Line data Source code
1 : use std::collections::HashSet;
2 : use std::str::FromStr;
3 : use std::sync::Arc;
4 : use std::time::Duration;
5 :
6 : use super::safekeeper_reconciler::ScheduleRequest;
7 : use crate::heartbeater::SafekeeperState;
8 : use crate::metrics;
9 : use crate::persistence::{
10 : DatabaseError, SafekeeperTimelineOpKind, TimelinePendingOpPersistence, TimelinePersistence,
11 : };
12 : use crate::safekeeper::Safekeeper;
13 : use crate::timeline_import::TimelineImportFinalizeError;
14 : use anyhow::Context;
15 : use http_utils::error::ApiError;
16 : use pageserver_api::controller_api::{
17 : SafekeeperDescribeResponse, SkSchedulingPolicy, TimelineImportRequest,
18 : };
19 : use pageserver_api::models::{SafekeeperInfo, SafekeepersInfo, TimelineInfo};
20 : use safekeeper_api::membership::{MemberSet, SafekeeperId};
21 : use tokio::task::JoinSet;
22 : use tokio_util::sync::CancellationToken;
23 : use utils::id::{NodeId, TenantId, TimelineId};
24 : use utils::logging::SecretString;
25 : use utils::lsn::Lsn;
26 :
27 : use super::Service;
28 :
29 : impl Service {
30 : /// Timeline creation on safekeepers
31 : ///
32 : /// Returns `Ok(left)` if the timeline has been created on a quorum of safekeepers,
33 : /// where `left` contains the list of safekeepers that didn't have a successful response.
34 : /// Assumes tenant lock is held while calling this function.
35 0 : pub(super) async fn tenant_timeline_create_safekeepers_quorum(
36 0 : &self,
37 0 : tenant_id: TenantId,
38 0 : timeline_id: TimelineId,
39 0 : pg_version: u32,
40 0 : timeline_persistence: &TimelinePersistence,
41 0 : ) -> Result<Vec<NodeId>, ApiError> {
42 0 : // If quorum is reached, return if we are outside of a specified timeout
43 0 : let jwt = self
44 0 : .config
45 0 : .safekeeper_jwt_token
46 0 : .clone()
47 0 : .map(SecretString::from);
48 0 : let mut joinset = JoinSet::new();
49 0 :
50 0 : // Prepare membership::Configuration from choosen safekeepers.
51 0 : let safekeepers = {
52 0 : let locked = self.inner.read().unwrap();
53 0 : locked.safekeepers.clone()
54 0 : };
55 0 :
56 0 : let mut members = Vec::new();
57 0 : for sk_id in timeline_persistence.sk_set.iter() {
58 0 : let sk_id = NodeId(*sk_id as u64);
59 0 : let Some(safekeeper) = safekeepers.get(&sk_id) else {
60 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
61 0 : "couldn't find entry for safekeeper with id {sk_id}"
62 0 : )))?;
63 : };
64 0 : members.push(SafekeeperId {
65 0 : id: sk_id,
66 0 : host: safekeeper.skp.host.clone(),
67 0 : pg_port: safekeeper.skp.port as u16,
68 0 : });
69 : }
70 0 : let mset = MemberSet::new(members).map_err(ApiError::InternalServerError)?;
71 0 : let mconf = safekeeper_api::membership::Configuration::new(mset);
72 0 :
73 0 : let req = safekeeper_api::models::TimelineCreateRequest {
74 0 : commit_lsn: None,
75 0 : mconf,
76 0 : pg_version,
77 0 : start_lsn: timeline_persistence.start_lsn.0,
78 0 : system_id: None,
79 0 : tenant_id,
80 0 : timeline_id,
81 0 : wal_seg_size: None,
82 0 : };
83 : const SK_CREATE_TIMELINE_RECONCILE_TIMEOUT: Duration = Duration::from_secs(30);
84 0 : for sk in timeline_persistence.sk_set.iter() {
85 0 : let sk_id = NodeId(*sk as u64);
86 0 : let safekeepers = safekeepers.clone();
87 0 : let http_client = self.http_client.clone();
88 0 : let jwt = jwt.clone();
89 0 : let req = req.clone();
90 0 : joinset.spawn(async move {
91 0 : // Unwrap is fine as we already would have returned error above
92 0 : let sk_p = safekeepers.get(&sk_id).unwrap();
93 0 : let res = sk_p
94 0 : .with_client_retries(
95 0 : |client| {
96 0 : let req = req.clone();
97 0 : async move { client.create_timeline(&req).await }
98 0 : },
99 0 : &http_client,
100 0 : &jwt,
101 0 : 3,
102 0 : 3,
103 0 : SK_CREATE_TIMELINE_RECONCILE_TIMEOUT,
104 0 : &CancellationToken::new(),
105 0 : )
106 0 : .await;
107 0 : (sk_id, sk_p.skp.host.clone(), res)
108 0 : });
109 0 : }
110 : // After we have built the joinset, we now wait for the tasks to complete,
111 : // but with a specified timeout to make sure we return swiftly, either with
112 : // a failure or success.
