Line data Source code
1 : use std::{
2 : borrow::Cow,
3 : cmp::Ordering,
4 : collections::{BTreeMap, HashMap, HashSet},
5 : ops::Deref,
6 : path::PathBuf,
7 : str::FromStr,
8 : sync::Arc,
9 : time::{Duration, Instant},
10 : };
11 :
12 : use crate::{
13 : background_node_operations::{
14 : Drain, Fill, Operation, OperationError, OperationHandler, MAX_RECONCILES_PER_OPERATION,
15 : },
16 : compute_hook::NotifyError,
17 : id_lock_map::{trace_exclusive_lock, trace_shared_lock, IdLockMap, TracingExclusiveGuard},
18 : metrics::LeadershipStatusGroup,
19 : persistence::{AbortShardSplitStatus, MetadataHealthPersistence, TenantFilter},
20 : reconciler::{ReconcileError, ReconcileUnits},
21 : scheduler::{MaySchedule, ScheduleContext, ScheduleMode},
22 : tenant_shard::{
23 : MigrateAttachment, ReconcileNeeded, ReconcilerStatus, ScheduleOptimization,
24 : ScheduleOptimizationAction,
25 : },
26 : };
27 : use anyhow::Context;
28 : use control_plane::storage_controller::{
29 : AttachHookRequest, AttachHookResponse, InspectRequest, InspectResponse,
30 : };
31 : use diesel::result::DatabaseErrorKind;
32 : use futures::{stream::FuturesUnordered, StreamExt};
33 : use itertools::Itertools;
34 : use pageserver_api::{
35 : controller_api::{
36 : MetadataHealthRecord, MetadataHealthUpdateRequest, NodeAvailability, NodeRegisterRequest,
37 : NodeSchedulingPolicy, PlacementPolicy, ShardSchedulingPolicy, TenantCreateRequest,
38 : TenantCreateResponse, TenantCreateResponseShard, TenantDescribeResponse,
39 : TenantDescribeResponseShard, TenantLocateResponse, TenantPolicyRequest,
40 : TenantShardMigrateRequest, TenantShardMigrateResponse, UtilizationScore,
41 : },
42 : models::{SecondaryProgress, TenantConfigRequest, TopTenantShardsRequest},
43 : };
44 : use reqwest::StatusCode;
45 : use tracing::{instrument, Instrument};
46 :
47 : use crate::pageserver_client::PageserverClient;
48 : use pageserver_api::{
49 : models::{
50 : self, LocationConfig, LocationConfigListResponse, LocationConfigMode,
51 : PageserverUtilization, ShardParameters, TenantConfig, TenantLocationConfigRequest,
52 : TenantLocationConfigResponse, TenantShardLocation, TenantShardSplitRequest,
53 : TenantShardSplitResponse, TenantTimeTravelRequest, TimelineCreateRequest, TimelineInfo,
54 : },
55 : shard::{ShardCount, ShardIdentity, ShardNumber, ShardStripeSize, TenantShardId},
56 : upcall_api::{
57 : ReAttachRequest, ReAttachResponse, ReAttachResponseTenant, ValidateRequest,
58 : ValidateResponse, ValidateResponseTenant,
59 : },
60 : };
61 : use pageserver_client::mgmt_api;
62 : use tokio::sync::mpsc::error::TrySendError;
63 : use tokio_util::sync::CancellationToken;
64 : use utils::{
65 : completion::Barrier,
66 : failpoint_support,
67 : generation::Generation,
68 : http::error::ApiError,
69 : id::{NodeId, TenantId, TimelineId},
70 : sync::gate::Gate,
71 : };
72 :
73 : use crate::{
74 : compute_hook::ComputeHook,
75 : heartbeater::{Heartbeater, PageserverState},
76 : node::{AvailabilityTransition, Node},
77 : persistence::{split_state::SplitState, DatabaseError, Persistence, TenantShardPersistence},
78 : reconciler::attached_location_conf,
79 : scheduler::Scheduler,
80 : tenant_shard::{
81 : IntentState, ObservedState, ObservedStateLocation, ReconcileResult, ReconcileWaitError,
82 : ReconcilerWaiter, TenantShard,
83 : },
84 : };
85 : use serde::{Deserialize, Serialize};
86 :
87 : pub mod chaos_injector;
88 :
89 : // For operations that should be quick, like attaching a new tenant
90 : const SHORT_RECONCILE_TIMEOUT: Duration = Duration::from_secs(5);
91 :
92 : // For operations that might be slow, like migrating a tenant with
93 : // some data in it.
