Line data Source code
1 : mod downloader;
2 : pub mod heatmap;
3 : mod heatmap_uploader;
4 : mod scheduler;
5 :
6 : use std::{sync::Arc, time::SystemTime};
7 :
8 : use crate::{
9 : context::RequestContext,
10 : disk_usage_eviction_task::DiskUsageEvictionInfo,
11 : metrics::SECONDARY_HEATMAP_TOTAL_SIZE,
12 : task_mgr::{self, TaskKind, BACKGROUND_RUNTIME},
13 : };
14 :
15 : use self::{
16 : downloader::{downloader_task, SecondaryDetail},
17 : heatmap_uploader::heatmap_uploader_task,
18 : };
19 :
20 : use super::{
21 : config::{SecondaryLocationConfig, TenantConfOpt},
22 : mgr::TenantManager,
23 : span::debug_assert_current_span_has_tenant_id,
24 : storage_layer::LayerName,
25 : };
26 :
27 : use crate::metrics::SECONDARY_RESIDENT_PHYSICAL_SIZE;
28 : use metrics::UIntGauge;
29 : use pageserver_api::{
30 : models,
31 : shard::{ShardIdentity, TenantShardId},
32 : };
33 : use remote_storage::GenericRemoteStorage;
34 :
35 : use tokio::task::JoinHandle;
36 : use tokio_util::sync::CancellationToken;
37 : use tracing::instrument;
38 : use utils::{completion::Barrier, id::TimelineId, sync::gate::Gate};
39 :
40 : enum DownloadCommand {
41 : Download(TenantShardId),
42 : }
43 : enum UploadCommand {
44 : Upload(TenantShardId),
45 : }
46 :
47 : impl UploadCommand {
48 0 : fn get_tenant_shard_id(&self) -> &TenantShardId {
49 0 : match self {
50 0 : Self::Upload(id) => id,
51 0 : }
52 0 : }
53 : }
54 :
55 : impl DownloadCommand {
56 0 : fn get_tenant_shard_id(&self) -> &TenantShardId {
57 0 : match self {
58 0 : Self::Download(id) => id,
59 0 : }
60 0 : }
61 : }
62 :
63 : struct CommandRequest<T> {
64 : payload: T,
65 : response_tx: tokio::sync::oneshot::Sender<CommandResponse>,
66 : }
67 :
68 : struct CommandResponse {
69 : result: anyhow::Result<()>,
70 : }
71 :
72 : // Whereas [`Tenant`] represents an attached tenant, this type represents the work
73 : // we do for secondary tenant locations: where we are not serving clients or
74 : // ingesting WAL, but we are maintaining a warm cache of layer files.
75 : //
76 : // This type is all about the _download_ path for secondary mode. The upload path
77 : // runs separately (see [`heatmap_uploader`]) while a regular attached `Tenant` exists.
78 : //
79 : // This structure coordinates TenantManager and SecondaryDownloader,
80 : // so that the downloader can indicate which tenants it is currently
81 : // operating on, and the manager can indicate when a particular
82 : // secondary tenant should cancel any work in flight.
83 : #[derive(Debug)]
84 : pub(crate) struct SecondaryTenant {
85 : /// Carrying a tenant shard ID simplifies callers such as the downloader
86 : /// which need to organize many of these objects by ID.
87 : tenant_shard_id: TenantShardId,
88 :
89 : /// Cancellation token indicates to SecondaryDownloader that it should stop doing
90 : /// any work for this tenant at the next opportunity.
91 : pub(crate) cancel: CancellationToken,
92 :
93 : pub(crate) gate: Gate,
94 :
95 : // Secondary mode does not need the full shard identity or the TenantConfOpt. However,
96 : // storing these enables us to report our full LocationConf, enabling convenient reconciliation
97 : // by the control plane (see [`Self::get_location_conf`])
98 : shard_identity: ShardIdentity,
99 : tenant_conf: std::sync::Mutex<TenantConfOpt>,
100 :
101 : // Internal state used by the Downloader.
