Line data Source code
1 : //! Periodically collect consumption metrics for all active tenants
2 : //! and push them to a HTTP endpoint.
3 : use crate::config::PageServerConf;
4 : use crate::consumption_metrics::metrics::MetricsKey;
5 : use crate::consumption_metrics::upload::KeyGen as _;
6 : use crate::context::{DownloadBehavior, RequestContext};
7 : use crate::task_mgr::{self, TaskKind, BACKGROUND_RUNTIME};
8 : use crate::tenant::size::CalculateSyntheticSizeError;
9 : use crate::tenant::tasks::BackgroundLoopKind;
10 : use crate::tenant::{mgr::TenantManager, LogicalSizeCalculationCause, Tenant};
11 : use camino::Utf8PathBuf;
12 : use consumption_metrics::EventType;
13 : use itertools::Itertools as _;
14 : use pageserver_api::models::TenantState;
15 : use remote_storage::{GenericRemoteStorage, RemoteStorageConfig};
16 : use reqwest::Url;
17 : use std::collections::HashMap;
18 : use std::sync::Arc;
19 : use std::time::{Duration, SystemTime};
20 : use tokio::time::Instant;
21 : use tokio_util::sync::CancellationToken;
22 : use tracing::*;
23 : use utils::id::NodeId;
24 :
25 : mod disk_cache;
26 : mod metrics;
27 : mod upload;
28 :
29 : const DEFAULT_HTTP_REPORTING_TIMEOUT: Duration = Duration::from_secs(60);
30 :
31 : /// Basically a key-value pair, but usually in a Vec except for [`Cache`].
32 : ///
33 : /// This is as opposed to `consumption_metrics::Event` which is the externally communicated form.
34 : /// Difference is basically the missing idempotency key, which lives only for the duration of
35 : /// upload attempts.
36 : type RawMetric = (MetricsKey, (EventType, u64));
37 :
38 : /// Caches the [`RawMetric`]s
39 : ///
40 : /// In practice, during startup, last sent values are stored here to be used in calculating new
41 : /// ones. After successful uploading, the cached values are updated to cache. This used to be used
42 : /// for deduplication, but that is no longer needed.
43 : type Cache = HashMap<MetricsKey, (EventType, u64)>;
44 :
45 0 : pub async fn run(
46 0 : conf: &'static PageServerConf,
47 0 : tenant_manager: Arc<TenantManager>,
48 0 : cancel: CancellationToken,
49 0 : ) {
50 0 : let Some(metric_collection_endpoint) = conf.metric_collection_endpoint.as_ref() else {
51 0 : return;
52 : };
53 :
54 0 : let local_disk_storage = conf.workdir.join("last_consumption_metrics.json");
55 0 :
56 0 : let metrics_ctx = RequestContext::todo_child(
57 0 : TaskKind::MetricsCollection,
58 0 : // This task itself shouldn't download anything.
59 0 : // The actual size calculation does need downloads, and
60 0 : // creates a child context with the right DownloadBehavior.