94 : pub const RECONCILE_TIMEOUT: Duration = Duration::from_secs(30);
95 :
96 : // If we receive a call using Secondary mode initially, it will omit generation. We will initialize
97 : // tenant shards into this generation, and as long as it remains in this generation, we will accept
98 : // input generation from future requests as authoritative.
99 : const INITIAL_GENERATION: Generation = Generation::new(0);
100 :
101 : /// How long [`Service::startup_reconcile`] is allowed to take before it should give
102 : /// up on unresponsive pageservers and proceed.
103 : pub(crate) const STARTUP_RECONCILE_TIMEOUT: Duration = Duration::from_secs(30);
104 :
105 : /// How long a node may be unresponsive to heartbeats before we declare it offline.
106 : /// This must be long enough to cover node restarts as well as normal operations: in future
107 : pub const MAX_OFFLINE_INTERVAL_DEFAULT: Duration = Duration::from_secs(30);
108 :
109 : /// How long a node may be unresponsive to heartbeats during start up before we declare it
110 : /// offline. This is much more lenient than [`MAX_OFFLINE_INTERVAL_DEFAULT`] since the pageserver's
111 : /// handling of the re-attach response may take a long time and blocks heartbeats from
112 : /// being handled on the pageserver side.
113 : pub const MAX_WARMING_UP_INTERVAL_DEFAULT: Duration = Duration::from_secs(300);
114 :
115 0 : #[derive(Clone, strum_macros::Display)]
116 : enum TenantOperations {
117 : Create,
118 : LocationConfig,
119 : ConfigSet,
120 : TimeTravelRemoteStorage,
121 : Delete,
122 : UpdatePolicy,
123 : ShardSplit,
124 : SecondaryDownload,
125 : TimelineCreate,
126 : TimelineDelete,
127 : AttachHook,
128 : TimelineDetachAncestor,
129 : }
130 :
131 0 : #[derive(Clone, strum_macros::Display)]
132 : enum NodeOperations {
133 : Register,
134 : Configure,
135 : Delete,
136 : }
137 :
138 : /// The leadership status for the storage controller process.
139 : /// Allowed transitions are:
140 : /// 1. Leader -> SteppedDown
141 : /// 2. Candidate -> Leader
142 0 : #[derive(Copy, Clone, strum_macros::Display, measured::FixedCardinalityLabel)]
143 : #[strum(serialize_all = "snake_case")]
144 : pub(crate) enum LeadershipStatus {
145 : /// This is the steady state where the storage controller can produce
146 : /// side effects in the cluster.
147 : Leader,
148 : /// We've been notified to step down by another candidate. No reconciliations
149 : /// take place in this state.
150 : SteppedDown,
151 : /// Initial state for a new storage controller instance. Will attempt to assume leadership.
152 : #[allow(unused)]
153 : Candidate,
154 : }
155 :
156 : pub const RECONCILER_CONCURRENCY_DEFAULT: usize = 128;
157 :
158 : // Depth of the channel used to enqueue shards for reconciliation when they can't do it immediately.
159 : // This channel is finite-size to avoid using excessive memory if we get into a state where reconciles are finishing more slowly
160 : // than they're being pushed onto the queue.
161 : const MAX_DELAYED_RECONCILES: usize = 10000;
162 :
163 : // Top level state available to all HTTP handlers
164 : struct ServiceState {
165 : leadership_status: LeadershipStatus,
166 :
167 : tenants: BTreeMap<TenantShardId, TenantShard>,
168 :
169 : nodes: Arc<HashMap<NodeId, Node>>,
170 :
171 : scheduler: Scheduler,
172 :
173 : /// Ongoing background operation on the cluster if any is running.
174 : /// Note that only one such operation may run at any given time,
175 : /// hence the type choice.
176 : ongoing_operation: Option<OperationHandler>,
177 :
178 : /// Queue of tenants who are waiting for concurrency limits to permit them to reconcile
179 : delayed_reconcile_rx: tokio::sync::mpsc::Receiver<TenantShardId>,
180 : }
181 :
182 : /// Transform an error from a pageserver into an error to return to callers of a storage
183 : /// controller API.
