Line data Source code
1 : //! Periodically collect consumption metrics for all active tenants
2 : //! and push them to a HTTP endpoint.
3 : use crate::context::{DownloadBehavior, RequestContext};
4 : use crate::task_mgr::{self, TaskKind, BACKGROUND_RUNTIME};
5 : use crate::tenant::tasks::BackgroundLoopKind;
6 : use crate::tenant::{
7 : mgr::TenantManager, LogicalSizeCalculationCause, PageReconstructError, Tenant,
8 : };
9 : use camino::Utf8PathBuf;
10 : use consumption_metrics::EventType;
11 : use pageserver_api::models::TenantState;
12 : use remote_storage::{GenericRemoteStorage, RemoteStorageConfig};
13 : use reqwest::Url;
14 : use std::collections::HashMap;
15 : use std::sync::Arc;
16 : use std::time::{Duration, SystemTime};
17 : use tokio::time::Instant;
18 : use tokio_util::sync::CancellationToken;
19 : use tracing::*;
20 : use utils::id::NodeId;
21 :
22 : mod metrics;
23 : use crate::consumption_metrics::metrics::MetricsKey;
24 : mod disk_cache;
25 : mod upload;
26 :
27 : const DEFAULT_HTTP_REPORTING_TIMEOUT: Duration = Duration::from_secs(60);
28 :
29 : /// Basically a key-value pair, but usually in a Vec except for [`Cache`].
30 : ///
31 : /// This is as opposed to `consumption_metrics::Event` which is the externally communicated form.
32 : /// Difference is basically the missing idempotency key, which lives only for the duration of
33 : /// upload attempts.
34 : type RawMetric = (MetricsKey, (EventType, u64));
35 :
36 : /// Caches the [`RawMetric`]s
37 : ///
38 : /// In practice, during startup, last sent values are stored here to be used in calculating new
39 : /// ones. After successful uploading, the cached values are updated to cache. This used to be used
40 : /// for deduplication, but that is no longer needed.
41 : type Cache = HashMap<MetricsKey, (EventType, u64)>;
42 :
43 : /// Main thread that serves metrics collection
44 : #[allow(clippy::too_many_arguments)]
45 0 : pub async fn collect_metrics(
46 0 : tenant_manager: Arc<TenantManager>,
47 0 : metric_collection_endpoint: &Url,
48 0 : metric_collection_bucket: &Option<RemoteStorageConfig>,
49 0 : metric_collection_interval: Duration,
50 0 : _cached_metric_collection_interval: Duration,
51 0 : synthetic_size_calculation_interval: Duration,
52 0 : node_id: NodeId,
53 0 : local_disk_storage: Utf8PathBuf,
54 0 : cancel: CancellationToken,
55 0 : ctx: RequestContext,
56 0 : ) -> anyhow::Result<()> {
57 0 : if _cached_metric_collection_interval != Duration::ZERO {
58 0 : tracing::warn!(
59 0 : "cached_metric_collection_interval is no longer used, please set it to zero."
60 : )
61 0 : }
62 :
63 : // spin up background worker that caclulates tenant sizes
64 0 : let worker_ctx =
65 0 : ctx.detached_child(TaskKind::CalculateSyntheticSize, DownloadBehavior::Download);
66 0 : task_mgr::spawn(
67 0 : BACKGROUND_RUNTIME.handle(),
68 0 : TaskKind::CalculateSyntheticSize,
69 0 : None,
70 0 : None,
71 0 : "synthetic size calculation",
72 0 : false,
73 0 : {
74 0 : let tenant_manager = tenant_manager.clone();
75 0 : async move {
76 0 : calculate_synthetic_size_worker(
77 0 : tenant_manager,
78 0 : synthetic_size_calculation_interval,
79 0 : &cancel,
80 0 : &worker_ctx,
81 0 : )
82 0 : .instrument(info_span!("synthetic_size_worker"))
83 0 : .await?;
84 0 : Ok(())
85 0 : }
86 0 : },
87 0 : );
88 0 :
89 0 : let path: Arc<Utf8PathBuf> = Arc::new(local_disk_storage);
90 0 :
91 0 : let cancel = task_mgr::shutdown_token();
92 0 :
93 0 : let restore_and_reschedule = restore_and_reschedule(&path, metric_collection_interval);
94 :
95 0 : let mut cached_metrics = tokio::select! {
96 : _ = cancel.cancelled() => return Ok(()),
97 : ret = restore_and_reschedule => ret,
98 : };
99 :
100 : // define client here to reuse it for all requests
101 0 : let client = reqwest::ClientBuilder::new()
102 0 : .timeout(DEFAULT_HTTP_REPORTING_TIMEOUT)
103 0 : .build()
104 0 : .expect("Failed to create http client with timeout");
105 :
106 0 : let bucket_client = if let Some(bucket_config) = metric_collection_bucket {
107 0 : match GenericRemoteStorage::from_config(bucket_config) {
108 0 : Ok(client) => Some(client),
109 0 : Err(e) => {
110 0 : // Non-fatal error: if we were given an invalid config, we will proceed
111 0 : // with sending metrics over the network, but not to S3.