61 0 : DownloadBehavior::Error,
62 0 : );
63 0 : let collect_metrics = BACKGROUND_RUNTIME.spawn(task_mgr::exit_on_panic_or_error(
64 0 : "consumption metrics collection",
65 0 : collect_metrics(
66 0 : tenant_manager.clone(),
67 0 : metric_collection_endpoint,
68 0 : &conf.metric_collection_bucket,
69 0 : conf.metric_collection_interval,
70 0 : conf.id,
71 0 : local_disk_storage,
72 0 : cancel.clone(),
73 0 : metrics_ctx,
74 0 : )
75 0 : .instrument(info_span!("metrics_collection")),
76 : ));
77 :
78 0 : let worker_ctx =
79 0 : RequestContext::todo_child(TaskKind::CalculateSyntheticSize, DownloadBehavior::Download);
80 0 : let synthetic_size_worker = BACKGROUND_RUNTIME.spawn(task_mgr::exit_on_panic_or_error(
81 0 : "synthetic size calculation",
82 0 : calculate_synthetic_size_worker(
83 0 : tenant_manager.clone(),
84 0 : conf.synthetic_size_calculation_interval,
85 0 : cancel.clone(),
86 0 : worker_ctx,
87 0 : )
88 0 : .instrument(info_span!("synthetic_size_worker")),
89 : ));
90 :
91 0 : let (collect_metrics, synthetic_size_worker) =
92 0 : futures::future::join(collect_metrics, synthetic_size_worker).await;
93 0 : collect_metrics
94 0 : .expect("unreachable: exit_on_panic_or_error would catch the panic and exit the process");
95 0 : synthetic_size_worker
96 0 : .expect("unreachable: exit_on_panic_or_error would catch the panic and exit the process");
97 0 : }
98 :
99 : /// Main thread that serves metrics collection
100 : #[allow(clippy::too_many_arguments)]
101 0 : async fn collect_metrics(
102 0 : tenant_manager: Arc<TenantManager>,
103 0 : metric_collection_endpoint: &Url,
104 0 : metric_collection_bucket: &Option<RemoteStorageConfig>,
105 0 : metric_collection_interval: Duration,
106 0 : node_id: NodeId,
107 0 : local_disk_storage: Utf8PathBuf,
108 0 : cancel: CancellationToken,
109 0 : ctx: RequestContext,
110 0 : ) -> anyhow::Result<()> {
111 0 : let path: Arc<Utf8PathBuf> = Arc::new(local_disk_storage);
112 0 :
113 0 : let restore_and_reschedule = restore_and_reschedule(&path, metric_collection_interval);
114 :
115 0 : let mut cached_metrics = tokio::select! {
116 0 : _ = cancel.cancelled() => return Ok(()),
117 0 : ret = restore_and_reschedule => ret,
118 0 : };
119 0 :
120 0 : // define client here to reuse it for all requests
121 0 : let client = reqwest::ClientBuilder::new()
122 0 : .timeout(DEFAULT_HTTP_REPORTING_TIMEOUT)
123 0 : .build()
124 0 : .expect("Failed to create http client with timeout");
125 :
126 0 : let bucket_client = if let Some(bucket_config) = metric_collection_bucket {
127 0 : match GenericRemoteStorage::from_config(bucket_config).await {
128 0 : Ok(client) => Some(client),
129 0 : Err(e) => {
130 0 : // Non-fatal error: if we were given an invalid config, we will proceed
131 0 : // with sending metrics over the network, but not to S3.
132 0 : tracing::warn!("Invalid configuration for metric_collection_bucket: {e}");
133 0 : None
134 : }
135 : }
136 : } else {
137 0 : None
138 : };
139 :
140 0 : let node_id = node_id.to_string();
141 :
142 : loop {
143 0 : let started_at = Instant::now();
144 :
145 : // these are point in time, with variable "now"
146 0 : let metrics = metrics::collect_all_metrics(&tenant_manager, &cached_metrics, &ctx).await;
147 :
148 : // Pre-generate event idempotency keys, to reuse them across the bucket
149 : // and HTTP sinks.
150 0 : let idempotency_keys = std::iter::repeat_with(|| node_id.as_str().generate())
151 0 : .take(metrics.len())
152 0 : .collect_vec();
153 0 :
154 0 : let metrics = Arc::new(metrics);
155 0 :
156 0 : // why not race cancellation here? because we are one of the last tasks, and if we are
157 0 : // already here, better to try to flush the new values.