184 0 : fn passthrough_api_error(node: &Node, e: mgmt_api::Error) -> ApiError {
185 0 : match e {
186 0 : mgmt_api::Error::SendRequest(e) => {
187 0 : // Presume errors sending requests are connectivity/availability issues
188 0 : ApiError::ResourceUnavailable(format!("{node} error sending request: {e}").into())
189 : }
190 0 : mgmt_api::Error::ReceiveErrorBody(str) => {
191 0 : // Presume errors receiving body are connectivity/availability issues
192 0 : ApiError::ResourceUnavailable(
193 0 : format!("{node} error receiving error body: {str}").into(),
194 0 : )
195 : }
196 0 : mgmt_api::Error::ReceiveBody(str) => {
197 0 : // Presume errors receiving body are connectivity/availability issues
198 0 : ApiError::ResourceUnavailable(format!("{node} error receiving body: {str}").into())
199 : }
200 0 : mgmt_api::Error::ApiError(StatusCode::NOT_FOUND, msg) => {
201 0 : ApiError::NotFound(anyhow::anyhow!(format!("{node}: {msg}")).into())
202 : }
203 0 : mgmt_api::Error::ApiError(StatusCode::SERVICE_UNAVAILABLE, msg) => {
204 0 : ApiError::ResourceUnavailable(format!("{node}: {msg}").into())
205 : }
206 0 : mgmt_api::Error::ApiError(status @ StatusCode::UNAUTHORIZED, msg)
207 0 : | mgmt_api::Error::ApiError(status @ StatusCode::FORBIDDEN, msg) => {
208 : // Auth errors talking to a pageserver are not auth errors for the caller: they are
209 : // internal server errors, showing that something is wrong with the pageserver or
210 : // storage controller's auth configuration.
211 0 : ApiError::InternalServerError(anyhow::anyhow!("{node} {status}: {msg}"))
212 : }
213 0 : mgmt_api::Error::ApiError(status, msg) => {
214 0 : // Presume general case of pageserver API errors is that we tried to do something
215 0 : // that can't be done right now.
216 0 : ApiError::Conflict(format!("{node} {status}: {status} {msg}"))
217 : }
218 0 : mgmt_api::Error::Cancelled => ApiError::ShuttingDown,
219 : }
220 0 : }
221 :
222 : impl ServiceState {
223 0 : fn new(
224 0 : nodes: HashMap<NodeId, Node>,
225 0 : tenants: BTreeMap<TenantShardId, TenantShard>,
226 0 : scheduler: Scheduler,
227 0 : delayed_reconcile_rx: tokio::sync::mpsc::Receiver<TenantShardId>,
228 0 : ) -> Self {
229 0 : let status = &crate::metrics::METRICS_REGISTRY
230 0 : .metrics_group
231 0 : .storage_controller_leadership_status;
232 0 :
233 0 : status.set(
234 0 : LeadershipStatusGroup {
235 0 : status: LeadershipStatus::Leader,
236 0 : },
237 0 : 1,
238 0 : );
239 0 :
240 0 : Self {
241 0 : // TODO: Starting up as Leader is a transient state. Once we enable rolling
242 0 : // upgrades on the k8s side, we should start up as Candidate.
243 0 : leadership_status: LeadershipStatus::Leader,
244 0 : tenants,
245 0 : nodes: Arc::new(nodes),
246 0 : scheduler,
247 0 : ongoing_operation: None,
248 0 : delayed_reconcile_rx,
249 0 : }
250 0 : }
251 :
252 0 : fn parts_mut(
253 0 : &mut self,
254 0 : ) -> (
255 0 : &mut Arc<HashMap<NodeId, Node>>,
256 0 : &mut BTreeMap<TenantShardId, TenantShard>,
257 0 : &mut Scheduler,
258 0 : ) {
259 0 : (&mut self.nodes, &mut self.tenants, &mut self.scheduler)
260 0 : }
261 :
262 0 : fn get_leadership_status(&self) -> LeadershipStatus {
263 0 : self.leadership_status
264 0 : }
265 :
266 0 : fn step_down(&mut self) {
267 0 : self.leadership_status = LeadershipStatus::SteppedDown;
268 0 :
269 0 : let status = &crate::metrics::METRICS_REGISTRY
270 0 : .metrics_group
271 0 : .storage_controller_leadership_status;
272 0 :
273 0 : status.set(
274 0 : LeadershipStatusGroup {
275 0 : status: LeadershipStatus::SteppedDown,
276 0 : },
277 0 : 1,
278 0 : );
279 0 : status.set(
280 0 : LeadershipStatusGroup {
281 0 : status: LeadershipStatus::Leader,
282 0 : },
283 0 : 0,
284 0 : );
285 0 : status.set(
286 0 : LeadershipStatusGroup {
287 0 : status: LeadershipStatus::Candidate,
288 0 : },
289 0 : 0,
290 0 : );
291 0 : }
292 : }
293 :
294 : #[derive(Clone)]
295 : pub struct Config {
296 : // All pageservers managed by one instance of this service must have
297 : // the same public key. This JWT token will be used to authenticate
298 : // this service to the pageservers it manages.