102 : detail: std::sync::Mutex<SecondaryDetail>,
103 :
104 : // Public state indicating overall progress of downloads relative to the last heatmap seen
105 : pub(crate) progress: std::sync::Mutex<models::SecondaryProgress>,
106 :
107 : // Sum of layer sizes on local disk
108 : pub(super) resident_size_metric: UIntGauge,
109 :
110 : // Sum of layer sizes in the most recently downloaded heatmap
111 : pub(super) heatmap_total_size_metric: UIntGauge,
112 : }
113 :
114 : impl SecondaryTenant {
115 0 : pub(crate) fn new(
116 0 : tenant_shard_id: TenantShardId,
117 0 : shard_identity: ShardIdentity,
118 0 : tenant_conf: TenantConfOpt,
119 0 : config: &SecondaryLocationConfig,
120 0 : ) -> Arc<Self> {
121 0 : let tenant_id = tenant_shard_id.tenant_id.to_string();
122 0 : let shard_id = format!("{}", tenant_shard_id.shard_slug());
123 0 : let resident_size_metric = SECONDARY_RESIDENT_PHYSICAL_SIZE
124 0 : .get_metric_with_label_values(&[&tenant_id, &shard_id])
125 0 : .unwrap();
126 0 :
127 0 : let heatmap_total_size_metric = SECONDARY_HEATMAP_TOTAL_SIZE
128 0 : .get_metric_with_label_values(&[&tenant_id, &shard_id])
129 0 : .unwrap();
130 0 :
131 0 : Arc::new(Self {
132 0 : tenant_shard_id,
133 0 : // todo: shall we make this a descendent of the
134 0 : // main cancellation token, or is it sufficient that
135 0 : // on shutdown we walk the tenants and fire their
136 0 : // individual cancellations?
137 0 : cancel: CancellationToken::new(),
138 0 : gate: Gate::default(),
139 0 :
140 0 : shard_identity,
141 0 : tenant_conf: std::sync::Mutex::new(tenant_conf),
142 0 :
143 0 : detail: std::sync::Mutex::new(SecondaryDetail::new(config.clone())),
144 0 :
145 0 : progress: std::sync::Mutex::default(),
146 0 :
147 0 : resident_size_metric,
148 0 : heatmap_total_size_metric,
149 0 : })
150 0 : }
151 :
152 0 : pub(crate) fn tenant_shard_id(&self) -> TenantShardId {
153 0 : self.tenant_shard_id
154 0 : }
155 :
156 0 : pub(crate) async fn shutdown(&self) {
157 0 : self.cancel.cancel();
158 0 :
159 0 : // Wait for any secondary downloader work to complete
160 0 : self.gate.close().await;
161 :
162 0 : self.validate_metrics();
163 0 :
164 0 : let tenant_id = self.tenant_shard_id.tenant_id.to_string();
165 0 : let shard_id = format!("{}", self.tenant_shard_id.shard_slug());
166 0 : let _ = SECONDARY_RESIDENT_PHYSICAL_SIZE.remove_label_values(&[&tenant_id, &shard_id]);
167 0 : let _ = SECONDARY_HEATMAP_TOTAL_SIZE.remove_label_values(&[&tenant_id, &shard_id]);
168 0 : }
169 :
170 0 : pub(crate) fn set_config(&self, config: &SecondaryLocationConfig) {
171 0 : self.detail.lock().unwrap().config = config.clone();
172 0 : }
173 :
174 0 : pub(crate) fn set_tenant_conf(&self, config: &TenantConfOpt) {
175 0 : *(self.tenant_conf.lock().unwrap()) = config.clone();
176 0 : }
177 :
178 : /// For API access: generate a LocationConfig equivalent to the one that would be used to
179 : /// create a Tenant in the same state. Do not use this in hot paths: it's for relatively
180 : /// rare external API calls, like a reconciliation at startup.