158 0 :
159 0 : let flush = async {
160 0 : match disk_cache::flush_metrics_to_disk(&metrics, &path).await {
161 : Ok(()) => {
162 0 : tracing::debug!("flushed metrics to disk");
163 : }
164 0 : Err(e) => {
165 0 : // idea here is that if someone creates a directory as our path, then they
166 0 : // might notice it from the logs before shutdown and remove it
167 0 : tracing::error!("failed to persist metrics to {path:?}: {e:#}");
168 : }
169 : }
170 :
171 0 : if let Some(bucket_client) = &bucket_client {
172 0 : let res = upload::upload_metrics_bucket(
173 0 : bucket_client,
174 0 : &cancel,
175 0 : &node_id,
176 0 : &metrics,
177 0 : &idempotency_keys,
178 0 : )
179 0 : .await;
180 0 : if let Err(e) = res {
181 0 : tracing::error!("failed to upload to remote storage: {e:#}");
182 0 : }
183 0 : }
184 0 : };
185 :
186 0 : let upload = async {
187 0 : let res = upload::upload_metrics_http(
188 0 : &client,
189 0 : metric_collection_endpoint,
190 0 : &cancel,
191 0 : &metrics,
192 0 : &mut cached_metrics,
193 0 : &idempotency_keys,
194 0 : )
195 0 : .await;
196 0 : if let Err(e) = res {
197 : // serialization error which should never happen
198 0 : tracing::error!("failed to upload via HTTP due to {e:#}");
199 0 : }
200 0 : };
201 :
202 : // let these run concurrently
203 0 : let (_, _) = tokio::join!(flush, upload);
204 :
205 0 : crate::tenant::tasks::warn_when_period_overrun(
206 0 : started_at.elapsed(),
207 0 : metric_collection_interval,
208 0 : BackgroundLoopKind::ConsumptionMetricsCollectMetrics,
209 0 : );
210 :
211 0 : let res =
212 0 : tokio::time::timeout_at(started_at + metric_collection_interval, cancel.cancelled())
213 0 : .await;
214 0 : if res.is_ok() {
215 0 : return Ok(());
216 0 : }
217 : }
218 0 : }
219 :
220 : /// Called on the first iteration in an attempt to join the metric uploading schedule from previous
221 : /// pageserver session. Pageserver is supposed to upload at intervals regardless of restarts.
222 : ///
223 : /// Cancellation safe.
224 0 : async fn restore_and_reschedule(
225 0 : path: &Arc<Utf8PathBuf>,
226 0 : metric_collection_interval: Duration,
227 0 : ) -> Cache {
228 0 : let (cached, earlier_metric_at) = match disk_cache::read_metrics_from_disk(path.clone()).await {
229 0 : Ok(found_some) => {
230 0 : // there is no min needed because we write these sequentially in
231 0 : // collect_all_metrics
232 0 : let earlier_metric_at = found_some
233 0 : .iter()
234 0 : .map(|(_, (et, _))| et.recorded_at())
235 0 : .copied()
236 0 : .next();
237 0 :
238 0 : let cached = found_some.into_iter().collect::<Cache>();
239 0 :
240 0 : (cached, earlier_metric_at)
241 : }
242 0 : Err(e) => {
243 : use std::io::{Error, ErrorKind};
244 :
245 0 : let root = e.root_cause();
246 0 : let maybe_ioerr = root.downcast_ref::<Error>();
247 0 : let is_not_found = maybe_ioerr.is_some_and(|e| e.kind() == ErrorKind::NotFound);
248 0 :
249 0 : if !is_not_found {
250 0 : tracing::info!("failed to read any previous metrics from {path:?}: {e:#}");
251 0 : }
252 :
253 0 : (HashMap::new(), None)
254 : }
255 : };
256 :
257 0 : if let Some(earlier_metric_at) = earlier_metric_at {
258 0 : let earlier_metric_at: SystemTime = earlier_metric_at.into();
259 :
260 0 : let error = reschedule(earlier_metric_at, metric_collection_interval).await;
261 :
262 0 : if let Some(error) = error {
263 0 : if error.as_secs() >= 60 {
264 0 : tracing::info!(
265 0 : error_ms = error.as_millis(),
266 0 : "startup scheduling error due to restart"
267 : )
268 0 : }
269 0 : }
270 0 : }
271 :
272 0 : cached
273 0 : }
274 :
275 0 : async fn reschedule(
276 0 : earlier_metric_at: SystemTime,
277 0 : metric_collection_interval: Duration,
278 0 : ) -> Option<Duration> {
279 0 : let now = SystemTime::now();
280 0 : match now.duration_since(earlier_metric_at) {
281 0 : Ok(from_last_send) if from_last_send < metric_collection_interval => {