112 0 : tracing::warn!("Invalid configuration for metric_collection_bucket: {e}");
113 0 : None
114 : }
115 : }
116 : } else {
117 0 : None
118 : };
119 :
120 0 : let node_id = node_id.to_string();
121 :
122 0 : loop {
123 0 : let started_at = Instant::now();
124 :
125 : // these are point in time, with variable "now"
126 0 : let metrics = metrics::collect_all_metrics(&tenant_manager, &cached_metrics, &ctx).await;
127 :
128 0 : let metrics = Arc::new(metrics);
129 0 :
130 0 : // why not race cancellation here? because we are one of the last tasks, and if we are
131 0 : // already here, better to try to flush the new values.
132 0 :
133 0 : let flush = async {
134 0 : match disk_cache::flush_metrics_to_disk(&metrics, &path).await {
135 : Ok(()) => {
136 0 : tracing::debug!("flushed metrics to disk");
137 : }
138 0 : Err(e) => {
139 0 : // idea here is that if someone creates a directory as our path, then they
140 0 : // might notice it from the logs before shutdown and remove it
141 0 : tracing::error!("failed to persist metrics to {path:?}: {e:#}");
142 : }
143 : }
144 :
145 0 : if let Some(bucket_client) = &bucket_client {
146 0 : let res =
147 0 : upload::upload_metrics_bucket(bucket_client, &cancel, &node_id, &metrics).await;
148 0 : if let Err(e) = res {
149 0 : tracing::error!("failed to upload to S3: {e:#}");
150 0 : }
151 0 : }
152 0 : };
153 :
154 0 : let upload = async {
155 0 : let res = upload::upload_metrics_http(
156 0 : &client,
157 0 : metric_collection_endpoint,
158 0 : &cancel,
159 0 : &node_id,
160 0 : &metrics,
161 0 : &mut cached_metrics,
162 0 : )
163 0 : .await;
164 0 : if let Err(e) = res {
165 : // serialization error which should never happen
166 0 : tracing::error!("failed to upload via HTTP due to {e:#}");
167 0 : }
168 0 : };
169 :
170 : // let these run concurrently
171 : let (_, _) = tokio::join!(flush, upload);
172 :
173 0 : crate::tenant::tasks::warn_when_period_overrun(
174 0 : started_at.elapsed(),
175 0 : metric_collection_interval,
176 0 : BackgroundLoopKind::ConsumptionMetricsCollectMetrics,
177 0 : );
178 :
179 0 : let res = tokio::time::timeout_at(
180 0 : started_at + metric_collection_interval,
181 0 : task_mgr::shutdown_token().cancelled(),
182 0 : )
183 0 : .await;
184 0 : if res.is_ok() {
185 0 : return Ok(());
186 0 : }
187 : }
188 0 : }
189 :
190 : /// Called on the first iteration in an attempt to join the metric uploading schedule from previous
191 : /// pageserver session. Pageserver is supposed to upload at intervals regardless of restarts.
192 : ///
193 : /// Cancellation safe.
194 0 : async fn restore_and_reschedule(
195 0 : path: &Arc<Utf8PathBuf>,
196 0 : metric_collection_interval: Duration,
197 0 : ) -> Cache {
198 0 : let (cached, earlier_metric_at) = match disk_cache::read_metrics_from_disk(path.clone()).await {
199 0 : Ok(found_some) => {
200 0 : // there is no min needed because we write these sequentially in
201 0 : // collect_all_metrics
202 0 : let earlier_metric_at = found_some
203 0 : .iter()
204 0 : .map(|(_, (et, _))| et.recorded_at())
205 0 : .copied()
206 0 : .next();
207 0 :
208 0 : let cached = found_some.into_iter().collect::<Cache>();
209 0 :
210 0 : (cached, earlier_metric_at)
211 : }
212 0 : Err(e) => {
213 0 : use std::io::{Error, ErrorKind};
214 0 :
215 0 : let root = e.root_cause();
216 0 : let maybe_ioerr = root.downcast_ref::<Error>();
217 0 : let is_not_found = maybe_ioerr.is_some_and(|e| e.kind() == ErrorKind::NotFound);
218 0 :
219 0 : if !is_not_found {
220 0 : tracing::info!("failed to read any previous metrics from {path:?}: {e:#}");
221 0 : }
222 :
223 0 : (HashMap::new(), None)
224 : }
225 : };
226 :
227 0 : if let Some(earlier_metric_at) = earlier_metric_at {
228 0 : let earlier_metric_at: SystemTime = earlier_metric_at.into();
229 :
230 0 : let error = reschedule(earlier_metric_at, metric_collection_interval).await;
231 :
232 0 : if let Some(error) = error {
233 0 : if error.as_secs() >= 60 {
234 0 : tracing::info!(
235 0 : error_ms = error.as_millis(),
236 0 : "startup scheduling error due to restart"
237 : )
238 0 : }
239 0 : }
240 0 : }
241 :
242 0 : cached
243 0 : }
244 :
245 0 : async fn reschedule(
246 0 : earlier_metric_at: SystemTime,
247 0 : metric_collection_interval: Duration,
248 0 : ) -> Option<Duration> {
