meta
1370 0 : .map(|m| m.metadata.generation_file_size())
1371 0 : .unwrap_or(remote_meta.metadata.generation_file_size());
1372 0 : // TODO: checksums for layers (https://github.com/neondatabase/neon/issues/2784)
1373 0 : if remote_meta.metadata.generation_file_size()
1374 0 : != last_meta_generation_file_size
1375 : {
1376 0 : tracing::info!(
1377 0 : "Removing local layer {name} as on-disk json metadata has different generation or file size from remote: {:?} -> {:?}",
1378 0 : last_meta_generation_file_size,
1379 0 : remote_meta.metadata.generation_file_size()
1380 : );
1381 0 : remove = true;
1382 0 : } else if local_meta.len() != remote_meta.metadata.file_size {
1383 : // This can happen in the presence of race conditions: the remote and on-disk metadata have changed, but we haven't had
1384 : // the chance yet to download the new layer to disk, before the process restarted.
1385 0 : tracing::info!(
1386 0 : "Removing local layer {name} with unexpected local size {} != {}",
1387 0 : local_meta.len(),
1388 : remote_meta.metadata.file_size
1389 : );
1390 0 : remove = true;
1391 0 : } else {
1392 0 : // We expect the access time to be initialized immediately afterwards, when
1393 0 : // the latest heatmap is applied to the state.
1394 0 : detail.touch_layer(
1395 0 : conf,
1396 0 : tenant_shard_id,
1397 0 : &heatmap.timeline_id,
1398 0 : remote_meta,
1399 0 : resident_metric,
1400 0 : || file_path,
1401 0 : );
1402 0 : }
1403 : }
1404 : None => {
1405 : // FIXME: consider some optimization when transitioning from attached to secondary: maybe
1406 : // wait until we have seen a heatmap that is more recent than the most recent on-disk state? Otherwise
1407 : // we will end up deleting any layers which were created+uploaded more recently than the heatmap.
1408 0 : tracing::info!(
1409 0 : "Removing secondary local layer {} because it's absent in heatmap",
1410 : name
1411 : );
1412 0 : remove = true;
1413 : }
1414 : }
1415 0 : if remove {
1416 0 : tokio::fs::remove_file(&dentry.path())
1417 0 : .await
1418 0 : .or_else(fs_ext::ignore_not_found)
1419 0 : .fatal_err(&format!(
1420 0 : "Removing layer {}",
1421 0 : dentry.path().to_string_lossy()
1422 0 : ));
1423 0 : }
1424 : }
1425 : Err(_) => {
1426 : // Ignore it.
1427 0 : tracing::warn!("Unexpected file in timeline directory: {file_name}");
1428 : }
1429 : }
1430 : }
1431 :
1432 0 : detail
1433 0 : }
1434 :
1435 : /// Loads a json-encoded heatmap file from the provided on-disk path
1436 0 : async fn load_heatmap(
1437 0 : path: &Utf8PathBuf,
1438 0 : ctx: &RequestContext,
1439 0 : ) -> Result<Option<HeatMapTenant>, anyhow::Error> {
1440 0 : let mut file = match VirtualFile::open(path, ctx).await {
1441 0 : Ok(file) => file,
1442 0 : Err(e) if e.kind() == std::io::ErrorKind::NotFound => return Ok(None),
1443 0 : Err(e) => Err(e)?,
1444 : };
1445 0 : let st = file.read_to_string(ctx).await?;
1446 0 : let htm = serde_json::from_str(&st)?;
1447 0 : Ok(Some(htm))
1448 0 : }
|