299 : pub jwt_token: Option<String>,
300 :
301 : // This JWT token will be used to authenticate this service to the control plane.
302 : pub control_plane_jwt_token: Option<String>,
303 :
304 : /// Where the compute hook should send notifications of pageserver attachment locations
305 : /// (this URL points to the control plane in prod). If this is None, the compute hook will
306 : /// assume it is running in a test environment and try to update neon_local.
307 : pub compute_hook_url: Option<String>,
308 :
309 : /// Grace period within which a pageserver does not respond to heartbeats, but is still
310 : /// considered active. Once the grace period elapses, the next heartbeat failure will
311 : /// mark the pagseserver offline.
312 : pub max_offline_interval: Duration,
313 :
314 : /// Extended grace period within which pageserver may not respond to heartbeats.
315 : /// This extended grace period kicks in after the node has been drained for restart
316 : /// and/or upon handling the re-attach request from a node.
317 : pub max_warming_up_interval: Duration,
318 :
319 : /// How many Reconcilers may be spawned concurrently
320 : pub reconciler_concurrency: usize,
321 :
322 : /// How large must a shard grow in bytes before we split it?
323 : /// None disables auto-splitting.
324 : pub split_threshold: Option<u64>,
325 :
326 : // TODO: make this cfg(feature = "testing")
327 : pub neon_local_repo_dir: Option<PathBuf>,
328 : }
329 :
330 : impl From<DatabaseError> for ApiError {
331 0 : fn from(err: DatabaseError) -> ApiError {
332 0 : match err {
333 0 : DatabaseError::Query(e) => ApiError::InternalServerError(e.into()),
334 : // FIXME: ApiError doesn't have an Unavailable variant, but ShuttingDown maps to 503.
335 : DatabaseError::Connection(_) | DatabaseError::ConnectionPool(_) => {
336 0 : ApiError::ShuttingDown
337 : }
338 0 : DatabaseError::Logical(reason) => {
339 0 : ApiError::InternalServerError(anyhow::anyhow!(reason))
340 : }
341 : }
342 0 : }
343 : }
344 :
345 : pub struct Service {
346 : inner: Arc<std::sync::RwLock<ServiceState>>,
347 : config: Config,
348 : persistence: Arc<Persistence>,
349 : compute_hook: Arc<ComputeHook>,
350 : result_tx: tokio::sync::mpsc::UnboundedSender<ReconcileResultRequest>,
351 :
352 : heartbeater: Heartbeater,
353 :
354 : // Channel for background cleanup from failed operations that require cleanup, such as shard split
355 : abort_tx: tokio::sync::mpsc::UnboundedSender<TenantShardSplitAbort>,
356 :
357 : // Locking on a tenant granularity (covers all shards in the tenant):
358 : // - Take exclusively for rare operations that mutate the tenant's persistent state (e.g. create/delete/split)
359 : // - Take in shared mode for operations that need the set of shards to stay the same to complete reliably (e.g. timeline CRUD)
360 : tenant_op_locks: IdLockMap<TenantId, TenantOperations>,
361 :
362 : // Locking for node-mutating operations: take exclusively for operations that modify the node's persistent state, or
363 : // that transition it to/from Active.
364 : node_op_locks: IdLockMap<NodeId, NodeOperations>,
365 :
366 : // Limit how many Reconcilers we will spawn concurrently
367 : reconciler_concurrency: Arc<tokio::sync::Semaphore>,
368 :
369 : /// Queue of tenants who are waiting for concurrency limits to permit them to reconcile
370 : /// Send into this queue to promptly attempt to reconcile this shard next time units are available.