113 0 : let reconcile_deadline = tokio::time::Instant::now() + SK_CREATE_TIMELINE_RECONCILE_TIMEOUT;
114 0 :
115 0 : // Wait until all tasks finish or timeout is hit, whichever occurs
116 0 : // first.
117 0 : let mut reconcile_results = Vec::new();
118 : loop {
119 0 : if let Ok(res) = tokio::time::timeout_at(reconcile_deadline, joinset.join_next()).await
120 : {
121 0 : let Some(res) = res else { break };
122 0 : match res {
123 0 : Ok(res) => {
124 0 : tracing::info!(
125 0 : "response from safekeeper id:{} at {}: {:?}",
126 : res.0,
127 : res.1,
128 : res.2
129 : );
130 0 : reconcile_results.push(res);
131 : }
132 0 : Err(join_err) => {
133 0 : tracing::info!("join_err for task in joinset: {join_err}");
134 : }
135 : }
136 : } else {
137 0 : tracing::info!(
138 0 : "timeout for creation call after {} responses",
139 0 : reconcile_results.len()
140 : );
141 0 : break;
142 : }
143 : }
144 :
145 : // Now check now if quorum was reached in reconcile_results.
146 0 : let total_result_count = reconcile_results.len();
147 0 : let remaining = reconcile_results
148 0 : .into_iter()
149 0 : .filter_map(|res| res.2.is_err().then_some(res.0))
150 0 : .collect::<Vec<_>>();
151 0 : tracing::info!(
152 0 : "Got {} non-successful responses from initial creation request of total {total_result_count} responses",
153 0 : remaining.len()
154 : );
155 0 : let target_sk_count = timeline_persistence.sk_set.len();
156 0 : let quorum_size = match target_sk_count {
157 : 0 => {
158 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
159 0 : "timeline configured without any safekeepers",
160 0 : )));
161 : }
162 : 1 | 2 => {
163 : #[cfg(feature = "testing")]
164 : {
165 : // In test settings, it is allowed to have one or two safekeepers
166 0 : target_sk_count
167 : }
168 : #[cfg(not(feature = "testing"))]
169 : {
170 : // The region is misconfigured: we need at least three safekeepers to be configured
171 : // in order to schedule work to them
172 : tracing::warn!(
173 : "couldn't find at least 3 safekeepers for timeline, found: {:?}",
174 : timeline_persistence.sk_set
175 : );
176 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
177 : "couldn't find at least 3 safekeepers to put timeline to"
178 : )));
179 : }
180 : }
181 0 : _ => target_sk_count / 2 + 1,
182 : };
183 0 : let success_count = target_sk_count - remaining.len();
184 0 : if success_count < quorum_size {
185 : // Failure
186 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
187 0 : "not enough successful reconciliations to reach quorum size: {success_count} of {quorum_size} of total {target_sk_count}"
188 0 : )));
189 0 : }
190 0 :
191 0 : Ok(remaining)
192 0 : }
193 :
194 : /// Create timeline in controller database and on safekeepers.