181 0 : pub(crate) fn get_location_conf(&self) -> models::LocationConfig {
182 0 : let conf = self.detail.lock().unwrap().config.clone();
183 0 :
184 0 : let conf = models::LocationConfigSecondary { warm: conf.warm };
185 0 :
186 0 : let tenant_conf = self.tenant_conf.lock().unwrap().clone();
187 0 : models::LocationConfig {
188 0 : mode: models::LocationConfigMode::Secondary,
189 0 : generation: None,
190 0 : secondary_conf: Some(conf),
191 0 : shard_number: self.tenant_shard_id.shard_number.0,
192 0 : shard_count: self.tenant_shard_id.shard_count.literal(),
193 0 : shard_stripe_size: self.shard_identity.stripe_size.0,
194 0 : tenant_conf: tenant_conf.into(),
195 0 : }
196 0 : }
197 :
198 0 : pub(crate) fn get_tenant_shard_id(&self) -> &TenantShardId {
199 0 : &self.tenant_shard_id
200 0 : }
201 :
202 0 : pub(crate) fn get_layers_for_eviction(self: &Arc<Self>) -> (DiskUsageEvictionInfo, usize) {
203 0 : self.detail.lock().unwrap().get_layers_for_eviction(self)
204 0 : }
205 :
206 : /// Cancellation safe, but on cancellation the eviction will go through
207 0 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug(), timeline_id=%timeline_id, name=%name))]
208 : pub(crate) async fn evict_layer(self: &Arc<Self>, timeline_id: TimelineId, name: LayerName) {
209 : debug_assert_current_span_has_tenant_id();
210 :
211 : let guard = match self.gate.enter() {
212 : Ok(g) => g,
213 : Err(_) => {
214 : tracing::debug!("Dropping layer evictions, secondary tenant shutting down",);
215 : return;
216 : }
217 : };
218 :
219 : let now = SystemTime::now();
220 : tracing::info!("Evicting secondary layer");
221 :
222 : let this = self.clone();
223 :
224 : // spawn it to be cancellation safe
225 0 : tokio::task::spawn_blocking(move || {
226 0 : let _guard = guard;
227 0 :
228 0 : // Update the timeline's state. This does not have to be synchronized with
229 0 : // the download process, because:
230 0 : // - If downloader is racing with us to remove a file (e.g. because it is
231 0 : // removed from heatmap), then our mutual .remove() operations will both
232 0 : // succeed.
233 0 : // - If downloader is racing with us to download the object (this would require
234 0 : // multiple eviction iterations to race with multiple download iterations), then
235 0 : // if we remove it from the state, the worst that happens is the downloader
236 0 : // downloads it again before re-inserting, or we delete the file but it remains
237 0 : // in the state map (in which case it will be downloaded if this secondary
238 0 : // tenant transitions to attached and tries to access it)
239 0 : //
240 0 : // The important assumption here is that the secondary timeline state does not
241 0 : // have to 100% match what is on disk, because it's a best-effort warming
242 0 : // of the cache.
243 0 : let mut detail = this.detail.lock().unwrap();
244 0 : if let Some(removed) =
245 0 : detail.evict_layer(name, &timeline_id, now, &this.resident_size_metric)
246 0 : {
247 0 : // We might race with removal of the same layer during downloads, so finding the layer we
248 0 : // were trying to remove is optional. Only issue the disk I/O to remove it if we found it.
249 0 : removed.remove_blocking();
250 0 : }
251 0 : })
252 : .await
253 : .expect("secondary eviction should not have panicked");
254 : }
255 :
256 : /// Exhaustive check that incrementally updated metrics match the actual state.
257 : #[cfg(feature = "testing")]
258 0 : fn validate_metrics(&self) {
259 0 : let detail = self.detail.lock().unwrap();
260 0 : let resident_size = detail.total_resident_size();
261 0 :
262 0 : assert_eq!(resident_size, self.resident_size_metric.get());
263 0 : }
264 :
265 : #[cfg(not(feature = "testing"))]
266 : fn validate_metrics(&self) {
267 : // No-op in non-testing builds
268 : }
269 : }
270 :
271 : /// The SecondaryController is a pseudo-rpc client for administrative control of secondary mode downloads,
272 : /// and heatmap uploads. This is not a hot data path: it's used for:
273 : /// - Live migrations, where we want to ensure a migration destination has the freshest possible
274 : /// content before trying to cut over.
275 : /// - Tests, where we want to immediately upload/download for a particular tenant.
276 : ///
277 : /// In normal operations, outside of migrations, uploads & downloads are autonomous and not driven by this interface.
278 : pub struct SecondaryController {
279 : upload_req_tx: tokio::sync::mpsc::Sender<CommandRequest<UploadCommand>>,
280 : download_req_tx: tokio::sync::mpsc::Sender<CommandRequest<DownloadCommand>>,
281 : }
282 :
283 : impl SecondaryController {
284 0 : async fn dispatch<T>(
285 0 : &self,
286 0 : queue: &tokio::sync::mpsc::Sender<CommandRequest<T>>,
287 0 : payload: T,
288 0 : ) -> anyhow::Result<()> {
289 0 : let (response_tx, response_rx) = tokio::sync::oneshot::channel();
290 0 :
291 0 : queue
292 0 : .send(CommandRequest {
293 0 : payload,
294 0 : response_tx,
295 0 : })
296 0 : .await
297 0 : .map_err(|_| anyhow::anyhow!("Receiver shut down"))?;
298 :
299 0 : let response = response_rx
300 0 : .await
301 0 : .map_err(|_| anyhow::anyhow!("Request dropped"))?;
302 :
303 0 : response.result
304 0 : }
305 :
306 0 : pub async fn upload_tenant(&self, tenant_shard_id: TenantShardId) -> anyhow::Result<()> {
307 0 : self.dispatch(&self.upload_req_tx, UploadCommand::Upload(tenant_shard_id))
308 0 : .await
309 0 : }
310 0 : pub async fn download_tenant(&self, tenant_shard_id: TenantShardId) -> anyhow::Result<()> {
311 0 : self.dispatch(
312 0 : &self.download_req_tx,
313 0 : DownloadCommand::Download(tenant_shard_id),
314 0 : )
315 0 : .await
316 0 : }
317 : }
318 :
319 : pub struct GlobalTasks {
320 : cancel: CancellationToken,
321 : uploader: JoinHandle<()>,
322 : downloader: JoinHandle<()>,
323 : }
324 :
325 : impl GlobalTasks {
326 : /// Caller is responsible for requesting shutdown via the cancellation token that was
327 : /// passed to [`spawn_tasks`].