371 : ///
372 : /// Note that this state logically lives inside ServiceInner, but carrying Sender here makes the code simpler
373 : /// by avoiding needing a &mut ref to something inside the ServiceInner. This could be optimized to
374 : /// use a VecDeque instead of a channel to reduce synchronization overhead, at the cost of some code complexity.
375 : delayed_reconcile_tx: tokio::sync::mpsc::Sender<TenantShardId>,
376 :
377 : // Process shutdown will fire this token
378 : cancel: CancellationToken,
379 :
380 : // Child token of [`Service::cancel`] used by reconcilers
381 : reconcilers_cancel: CancellationToken,
382 :
383 : // Background tasks will hold this gate
384 : gate: Gate,
385 :
386 : // Reconcilers background tasks will hold this gate
387 : reconcilers_gate: Gate,
388 :
389 : /// This waits for initial reconciliation with pageservers to complete. Until this barrier
390 : /// passes, it isn't safe to do any actions that mutate tenants.
391 : pub(crate) startup_complete: Barrier,
392 : }
393 :
394 : impl From<ReconcileWaitError> for ApiError {
395 0 : fn from(value: ReconcileWaitError) -> Self {
396 0 : match value {
397 0 : ReconcileWaitError::Shutdown => ApiError::ShuttingDown,
398 0 : e @ ReconcileWaitError::Timeout(_) => ApiError::Timeout(format!("{e}").into()),
399 0 : e @ ReconcileWaitError::Failed(..) => ApiError::InternalServerError(anyhow::anyhow!(e)),
400 : }
401 0 : }
402 : }
403 :
404 : impl From<OperationError> for ApiError {
405 0 : fn from(value: OperationError) -> Self {
406 0 : match value {
407 0 : OperationError::NodeStateChanged(err) | OperationError::FinalizeError(err) => {
408 0 : ApiError::InternalServerError(anyhow::anyhow!(err))
409 : }
410 0 : OperationError::Cancelled => ApiError::Conflict("Operation was cancelled".into()),
411 : }
412 0 : }
413 : }
414 :
415 : #[allow(clippy::large_enum_variant)]
416 : enum TenantCreateOrUpdate {
417 : Create(TenantCreateRequest),
418 : Update(Vec<ShardUpdate>),
419 : }
420 :
421 : struct ShardSplitParams {
422 : old_shard_count: ShardCount,
423 : new_shard_count: ShardCount,
424 : new_stripe_size: Option<ShardStripeSize>,
425 : targets: Vec<ShardSplitTarget>,
426 : policy: PlacementPolicy,
427 : config: TenantConfig,
428 : shard_ident: ShardIdentity,
429 : }
430 :
431 : // When preparing for a shard split, we may either choose to proceed with the split,
432 : // or find that the work is already done and return NoOp.
433 : enum ShardSplitAction {
434 : Split(ShardSplitParams),
435 : NoOp(TenantShardSplitResponse),
436 : }
437 :
438 : // A parent shard which will be split
439 : struct ShardSplitTarget {
440 : parent_id: TenantShardId,
441 : node: Node,
442 : child_ids: Vec<TenantShardId>,
443 : }
444 :
445 : /// When we tenant shard split operation fails, we may not be able to clean up immediately, because nodes
446 : /// might not be available. We therefore use a queue of abort operations processed in the background.
447 : struct TenantShardSplitAbort {
448 : tenant_id: TenantId,
449 : /// The target values from the request that failed
450 : new_shard_count: ShardCount,
451 : new_stripe_size: Option<ShardStripeSize>,
452 : /// Until this abort op is complete, no other operations may be done on the tenant
453 : _tenant_lock: TracingExclusiveGuard<TenantOperations>,
454 : }
455 :
456 0 : #[derive(thiserror::Error, Debug)]
457 : enum TenantShardSplitAbortError {
458 : #[error(transparent)]
459 : Database(#[from] DatabaseError),
460 : #[error(transparent)]
461 : Remote(#[from] mgmt_api::Error),
462 : #[error("Unavailable")]
463 : Unavailable,
464 : }
465 :
466 : struct ShardUpdate {
467 : tenant_shard_id: TenantShardId,
468 : placement_policy: PlacementPolicy,
469 : tenant_config: TenantConfig,
470 :
471 : /// If this is None, generation is not updated.
