Line data Source code
1 : use futures::{stream::FuturesUnordered, StreamExt};
2 : use safekeeper_api::models::SafekeeperUtilization;
3 : use safekeeper_client::mgmt_api;
4 : use std::{
5 : collections::HashMap,
6 : fmt::Debug,
7 : future::Future,
8 : sync::Arc,
9 : time::{Duration, Instant},
10 : };
11 : use tokio_util::sync::CancellationToken;
12 :
13 : use pageserver_api::{
14 : controller_api::{NodeAvailability, SkSchedulingPolicy},
15 : models::PageserverUtilization,
16 : };
17 :
18 : use thiserror::Error;
19 : use utils::{id::NodeId, logging::SecretString};
20 :
21 : use crate::{node::Node, safekeeper::Safekeeper};
22 :
23 : struct HeartbeaterTask<Server, State> {
24 : receiver: tokio::sync::mpsc::UnboundedReceiver<HeartbeatRequest<Server, State>>,
25 : cancel: CancellationToken,
26 :
27 : state: HashMap<NodeId, State>,
28 :
29 : max_offline_interval: Duration,
30 : max_warming_up_interval: Duration,
31 : jwt_token: Option<String>,
32 : }
33 :
34 : #[derive(Debug, Clone)]
35 : pub(crate) enum PageserverState {
36 : Available {
37 : last_seen_at: Instant,
38 : utilization: PageserverUtilization,
39 : },
40 : WarmingUp {
41 : started_at: Instant,
42 : },
43 : Offline,
44 : }
45 :
46 : #[derive(Debug, Clone)]
47 : pub(crate) enum SafekeeperState {
48 : Available {
49 : last_seen_at: Instant,
50 : utilization: SafekeeperUtilization,
51 : },
52 : Offline,
53 : }
54 :
55 : #[derive(Debug)]
56 : pub(crate) struct AvailablityDeltas<State>(pub Vec<(NodeId, State)>);
57 :
58 : #[derive(Debug, Error)]
59 : pub(crate) enum HeartbeaterError {
60 : #[error("Cancelled")]
61 : Cancel,
62 : }
63 :
64 : struct HeartbeatRequest<Server, State> {
65 : servers: Arc<HashMap<NodeId, Server>>,
66 : reply: tokio::sync::oneshot::Sender<Result<AvailablityDeltas<State>, HeartbeaterError>>,
67 : }
68 :
69 : pub(crate) struct Heartbeater<Server, State> {
70 : sender: tokio::sync::mpsc::UnboundedSender<HeartbeatRequest<Server, State>>,
71 : }
72 :
73 : #[allow(private_bounds)]
74 : impl<Server: Send + Sync + 'static, State: Debug + Send + 'static> Heartbeater<Server, State>
75 : where
76 : HeartbeaterTask<Server, State>: HeartBeat<Server, State>,
77 : {
78 0 : pub(crate) fn new(
79 0 : jwt_token: Option<String>,
80 0 : max_offline_interval: Duration,
81 0 : max_warming_up_interval: Duration,
82 0 : cancel: CancellationToken,
83 0 : ) -> Self {
84 0 : let (sender, receiver) =
85 0 : tokio::sync::mpsc::unbounded_channel::<HeartbeatRequest<Server, State>>();
86 0 : let mut heartbeater = HeartbeaterTask::new(
87 0 : receiver,
88 0 : jwt_token,
89 0 : max_offline_interval,
90 0 : max_warming_up_interval,
91 0 : cancel,
92 0 : );
93 0 : tokio::task::spawn(async move { heartbeater.run().await });
94 0 :
95 0 : Self { sender }
96 0 : }
97 :
98 0 : pub(crate) async fn heartbeat(
99 0 : &self,
100 0 : servers: Arc<HashMap<NodeId, Server>>,
101 0 : ) -> Result<AvailablityDeltas<State>, HeartbeaterError> {
102 0 : let (sender, receiver) = tokio::sync::oneshot::channel();
103 0 : self.sender
104 0 : .send(HeartbeatRequest {
105 0 : servers,
106 0 : reply: sender,
107 0 : })