195 : /// `timeline_info` is result of timeline creation on pageserver.
196 : ///
197 : /// All actions must be idempotent as the call is retried until success. It
198 : /// tries to create timeline in the db and on at least majority of
199 : /// safekeepers + queue creation for safekeepers which missed it in the db
200 : /// for infinite retries; after that, call returns Ok.
201 : ///
202 : /// The idea is that once this is reached as long as we have alive majority
203 : /// of safekeepers it is expected to get eventually operational as storcon
204 : /// will be able to seed timeline on nodes which missed creation by making
205 : /// pull_timeline from peers. On the other hand we don't want to fail
206 : /// timeline creation if one safekeeper is down.
207 0 : pub(super) async fn tenant_timeline_create_safekeepers(
208 0 : self: &Arc<Self>,
209 0 : tenant_id: TenantId,
210 0 : timeline_info: &TimelineInfo,
211 0 : read_only: bool,
212 0 : ) -> Result<SafekeepersInfo, ApiError> {
213 0 : let timeline_id = timeline_info.timeline_id;
214 0 : let pg_version = timeline_info.pg_version * 10000;
215 0 : // Initially start_lsn is determined by last_record_lsn in pageserver
216 0 : // response as it does initdb. However, later we persist it and in sk
217 0 : // creation calls replace with the value from the timeline row if it
218 0 : // previously existed as on retries in theory endpoint might have
219 0 : // already written some data and advanced last_record_lsn, while we want
220 0 : // safekeepers to have consistent start_lsn.
221 0 : let start_lsn = timeline_info.last_record_lsn;
222 :
223 : // Choose initial set of safekeepers respecting affinity
224 0 : let sks = if !read_only {
225 0 : self.safekeepers_for_new_timeline().await?
226 : } else {
227 0 : Vec::new()
228 : };
229 0 : let sks_persistence = sks.iter().map(|sk| sk.id.0 as i64).collect::<Vec<_>>();
230 0 : // Add timeline to db
231 0 : let mut timeline_persist = TimelinePersistence {
232 0 : tenant_id: tenant_id.to_string(),
233 0 : timeline_id: timeline_id.to_string(),
234 0 : start_lsn: start_lsn.into(),
235 0 : generation: 1,
236 0 : sk_set: sks_persistence.clone(),
237 0 : new_sk_set: None,
238 0 : cplane_notified_generation: 0,
239 0 : deleted_at: None,
240 0 : };
241 0 : let inserted = self
242 0 : .persistence
243 0 : .insert_timeline(timeline_persist.clone())
244 0 : .await?;
245 0 : if !inserted {
246 0 : if let Some(existent_persist) = self
247 0 : .persistence
248 0 : .get_timeline(tenant_id, timeline_id)
249 0 : .await?
250 0 : {
251 0 : // Replace with what we have in the db, to get stuff like the generation right.
252 0 : // We do still repeat the http calls to the safekeepers. After all, we could have
253 0 : // crashed right after the wrote to the DB.