472 : generation: Option<Generation>,
473 : }
474 :
475 : enum StopReconciliationsReason {
476 : ShuttingDown,
477 : SteppingDown,
478 : }
479 :
480 : impl std::fmt::Display for StopReconciliationsReason {
481 0 : fn fmt(&self, writer: &mut std::fmt::Formatter) -> std::fmt::Result {
482 0 : let s = match self {
483 0 : Self::ShuttingDown => "Shutting down",
484 0 : Self::SteppingDown => "Stepping down",
485 : };
486 0 : write!(writer, "{}", s)
487 0 : }
488 : }
489 :
490 : pub(crate) enum ReconcileResultRequest {
491 : ReconcileResult(ReconcileResult),
492 : Stop,
493 : }
494 :
495 : // TODO: move this into the storcon peer client when that gets added
496 0 : #[derive(Serialize, Deserialize, Debug, Default)]
497 : pub(crate) struct GlobalObservedState(HashMap<TenantShardId, ObservedState>);
498 :
499 : impl Service {
500 0 : pub fn get_config(&self) -> &Config {
501 0 : &self.config
502 0 : }
503 :
504 : /// Called once on startup, this function attempts to contact all pageservers to build an up-to-date
505 : /// view of the world, and determine which pageservers are responsive.
506 0 : #[instrument(skip_all)]
507 : async fn startup_reconcile(
508 : self: &Arc<Service>,
509 : bg_compute_notify_result_tx: tokio::sync::mpsc::Sender<
510 : Result<(), (TenantShardId, NotifyError)>,
511 : >,
512 : ) {
513 : // For all tenant shards, a vector of observed states on nodes (where None means
514 : // indeterminate, same as in [`ObservedStateLocation`])
515 : let mut observed: HashMap<TenantShardId, Vec<(NodeId, Option<LocationConfig>)>> =
516 : HashMap::new();
517 :
518 : // Startup reconciliation does I/O to other services: whether they
519 : // are responsive or not, we should aim to finish within our deadline, because:
520 : // - If we don't, a k8s readiness hook watching /ready will kill us.
521 : // - While we're waiting for startup reconciliation, we are not fully
522 : // available for end user operations like creating/deleting tenants and timelines.
523 : //
524 : // We set multiple deadlines to break up the time available between the phases of work: this is
525 : // arbitrary, but avoids a situation where the first phase could burn our entire timeout period.
526 : let start_at = Instant::now();
527 : let node_scan_deadline = start_at
528 : .checked_add(STARTUP_RECONCILE_TIMEOUT / 2)
529 : .expect("Reconcile timeout is a modest constant");
530 :
531 : // Accumulate a list of any tenant locations that ought to be detached
532 : let mut cleanup = Vec::new();
533 :
534 : let node_listings = self.scan_node_locations(node_scan_deadline).await;
535 : // Send initial heartbeat requests to nodes that replied to the location listing above.
536 : let nodes_online = self.initial_heartbeat_round(node_listings.keys()).await;
537 :
538 : for (node_id, list_response) in node_listings {
539 : let tenant_shards = list_response.tenant_shards;
540 : tracing::info!(
541 : "Received {} shard statuses from pageserver {}, setting it to Active",
542 : tenant_shards.len(),
543 : node_id
544 : );
545 :
546 : for (tenant_shard_id, conf_opt) in tenant_shards {
547 : let shard_observations = observed.entry(tenant_shard_id).or_default();
548 : shard_observations.push((node_id, conf_opt));
549 : }
550 : }
551 :
552 : // List of tenants for which we will attempt to notify compute of their location at startup
553 : let mut compute_notifications = Vec::new();
554 :
555 : // Populate intent and observed states for all tenants, based on reported state on pageservers
556 : tracing::info!("Populating tenant shards' states from initial pageserver scan...");
557 : let shard_count = {
558 : let mut locked = self.inner.write().unwrap();
559 : let (nodes, tenants, scheduler) = locked.parts_mut();
560 :
561 : // Mark nodes online if they responded to us: nodes are offline by default after a restart.