108 0 : .map_err(|_| HeartbeaterError::Cancel)?;
109 :
110 0 : receiver
111 0 : .await
112 0 : .map_err(|_| HeartbeaterError::Cancel)
113 0 : .and_then(|x| x)
114 0 : }
115 : }
116 :
117 : impl<Server, State: Debug> HeartbeaterTask<Server, State>
118 : where
119 : HeartbeaterTask<Server, State>: HeartBeat<Server, State>,
120 : {
121 0 : fn new(
122 0 : receiver: tokio::sync::mpsc::UnboundedReceiver<HeartbeatRequest<Server, State>>,
123 0 : jwt_token: Option<String>,
124 0 : max_offline_interval: Duration,
125 0 : max_warming_up_interval: Duration,
126 0 : cancel: CancellationToken,
127 0 : ) -> Self {
128 0 : Self {
129 0 : receiver,
130 0 : cancel,
131 0 : state: HashMap::new(),
132 0 : max_offline_interval,
133 0 : max_warming_up_interval,
134 0 : jwt_token,
135 0 : }
136 0 : }
137 0 : async fn run(&mut self) {
138 : loop {
139 0 : tokio::select! {
140 0 : request = self.receiver.recv() => {
141 0 : match request {
142 0 : Some(req) => {
143 0 : if req.reply.is_closed() {
144 : // Prevent a possibly infinite buildup of the receiver channel, if requests arrive faster than we can handle them
145 0 : continue;
146 0 : }
147 0 : let res = self.heartbeat(req.servers).await;
148 : // Ignore the return value in order to not panic if the heartbeat function's future was cancelled
149 0 : _ = req.reply.send(res);
150 : },
151 0 : None => { return; }
152 : }
153 : },
154 0 : _ = self.cancel.cancelled() => return
155 : }
156 : }
157 0 : }
158 : }
159 :
160 : pub(crate) trait HeartBeat<Server, State> {
161 : fn heartbeat(
162 : &mut self,
163 : pageservers: Arc<HashMap<NodeId, Server>>,
164 : ) -> impl Future<Output = Result<AvailablityDeltas<State>, HeartbeaterError>> + Send;
165 : }
166 :
167 : impl HeartBeat<Node, PageserverState> for HeartbeaterTask<Node, PageserverState> {
168 0 : async fn heartbeat(
169 0 : &mut self,
170 0 : pageservers: Arc<HashMap<NodeId, Node>>,
171 0 : ) -> Result<AvailablityDeltas<PageserverState>, HeartbeaterError> {
172 0 : let mut new_state = HashMap::new();
173 0 :
174 0 : let mut heartbeat_futs = FuturesUnordered::new();
175 0 : for (node_id, node) in &*pageservers {
176 0 : heartbeat_futs.push({
177 0 : let jwt_token = self.jwt_token.clone();
178 0 : let cancel = self.cancel.clone();
179 0 :
180 0 : // Clone the node and mark it as available such that the request
181 0 : // goes through to the pageserver even when the node is marked offline.
182 0 : // This doesn't impact the availability observed by [`crate::service::Service`].
183 0 : let mut node_clone = node.clone();
184 0 : node_clone
185 0 : .set_availability(NodeAvailability::Active(PageserverUtilization::full()));
186 0 :
187 0 : async move {
188 0 : let response = node_clone
189 0 : .with_client_retries(
190 0 : |client| async move { client.get_utilization().await },
191 0 : &jwt_token,
192 0 : 3,
193 0 : 3,
194 0 : Duration::from_secs(1),
195 0 : &cancel,
196 0 : )
197 0 : .await;
198 :
199 0 : let response = match response {
200 0 : Some(r) => r,
201 : None => {
202 : // This indicates cancellation of the request.
203 : // We ignore the node in this case.