254 0 : timeline_persist = existent_persist;
255 0 : } else {
256 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
257 0 : "insertion said timeline already in db, but looking it up, it was gone"
258 0 : )));
259 : }
260 0 : }
261 0 : let ret = SafekeepersInfo {
262 0 : generation: timeline_persist.generation as u32,
263 0 : safekeepers: sks.clone(),
264 0 : tenant_id,
265 0 : timeline_id,
266 0 : };
267 0 : if read_only {
268 0 : return Ok(ret);
269 0 : }
270 :
271 : // Create the timeline on a quorum of safekeepers
272 0 : let remaining = self
273 0 : .tenant_timeline_create_safekeepers_quorum(
274 0 : tenant_id,
275 0 : timeline_id,
276 0 : pg_version,
277 0 : &timeline_persist,
278 0 : )
279 0 : .await?;
280 :
281 : // For the remaining safekeepers, take care of their reconciliation asynchronously
282 0 : for &remaining_id in remaining.iter() {
283 0 : let pending_op = TimelinePendingOpPersistence {
284 0 : tenant_id: tenant_id.to_string(),
285 0 : timeline_id: timeline_id.to_string(),
286 0 : generation: timeline_persist.generation,
287 0 : op_kind: crate::persistence::SafekeeperTimelineOpKind::Pull,
288 0 : sk_id: remaining_id.0 as i64,
289 0 : };
290 0 : tracing::info!("writing pending op for sk id {remaining_id}");
291 0 : self.persistence.insert_pending_op(pending_op).await?;
292 : }
293 0 : if !remaining.is_empty() {
294 0 : let locked = self.inner.read().unwrap();
295 0 : for remaining_id in remaining {
296 0 : let Some(sk) = locked.safekeepers.get(&remaining_id) else {
297 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
298 0 : "Couldn't find safekeeper with id {remaining_id}"
299 0 : )));
300 : };
301 0 : let Ok(host_list) = sks
302 0 : .iter()
303 0 : .map(|sk| {
304 0 : Ok((
305 0 : sk.id,
306 0 : locked
307 0 : .safekeepers
308 0 : .get(&sk.id)
309 0 : .ok_or_else(|| {
310 0 : ApiError::InternalServerError(anyhow::anyhow!(
311 0 : "Couldn't find safekeeper with id {} to pull from",
312 0 : sk.id
313 0 : ))
314 0 : })?
315 0 : .base_url(),
316 : ))
317 0 : })
318 0 : .collect::<Result<_, ApiError>>()
319 : else {
320 0 : continue;
321 : };
322 0 : let req = ScheduleRequest {
323 0 : safekeeper: Box::new(sk.clone()),
324 0 : host_list,
325 0 : tenant_id,
326 0 : timeline_id: Some(timeline_id),
327 0 : generation: timeline_persist.generation as u32,
328 0 : kind: crate::persistence::SafekeeperTimelineOpKind::Pull,
329 0 : };
330 0 : locked.safekeeper_reconcilers.schedule_request(req);
331 : }
332 0 : }
333 :
334 0 : Ok(ret)
335 0 : }
336 :
337 0 : pub(crate) async fn tenant_timeline_create_safekeepers_until_success(
338 0 : self: &Arc<Self>,
339 0 : tenant_id: TenantId,
340 0 : timeline_info: TimelineInfo,
341 0 : ) -> Result<(), TimelineImportFinalizeError> {
342 : const BACKOFF: Duration = Duration::from_secs(5);
343 :
344 : loop {
345 0 : if self.cancel.is_cancelled() {
346 0 : return Err(TimelineImportFinalizeError::ShuttingDown);
347 0 : }
348 0 :
349 0 : // This function is only used in non-read-only scenarios
350 0 : let read_only = false;
351 0 : let res = self
352 0 : .tenant_timeline_create_safekeepers(tenant_id, &timeline_info, read_only)
353 0 : .await;
354 :
355 0 : match res {
356 : Ok(_) => {
357 0 : tracing::info!("Timeline created on safekeepers");
358 0 : break;
359 : }
360 0 : Err(err) => {
361 0 : tracing::error!("Failed to create timeline on safekeepers: {err}");
362 0 : tokio::select! {
363 0 : _ = self.cancel.cancelled() => {
364 0 : return Err(TimelineImportFinalizeError::ShuttingDown);
365 : },
366 0 : _ = tokio::time::sleep(BACKOFF) => {}
367 : };
368 : }
369 : }
370 : }
371 :
372 0 : Ok(())
373 0 : }
374 :
375 : /// Directly insert the timeline into the database without reconciling it with safekeepers.
376 : ///
377 : /// Useful if the timeline already exists on the specified safekeepers,
378 : /// but we want to make it storage controller managed.