249 0 : let now = SystemTime::now();
250 0 : match now.duration_since(earlier_metric_at) {
251 0 : Ok(from_last_send) if from_last_send < metric_collection_interval => {
252 0 : let sleep_for = metric_collection_interval - from_last_send;
253 0 :
254 0 : let deadline = std::time::Instant::now() + sleep_for;
255 0 :
256 0 : tokio::time::sleep_until(deadline.into()).await;
257 :
258 0 : let now = std::time::Instant::now();
259 0 :
260 0 : // executor threads might be busy, add extra measurements
261 0 : Some(if now < deadline {
262 0 : deadline - now
263 : } else {
264 0 : now - deadline
265 : })
266 : }
267 0 : Ok(from_last_send) => Some(from_last_send.saturating_sub(metric_collection_interval)),
268 : Err(_) => {
269 0 : tracing::warn!(
270 : ?now,
271 : ?earlier_metric_at,
272 0 : "oldest recorded metric is in future; first values will come out with inconsistent timestamps"
273 : );
274 0 : earlier_metric_at.duration_since(now).ok()
275 : }
276 : }
277 0 : }
278 :
279 : /// Caclculate synthetic size for each active tenant
280 0 : async fn calculate_synthetic_size_worker(
281 0 : tenant_manager: Arc<TenantManager>,
282 0 : synthetic_size_calculation_interval: Duration,
283 0 : cancel: &CancellationToken,
284 0 : ctx: &RequestContext,
285 0 : ) -> anyhow::Result<()> {
286 0 : info!("starting calculate_synthetic_size_worker");
287 : scopeguard::defer! {
288 : info!("calculate_synthetic_size_worker stopped");
289 : };
290 :
291 0 : loop {
292 0 : let started_at = Instant::now();
293 :
294 0 : let tenants = match tenant_manager.list_tenants() {
295 0 : Ok(tenants) => tenants,
296 0 : Err(e) => {
297 0 : warn!("cannot get tenant list: {e:#}");
298 0 : continue;
299 : }
300 : };
301 :
302 0 : for (tenant_shard_id, tenant_state, _gen) in tenants {
303 0 : if tenant_state != TenantState::Active {
304 0 : continue;
305 0 : }
306 0 :
307 0 : if !tenant_shard_id.is_shard_zero() {
308 : // We only send consumption metrics from shard 0, so don't waste time calculating
309 : // synthetic size on other shards.
310 0 : continue;
311 0 : }
312 :
313 0 : let Ok(tenant) = tenant_manager.get_attached_tenant_shard(tenant_shard_id) else {
314 0 : continue;
315 : };
316 :
317 0 : if !tenant.is_active() {
318 0 : continue;
319 0 : }
320 0 :
321 0 : // there is never any reason to exit calculate_synthetic_size_worker following any
322 0 : // return value -- we don't need to care about shutdown because no tenant is found when
323 0 : // pageserver is shut down.
324 0 : calculate_and_log(&tenant, cancel, ctx).await;
325 : }
326 :
327 0 : crate::tenant::tasks::warn_when_period_overrun(
328 0 : started_at.elapsed(),
329 0 : synthetic_size_calculation_interval,
330 0 : BackgroundLoopKind::ConsumptionMetricsSyntheticSizeWorker,
331 0 : );
332 :
333 0 : let res = tokio::time::timeout_at(
334 0 : started_at + synthetic_size_calculation_interval,
335 0 : cancel.cancelled(),
336 0 : )
337 0 : .await;
338 0 : if res.is_ok() {
339 0 : return Ok(());
340 0 : }
341 : }
342 0 : }
343 :
344 0 : async fn calculate_and_log(tenant: &Tenant, cancel: &CancellationToken, ctx: &RequestContext) {
345 : const CAUSE: LogicalSizeCalculationCause =
346 : LogicalSizeCalculationCause::ConsumptionMetricsSyntheticSize;
347 :
348 : // TODO should we use concurrent_background_tasks_rate_limit() here, like the other background tasks?
349 : // We can put in some prioritization for consumption metrics.
350 : // Same for the loop that fetches computed metrics.
351 : // By using the same limiter, we centralize metrics collection for "start" and "finished" counters,
352 : // which turns out is really handy to understand the system.
353 0 : let Err(e) = tenant.calculate_synthetic_size(CAUSE, cancel, ctx).await else {
354 0 : return;
355 : };
356 :
357 : // this error can be returned if timeline is shutting down, but it does not
358 : // mean the synthetic size worker should terminate.
359 0 : let shutting_down = matches!(
360 0 : e.downcast_ref::<PageReconstructError>(),
361 : Some(PageReconstructError::Cancelled | PageReconstructError::AncestorStopping(_))
362 : );
363 :
364 0 : if !shutting_down {
365 0 : let tenant_shard_id = tenant.tenant_shard_id();
366 0 : error!("failed to calculate synthetic size for tenant {tenant_shard_id}: {e:#}");
367 0 : }
368 0 : }
|