204 0 : return None;
205 : }
206 : };
207 :
208 0 : let status = if let Ok(utilization) = response {
209 0 : PageserverState::Available {
210 0 : last_seen_at: Instant::now(),
211 0 : utilization,
212 0 : }
213 0 : } else if let NodeAvailability::WarmingUp(last_seen_at) =
214 0 : node.get_availability()
215 : {
216 0 : PageserverState::WarmingUp {
217 0 : started_at: *last_seen_at,
218 0 : }
219 : } else {
220 0 : PageserverState::Offline
221 : };
222 :
223 0 : Some((*node_id, status))
224 0 : }
225 0 : });
226 :
227 : loop {
228 0 : let maybe_status = tokio::select! {
229 0 : next = heartbeat_futs.next() => {
230 0 : match next {
231 0 : Some(result) => result,
232 0 : None => { break; }
233 : }
234 : },
235 0 : _ = self.cancel.cancelled() => { return Err(HeartbeaterError::Cancel); }
236 : };
237 :
238 0 : if let Some((node_id, status)) = maybe_status {
239 0 : new_state.insert(node_id, status);
240 0 : }
241 : }
242 : }
243 :
244 0 : let mut warming_up = 0;
245 0 : let mut offline = 0;
246 0 : for state in new_state.values() {
247 0 : match state {
248 0 : PageserverState::WarmingUp { .. } => {
249 0 : warming_up += 1;
250 0 : }
251 0 : PageserverState::Offline { .. } => offline += 1,
252 0 : PageserverState::Available { .. } => {}
253 : }
254 : }
255 :
256 0 : tracing::info!(
257 0 : "Heartbeat round complete for {} nodes, {} warming-up, {} offline",
258 0 : new_state.len(),
259 : warming_up,
260 : offline
261 : );
262 :
263 0 : let mut deltas = Vec::new();
264 0 : let now = Instant::now();
265 0 : for (node_id, ps_state) in new_state.iter_mut() {
266 : use std::collections::hash_map::Entry::*;
267 0 : let entry = self.state.entry(*node_id);
268 0 :
269 0 : let mut needs_update = false;
270 0 : match entry {
271 0 : Occupied(ref occ) => match (occ.get(), &ps_state) {
272 0 : (PageserverState::Offline, PageserverState::Offline) => {}
273 0 : (PageserverState::Available { last_seen_at, .. }, PageserverState::Offline) => {
274 0 : if now - *last_seen_at >= self.max_offline_interval {
275 0 : deltas.push((*node_id, ps_state.clone()));
276 0 : needs_update = true;
277 0 : }
278 : }
279 0 : (_, PageserverState::WarmingUp { started_at }) => {
280 0 : if now - *started_at >= self.max_warming_up_interval {
281 0 : *ps_state = PageserverState::Offline;
282 0 : }
283 :
284 0 : deltas.push((*node_id, ps_state.clone()));
285 0 : needs_update = true;
286 : }
287 0 : _ => {
288 0 : deltas.push((*node_id, ps_state.clone()));
289 0 : needs_update = true;
290 0 : }
291 : },
292 0 : Vacant(_) => {
293 0 : // This is a new node. Don't generate a delta for it.
294 0 : deltas.push((*node_id, ps_state.clone()));
295 0 : }
296 : }
297 :
298 0 : match entry {
299 0 : Occupied(mut occ) if needs_update => {
300 0 : (*occ.get_mut()) = ps_state.clone();
301 0 : }
302 0 : Vacant(vac) => {
303 0 : vac.insert(ps_state.clone());
304 0 : }
305 0 : _ => {}
306 : }
307 : }
308 :
309 0 : Ok(AvailablityDeltas(deltas))
310 0 : }
311 : }
312 :
313 : impl HeartBeat<Safekeeper, SafekeeperState> for HeartbeaterTask<Safekeeper, SafekeeperState> {
314 0 : async fn heartbeat(
315 0 : &mut self,
316 0 : safekeepers: Arc<HashMap<NodeId, Safekeeper>>,
317 0 : ) -> Result<AvailablityDeltas<SafekeeperState>, HeartbeaterError> {
318 0 : let mut new_state = HashMap::new();
319 0 :
320 0 : let mut heartbeat_futs = FuturesUnordered::new();
321 0 : for (node_id, sk) in &*safekeepers {
322 0 : if sk.scheduling_policy() == SkSchedulingPolicy::Decomissioned {
323 0 : continue;
324 0 : }
325 0 : heartbeat_futs.push({
326 0 : let jwt_token = self
327 0 : .jwt_token
328 0 : .as_ref()
329 0 : .map(|t| SecretString::from(t.to_owned()));
330 0 : let cancel = self.cancel.clone();
331 0 :
332 0 : async move {
333 0 : let response = sk
334 0 : .with_client_retries(
335 0 : |client| async move { client.get_utilization().await },
336 0 : &jwt_token,
337 0 : 3,
338 0 : 3,
339 0 : Duration::from_secs(1),
340 0 : &cancel,
341 0 : )
342 0 : .await;
343 :
344 0 : let status = match response {
345 0 : Ok(utilization) => SafekeeperState::Available {
346 0 : last_seen_at: Instant::now(),
347 0 : utilization,
348 0 : },
349 : Err(mgmt_api::Error::Cancelled) => {
350 : // This indicates cancellation of the request.