379 0 : pub(crate) async fn timeline_import(&self, req: TimelineImportRequest) -> Result<(), ApiError> {
380 0 : let persistence = TimelinePersistence {
381 0 : tenant_id: req.tenant_id.to_string(),
382 0 : timeline_id: req.timeline_id.to_string(),
383 0 : start_lsn: Lsn::INVALID.into(),
384 0 : generation: 1,
385 0 : sk_set: req.sk_set.iter().map(|sk_id| sk_id.0 as i64).collect(),
386 0 : new_sk_set: None,
387 0 : cplane_notified_generation: 1,
388 0 : deleted_at: None,
389 0 : };
390 0 : let inserted = self.persistence.insert_timeline(persistence).await?;
391 0 : if inserted {
392 0 : tracing::info!("imported timeline into db");
393 : } else {
394 0 : tracing::info!("didn't import timeline into db, as it is already present in db");
395 : }
396 0 : Ok(())
397 0 : }
398 :
399 : /// Perform timeline deletion on safekeepers. Will return success: we persist the deletion into the reconciler.
400 0 : pub(super) async fn tenant_timeline_delete_safekeepers(
401 0 : self: &Arc<Self>,
402 0 : tenant_id: TenantId,
403 0 : timeline_id: TimelineId,
404 0 : ) -> Result<(), ApiError> {
405 0 : let tl = self
406 0 : .persistence
407 0 : .get_timeline(tenant_id, timeline_id)
408 0 : .await?;
409 0 : let Some(tl) = tl else {
410 0 : tracing::info!(
411 0 : "timeline {tenant_id}/{timeline_id} doesn't exist in timelines table, no deletions on safekeepers needed"
412 : );
413 0 : return Ok(());
414 : };
415 0 : self.persistence
416 0 : .timeline_set_deleted_at(tenant_id, timeline_id)
417 0 : .await?;
418 0 : let all_sks = tl
419 0 : .new_sk_set
420 0 : .iter()
421 0 : .flatten()
422 0 : .chain(tl.sk_set.iter())
423 0 : .collect::<HashSet<_>>();
424 0 :
425 0 : // The timeline has no safekeepers: we need to delete it from the db manually,
426 0 : // as no safekeeper reconciler will get to it
427 0 : if all_sks.is_empty() {
428 0 : if let Err(err) = self
429 0 : .persistence
430 0 : .delete_timeline(tenant_id, timeline_id)
431 0 : .await
432 : {
433 0 : tracing::warn!(%tenant_id, %timeline_id, "couldn't delete timeline from db: {err}");
434 0 : }
435 0 : }
436 :
437 : // Schedule reconciliations
438 0 : for &sk_id in all_sks.iter() {
439 0 : let pending_op = TimelinePendingOpPersistence {
440 0 : tenant_id: tenant_id.to_string(),
441 0 : timeline_id: timeline_id.to_string(),
442 0 : generation: i32::MAX,
443 0 : op_kind: SafekeeperTimelineOpKind::Delete,
444 0 : sk_id: *sk_id,
445 0 : };
446 0 : tracing::info!("writing pending op for sk id {sk_id}");
447 0 : self.persistence.insert_pending_op(pending_op).await?;
448 : }
449 : {
450 0 : let locked = self.inner.read().unwrap();
451 0 : for sk_id in all_sks {
452 0 : let sk_id = NodeId(*sk_id as u64);
453 0 : let Some(sk) = locked.safekeepers.get(&sk_id) else {
454 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
455 0 : "Couldn't find safekeeper with id {sk_id}"
456 0 : )));
457 : };
458 :
459 0 : let req = ScheduleRequest {
460 0 : safekeeper: Box::new(sk.clone()),
461 0 : // we don't use this for this kind, put a dummy value
462 0 : host_list: Vec::new(),
463 0 : tenant_id,
464 0 : timeline_id: Some(timeline_id),
465 0 : generation: tl.generation as u32,
466 0 : kind: SafekeeperTimelineOpKind::Delete,
467 0 : };
468 0 : locked.safekeeper_reconcilers.schedule_request(req);
469 : }
470 : }
471 0 : Ok(())
472 0 : }
473 :
474 : /// Perform tenant deletion on safekeepers.