328 : ///
329 : /// # Panics
330 : ///
331 : /// This method panics if that token is not cancelled.
332 : /// This is low-risk because we're calling this during process shutdown, so, a panic
333 : /// will be informative but not cause undue downtime.
334 0 : pub async fn wait(self) {
335 0 : let Self {
336 0 : cancel,
337 0 : uploader,
338 0 : downloader,
339 0 : } = self;
340 0 : assert!(
341 0 : cancel.is_cancelled(),
342 0 : "must cancel cancellation token, otherwise the tasks will not shut down"
343 : );
344 :
345 0 : let (uploader, downloader) = futures::future::join(uploader, downloader).await;
346 0 : uploader.expect(
347 0 : "unreachable: exit_on_panic_or_error would catch the panic and exit the process",
348 0 : );
349 0 : downloader.expect(
350 0 : "unreachable: exit_on_panic_or_error would catch the panic and exit the process",
351 0 : );
352 0 : }
353 : }
354 :
355 0 : pub fn spawn_tasks(
356 0 : tenant_manager: Arc<TenantManager>,
357 0 : remote_storage: GenericRemoteStorage,
358 0 : background_jobs_can_start: Barrier,
359 0 : cancel: CancellationToken,
360 0 : ) -> (SecondaryController, GlobalTasks) {
361 0 : let mgr_clone = tenant_manager.clone();
362 0 : let storage_clone = remote_storage.clone();
363 0 : let bg_jobs_clone = background_jobs_can_start.clone();
364 0 :
365 0 : let (download_req_tx, download_req_rx) =
366 0 : tokio::sync::mpsc::channel::<CommandRequest<DownloadCommand>>(16);
367 0 : let (upload_req_tx, upload_req_rx) =
368 0 : tokio::sync::mpsc::channel::<CommandRequest<UploadCommand>>(16);
369 0 :
370 0 : let cancel_clone = cancel.clone();
371 0 : let downloader = BACKGROUND_RUNTIME.spawn(task_mgr::exit_on_panic_or_error(
372 0 : "secondary tenant downloads",
373 0 : async move {
374 0 : downloader_task(
375 0 : mgr_clone,
376 0 : storage_clone,
377 0 : download_req_rx,
378 0 : bg_jobs_clone,
379 0 : cancel_clone,
380 0 : RequestContext::new(
381 0 : TaskKind::SecondaryDownloads,
382 0 : crate::context::DownloadBehavior::Download,
383 0 : ),
384 0 : )
385 0 : .await;
386 0 : anyhow::Ok(())
387 0 : },
388 0 : ));
389 0 :
390 0 : let cancel_clone = cancel.clone();
391 0 : let uploader = BACKGROUND_RUNTIME.spawn(task_mgr::exit_on_panic_or_error(
392 0 : "heatmap uploads",
393 0 : async move {
394 0 : heatmap_uploader_task(
395 0 : tenant_manager,
396 0 : remote_storage,
397 0 : upload_req_rx,
398 0 : background_jobs_can_start,
399 0 : cancel_clone,
400 0 : )
401 0 : .await;
402 0 : anyhow::Ok(())
403 0 : },
404 0 : ));
405 0 :
406 0 : (
407 0 : SecondaryController {
408 0 : upload_req_tx,
409 0 : download_req_tx,
410 0 : },
411 0 : GlobalTasks {
412 0 : cancel,
413 0 : uploader,
414 0 : downloader,
415 0 : },
416 0 : )
417 0 : }
|