351 : // We ignore the node in this case.
352 0 : return None;
353 : }
354 0 : Err(e) => {
355 0 : tracing::info!(
356 0 : "Marking safekeeper {} at as offline: {e}",
357 0 : sk.base_url()
358 : );
359 0 : SafekeeperState::Offline
360 : }
361 : };
362 :
363 0 : Some((*node_id, status))
364 0 : }
365 0 : });
366 :
367 : loop {
368 0 : let maybe_status = tokio::select! {
369 0 : next = heartbeat_futs.next() => {
370 0 : match next {
371 0 : Some(result) => result,
372 0 : None => { break; }
373 : }
374 : },
375 0 : _ = self.cancel.cancelled() => { return Err(HeartbeaterError::Cancel); }
376 : };
377 :
378 0 : if let Some((node_id, status)) = maybe_status {
379 0 : new_state.insert(node_id, status);
380 0 : }
381 : }
382 : }
383 :
384 0 : let mut offline = 0;
385 0 : for state in new_state.values() {
386 0 : match state {
387 0 : SafekeeperState::Offline { .. } => offline += 1,
388 0 : SafekeeperState::Available { .. } => {}
389 : }
390 : }
391 :
392 0 : tracing::info!(
393 0 : "Heartbeat round complete for {} safekeepers, {} offline",
394 0 : new_state.len(),
395 : offline
396 : );
397 :
398 0 : let mut deltas = Vec::new();
399 0 : let now = Instant::now();
400 0 : for (node_id, sk_state) in new_state.iter_mut() {
401 : use std::collections::hash_map::Entry::*;
402 0 : let entry = self.state.entry(*node_id);
403 0 :
404 0 : let mut needs_update = false;
405 0 : match entry {
406 0 : Occupied(ref occ) => match (occ.get(), &sk_state) {
407 0 : (SafekeeperState::Offline, SafekeeperState::Offline) => {}
408 0 : (SafekeeperState::Available { last_seen_at, .. }, SafekeeperState::Offline) => {
409 0 : if now - *last_seen_at >= self.max_offline_interval {
410 0 : deltas.push((*node_id, sk_state.clone()));
411 0 : needs_update = true;
412 0 : }
413 : }
414 0 : _ => {
415 0 : deltas.push((*node_id, sk_state.clone()));
416 0 : needs_update = true;
417 0 : }
418 : },
419 0 : Vacant(_) => {
420 0 : // This is a new node. Don't generate a delta for it.
421 0 : deltas.push((*node_id, sk_state.clone()));
422 0 : }
423 : }
424 :
425 0 : match entry {
426 0 : Occupied(mut occ) if needs_update => {
427 0 : (*occ.get_mut()) = sk_state.clone();
428 0 : }
429 0 : Vacant(vac) => {
430 0 : vac.insert(sk_state.clone());
431 0 : }
432 0 : _ => {}
433 : }
434 : }
435 :
436 0 : Ok(AvailablityDeltas(deltas))
437 0 : }
438 : }
|