475 0 : pub(super) async fn tenant_delete_safekeepers(
476 0 : self: &Arc<Self>,
477 0 : tenant_id: TenantId,
478 0 : ) -> Result<(), ApiError> {
479 0 : let timeline_list = self
480 0 : .persistence
481 0 : .list_timelines_for_tenant(tenant_id)
482 0 : .await?;
483 :
484 0 : if timeline_list.is_empty() {
485 : // Early exit: the tenant is either empty or not migrated to the storcon yet
486 0 : tracing::info!("Skipping tenant delete as the timeline doesn't exist in db");
487 0 : return Ok(());
488 0 : }
489 :
490 0 : let timeline_list = timeline_list
491 0 : .into_iter()
492 0 : .map(|timeline| {
493 0 : let timeline_id = TimelineId::from_str(&timeline.timeline_id)
494 0 : .context("timeline id loaded from db")
495 0 : .map_err(ApiError::InternalServerError)?;
496 0 : Ok((timeline_id, timeline))
497 0 : })
498 0 : .collect::<Result<Vec<_>, ApiError>>()?;
499 :
500 : // Remove pending ops from db, and set `deleted_at`.
501 : // We cancel them in a later iteration once we hold the state lock.
502 0 : for (timeline_id, _timeline) in timeline_list.iter() {
503 0 : self.persistence
504 0 : .remove_pending_ops_for_timeline(tenant_id, Some(*timeline_id))
505 0 : .await?;
506 0 : self.persistence
507 0 : .timeline_set_deleted_at(tenant_id, *timeline_id)
508 0 : .await?;
509 : }
510 :
511 : // The list of safekeepers that have any of the timelines
512 0 : let mut sk_list = HashSet::new();
513 :
514 : // List all pending ops for all timelines, cancel them
515 0 : for (_timeline_id, timeline) in timeline_list.iter() {
516 0 : let sk_iter = timeline
517 0 : .sk_set
518 0 : .iter()
519 0 : .chain(timeline.new_sk_set.iter().flatten())
520 0 : .map(|id| NodeId(*id as u64));
521 0 : sk_list.extend(sk_iter);
522 0 : }
523 :
524 0 : for &sk_id in sk_list.iter() {
525 0 : let pending_op = TimelinePendingOpPersistence {
526 0 : tenant_id: tenant_id.to_string(),
527 0 : timeline_id: String::new(),
528 0 : generation: i32::MAX,
529 0 : op_kind: SafekeeperTimelineOpKind::Delete,
530 0 : sk_id: sk_id.0 as i64,
531 0 : };
532 0 : tracing::info!("writing pending op for sk id {sk_id}");
533 0 : self.persistence.insert_pending_op(pending_op).await?;
534 : }
535 :
536 0 : let mut locked = self.inner.write().unwrap();
537 :
538 0 : for (timeline_id, _timeline) in timeline_list.iter() {
539 0 : for sk_id in sk_list.iter() {
540 0 : locked
541 0 : .safekeeper_reconcilers
542 0 : .cancel_reconciles_for_timeline(*sk_id, tenant_id, Some(*timeline_id));
543 0 : }
544 : }
545 :
546 : // unwrap is safe: we return above for an empty timeline list
547 0 : let max_generation = timeline_list
548 0 : .iter()
549 0 : .map(|(_tl_id, tl)| tl.generation as u32)
550 0 : .max()
551 0 : .unwrap();
552 :
553 0 : for sk_id in sk_list {
554 0 : let Some(safekeeper) = locked.safekeepers.get(&sk_id) else {
555 0 : tracing::warn!("Couldn't find safekeeper with id {sk_id}");
556 0 : continue;
557 : };
558 : // Add pending op for tenant deletion
559 0 : let req = ScheduleRequest {
560 0 : generation: max_generation,
561 0 : host_list: Vec::new(),
562 0 : kind: SafekeeperTimelineOpKind::Delete,
563 0 : safekeeper: Box::new(safekeeper.clone()),
564 0 : tenant_id,
565 0 : timeline_id: None,
566 0 : };
567 0 : locked.safekeeper_reconcilers.schedule_request(req);
568 : }
569 0 : Ok(())
570 0 : }
571 :
572 : /// Choose safekeepers for the new timeline: 3 in different azs.