282 0 : let sleep_for = metric_collection_interval - from_last_send;
283 0 :
284 0 : let deadline = std::time::Instant::now() + sleep_for;
285 0 :
286 0 : tokio::time::sleep_until(deadline.into()).await;
287 :
288 0 : let now = std::time::Instant::now();
289 0 :
290 0 : // executor threads might be busy, add extra measurements
291 0 : Some(if now < deadline {
292 0 : deadline - now
293 : } else {
294 0 : now - deadline
295 : })
296 : }
297 0 : Ok(from_last_send) => Some(from_last_send.saturating_sub(metric_collection_interval)),
298 : Err(_) => {
299 0 : tracing::warn!(
300 : ?now,
301 : ?earlier_metric_at,
302 0 : "oldest recorded metric is in future; first values will come out with inconsistent timestamps"
303 : );
304 0 : earlier_metric_at.duration_since(now).ok()
305 : }
306 : }
307 0 : }
308 :
309 : /// Caclculate synthetic size for each active tenant
310 0 : async fn calculate_synthetic_size_worker(
311 0 : tenant_manager: Arc<TenantManager>,
312 0 : synthetic_size_calculation_interval: Duration,
313 0 : cancel: CancellationToken,
314 0 : ctx: RequestContext,
315 0 : ) -> anyhow::Result<()> {
316 0 : info!("starting calculate_synthetic_size_worker");
317 0 : scopeguard::defer! {
318 0 : info!("calculate_synthetic_size_worker stopped");
319 0 : };
320 :
321 : loop {
322 0 : let started_at = Instant::now();
323 :
324 0 : let tenants = match tenant_manager.list_tenants() {
325 0 : Ok(tenants) => tenants,
326 0 : Err(e) => {
327 0 : warn!("cannot get tenant list: {e:#}");
328 0 : continue;
329 : }
330 : };
331 :
332 0 : for (tenant_shard_id, tenant_state, _gen) in tenants {
333 0 : if tenant_state != TenantState::Active {
334 0 : continue;
335 0 : }
336 0 :
337 0 : if !tenant_shard_id.is_shard_zero() {
338 : // We only send consumption metrics from shard 0, so don't waste time calculating
339 : // synthetic size on other shards.
340 0 : continue;
341 0 : }
342 :
343 0 : let Ok(tenant) = tenant_manager.get_attached_tenant_shard(tenant_shard_id) else {
344 0 : continue;
345 : };
346 :
347 0 : if !tenant.is_active() {
348 0 : continue;
349 0 : }
350 0 :
351 0 : // there is never any reason to exit calculate_synthetic_size_worker following any
352 0 : // return value -- we don't need to care about shutdown because no tenant is found when
353 0 : // pageserver is shut down.
354 0 : calculate_and_log(&tenant, &cancel, &ctx).await;
355 : }
356 :
357 0 : crate::tenant::tasks::warn_when_period_overrun(
358 0 : started_at.elapsed(),
359 0 : synthetic_size_calculation_interval,
360 0 : BackgroundLoopKind::ConsumptionMetricsSyntheticSizeWorker,
361 0 : );
362 :
363 0 : let res = tokio::time::timeout_at(
364 0 : started_at + synthetic_size_calculation_interval,
365 0 : cancel.cancelled(),
366 0 : )
367 0 : .await;
368 0 : if res.is_ok() {
369 0 : return Ok(());
370 0 : }
371 : }
372 0 : }
373 :
374 0 : async fn calculate_and_log(tenant: &Tenant, cancel: &CancellationToken, ctx: &RequestContext) {
375 : const CAUSE: LogicalSizeCalculationCause =
376 : LogicalSizeCalculationCause::ConsumptionMetricsSyntheticSize;
377 :
378 : // TODO should we use concurrent_background_tasks_rate_limit() here, like the other background tasks?
379 : // We can put in some prioritization for consumption metrics.
380 : // Same for the loop that fetches computed metrics.
381 : // By using the same limiter, we centralize metrics collection for "start" and "finished" counters,
382 : // which turns out is really handy to understand the system.
383 0 : match tenant.calculate_synthetic_size(CAUSE, cancel, ctx).await {
384 0 : Ok(_) => {}
385 0 : Err(CalculateSyntheticSizeError::Cancelled) => {}
386 0 : Err(e) => {
387 0 : let tenant_shard_id = tenant.tenant_shard_id();
388 0 : error!("failed to calculate synthetic size for tenant {tenant_shard_id}: {e:#}");
389 : }
390 : }
391 0 : }
|