573 0 : pub(crate) async fn safekeepers_for_new_timeline(
574 0 : &self,
575 0 : ) -> Result<Vec<SafekeeperInfo>, ApiError> {
576 0 : let mut all_safekeepers = {
577 0 : let locked = self.inner.read().unwrap();
578 0 : locked
579 0 : .safekeepers
580 0 : .iter()
581 0 : .filter_map(|sk| {
582 0 : if sk.1.scheduling_policy() != SkSchedulingPolicy::Active {
583 : // If we don't want to schedule stuff onto the safekeeper, respect that.
584 0 : return None;
585 0 : }
586 0 : let utilization_opt = if let SafekeeperState::Available {
587 : last_seen_at: _,
588 0 : utilization,
589 0 : } = sk.1.availability()
590 : {
591 0 : Some(utilization)
592 : } else {
593 : // non-available safekeepers still get a chance for new timelines,
594 : // but put them last in the list.
595 0 : None
596 : };
597 0 : let info = SafekeeperInfo {
598 0 : hostname: sk.1.skp.host.clone(),
599 0 : id: NodeId(sk.1.skp.id as u64),
600 0 : };
601 0 : Some((utilization_opt, info, sk.1.skp.availability_zone_id.clone()))
602 0 : })
603 0 : .collect::<Vec<_>>()
604 0 : };
605 0 : all_safekeepers.sort_by_key(|sk| {
606 0 : (
607 0 : sk.0.as_ref()
608 0 : .map(|ut| ut.timeline_count)
609 0 : .unwrap_or(u64::MAX),
610 0 : // Use the id to decide on equal scores for reliability
611 0 : sk.1.id.0,
612 0 : )
613 0 : });
614 : // Number of safekeepers in different AZs we are looking for
615 0 : let wanted_count = match all_safekeepers.len() {
616 : 0 => {
617 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
618 0 : "couldn't find any active safekeeper for new timeline",
619 0 : )));
620 : }
621 : // Have laxer requirements on testig mode as we don't want to
622 : // spin up three safekeepers for every single test
623 : #[cfg(feature = "testing")]
624 0 : 1 | 2 => all_safekeepers.len(),
625 0 : _ => 3,
626 : };
627 0 : let mut sks = Vec::new();
628 0 : let mut azs = HashSet::new();
629 0 : for (_sk_util, sk_info, az_id) in all_safekeepers.iter() {
630 0 : if !azs.insert(az_id) {
631 0 : continue;
632 0 : }
633 0 : sks.push(sk_info.clone());
634 0 : if sks.len() == wanted_count {
635 0 : break;
636 0 : }
637 : }
638 0 : if sks.len() == wanted_count {
639 0 : Ok(sks)
640 : } else {
641 0 : Err(ApiError::InternalServerError(anyhow::anyhow!(
642 0 : "couldn't find {wanted_count} safekeepers in different AZs for new timeline (found: {}, total active: {})",
643 0 : sks.len(),
644 0 : all_safekeepers.len(),
645 0 : )))
646 : }
647 0 : }
648 :
649 0 : pub(crate) async fn safekeepers_list(
650 0 : &self,
651 0 : ) -> Result<Vec<SafekeeperDescribeResponse>, DatabaseError> {
652 0 : let locked = self.inner.read().unwrap();
653 0 : let mut list = locked
654 0 : .safekeepers
655 0 : .iter()
656 0 : .map(|sk| sk.1.describe_response())
657 0 : .collect::<Result<Vec<_>, _>>()?;
658 0 : list.sort_by_key(|v| v.id);
659 0 : Ok(list)
660 0 : }
661 :
662 0 : pub(crate) async fn get_safekeeper(
663 0 : &self,
664 0 : id: i64,
665 0 : ) -> Result<SafekeeperDescribeResponse, DatabaseError> {
666 0 : let locked = self.inner.read().unwrap();
667 0 : let sk = locked
668 0 : .safekeepers
669 0 : .get(&NodeId(id as u64))
670 0 : .ok_or(diesel::result::Error::NotFound)?;
671 0 : sk.describe_response()
672 0 : }
673 :
674 0 : pub(crate) async fn upsert_safekeeper(
675 0 : self: &Arc<Service>,
676 0 : record: crate::persistence::SafekeeperUpsert,
677 0 : ) -> Result<(), ApiError> {
678 0 : let node_id = NodeId(record.id as u64);
679 0 : let use_https = self.config.use_https_safekeeper_api;
680 0 :
681 0 : if use_https && record.https_port.is_none() {
682 0 : return Err(ApiError::PreconditionFailed(
683 0 : format!(
684 0 : "cannot upsert safekeeper {node_id}: \
685 0 : https is enabled, but https port is not specified"
686 0 : )
687 0 : .into(),
688 0 : ));
689 0 : }
690 0 :
691 0 : self.persistence.safekeeper_upsert(record.clone()).await?;
692 : {
693 0 : let mut locked = self.inner.write().unwrap();
694 0 : let mut safekeepers = (*locked.safekeepers).clone();
695 0 : match safekeepers.entry(node_id) {
696 0 : std::collections::hash_map::Entry::Occupied(mut entry) => entry
697 0 : .get_mut()
698 0 : .update_from_record(record)
699 0 : .expect("all preconditions should be checked before upsert to database"),
700 0 : std::collections::hash_map::Entry::Vacant(entry) => {
701 0 : entry.insert(
702 0 : Safekeeper::from_persistence(
703 0 : crate::persistence::SafekeeperPersistence::from_upsert(
704 0 : record,
705 0 : SkSchedulingPolicy::Pause,
706 0 : ),
707 0 : CancellationToken::new(),
708 0 : use_https,
709 0 : )
710 0 : .expect("all preconditions should be checked before upsert to database"),
711 0 : );
712 0 : }
713 : }
714 0 : locked
715 0 : .safekeeper_reconcilers
716 0 : .start_reconciler(node_id, self);
717 0 : locked.safekeepers = Arc::new(safekeepers);
718 0 : metrics::METRICS_REGISTRY
719 0 : .metrics_group
720 0 : .storage_controller_safekeeper_nodes
721 0 : .set(locked.safekeepers.len() as i64);
722 0 : metrics::METRICS_REGISTRY
723 0 : .metrics_group
724 0 : .storage_controller_https_safekeeper_nodes
725 0 : .set(
726 0 : locked
727 0 : .safekeepers
728 0 : .values()
729 0 : .filter(|s| s.has_https_port())
730 0 : .count() as i64,
731 0 : );
732 0 : }
733 0 : Ok(())
734 0 : }
735 :
736 0 : pub(crate) async fn set_safekeeper_scheduling_policy(
737 0 : self: &Arc<Service>,
738 0 : id: i64,
739 0 : scheduling_policy: SkSchedulingPolicy,
740 0 : ) -> Result<(), DatabaseError> {
741 0 : self.persistence
742 0 : .set_safekeeper_scheduling_policy(id, scheduling_policy)
743 0 : .await?;
744 0 : let node_id = NodeId(id as u64);
745 0 : // After the change has been persisted successfully, update the in-memory state
746 0 : {
747 0 : let mut locked = self.inner.write().unwrap();
748 0 : let mut safekeepers = (*locked.safekeepers).clone();
749 0 : let sk = safekeepers
750 0 : .get_mut(&node_id)
751 0 : .ok_or(DatabaseError::Logical("Not found".to_string()))?;
752 0 : sk.set_scheduling_policy(scheduling_policy);
753 0 :
754 0 : match scheduling_policy {
755 0 : SkSchedulingPolicy::Active => {
756 0 : locked
757 0 : .safekeeper_reconcilers
758 0 : .start_reconciler(node_id, self);
759 0 : }
760 0 : SkSchedulingPolicy::Decomissioned | SkSchedulingPolicy::Pause => {
761 0 : locked.safekeeper_reconcilers.stop_reconciler(node_id);
762 0 : }
763 : }
764 :
765 0 : locked.safekeepers = Arc::new(safekeepers);
766 0 : }
767 0 : Ok(())
768 0 : }
769 : }
|