562 : let mut new_nodes = (**nodes).clone();
563 : for (node_id, node) in new_nodes.iter_mut() {
564 : if let Some(utilization) = nodes_online.get(node_id) {
565 : node.set_availability(NodeAvailability::Active(UtilizationScore(
566 : utilization.utilization_score,
567 : )));
568 : scheduler.node_upsert(node);
569 : }
570 : }
571 : *nodes = Arc::new(new_nodes);
572 :
573 : for (tenant_shard_id, shard_observations) in observed {
574 : for (node_id, observed_loc) in shard_observations {
575 : let Some(tenant_shard) = tenants.get_mut(&tenant_shard_id) else {
576 : cleanup.push((tenant_shard_id, node_id));
577 : continue;
578 : };
579 : tenant_shard
580 : .observed
581 : .locations
582 : .insert(node_id, ObservedStateLocation { conf: observed_loc });
583 : }
584 : }
585 :
586 : // Populate each tenant's intent state
587 : let mut schedule_context = ScheduleContext::default();
588 : for (tenant_shard_id, tenant_shard) in tenants.iter_mut() {
589 : if tenant_shard_id.shard_number == ShardNumber(0) {
590 : // Reset scheduling context each time we advance to the next Tenant
591 : schedule_context = ScheduleContext::default();
592 : }
593 :
594 : tenant_shard.intent_from_observed(scheduler);
595 : if let Err(e) = tenant_shard.schedule(scheduler, &mut schedule_context) {
596 : // Non-fatal error: we are unable to properly schedule the tenant, perhaps because
597 : // not enough pageservers are available. The tenant may well still be available
598 : // to clients.
599 : tracing::error!("Failed to schedule tenant {tenant_shard_id} at startup: {e}");
600 : } else {
601 : // If we're both intending and observed to be attached at a particular node, we will
602 : // emit a compute notification for this. In the case where our observed state does not
603 : // yet match our intent, we will eventually reconcile, and that will emit a compute notification.
604 : if let Some(attached_at) = tenant_shard.stably_attached() {
605 : compute_notifications.push((
606 : *tenant_shard_id,
607 : attached_at,
608 : tenant_shard.shard.stripe_size,
609 : ));
610 : }
611 : }
612 : }
613 :
614 : tenants.len()
615 : };
616 :
617 : // TODO: if any tenant's intent now differs from its loaded generation_pageserver, we should clear that
618 : // generation_pageserver in the database.
619 :
620 : // Emit compute hook notifications for all tenants which are already stably attached. Other tenants
621 : // will emit compute hook notifications when they reconcile.
622 : //
623 : // Ordering: our calls to notify_background synchronously establish a relative order for these notifications vs. any later
624 : // calls into the ComputeHook for the same tenant: we can leave these to run to completion in the background and any later
625 : // calls will be correctly ordered wrt these.
626 : //
627 : // Concurrency: we call notify_background for all tenants, which will create O(N) tokio tasks, but almost all of them
628 : // will just wait on the ComputeHook::API_CONCURRENCY semaphore immediately, so very cheap until they get that semaphore
629 : // unit and start doing I/O.
630 : tracing::info!(
631 : "Sending {} compute notifications",
632 : compute_notifications.len()
633 : );
634 : self.compute_hook.notify_background(
635 : compute_notifications,
636 : bg_compute_notify_result_tx.clone(),
637 : &self.cancel,
638 : );
639 :
640 : // Finally, now that the service is up and running, launch reconcile operations for any tenants
641 : // which require it: under normal circumstances this should only include tenants that were in some
642 : // transient state before we restarted, or any tenants whose compute hooks failed above.
643 : tracing::info!("Checking for shards in need of reconciliation...");
644 : let reconcile_tasks = self.reconcile_all();
645 : // We will not wait for these reconciliation tasks to run here: we're now done with startup and
646 : // normal operations may proceed.
647 :
648 : // Clean up any tenants that were found on pageservers but are not known to us. Do this in the
649 : // background because it does not need to complete in order to proceed with other work.
650 : if !cleanup.is_empty() {
651 : tracing::info!("Cleaning up {} locations in the background", cleanup.len());
652 : tokio::task::spawn({
653 : let cleanup_self = self.clone();
654 0 : async move { cleanup_self.cleanup_locations(cleanup).await }
655 : });
656 : }
657 :
658 : tracing::info!("Startup complete, spawned {reconcile_tasks} reconciliation tasks ({shard_count} shards total)");
659 : }
660 :
661 0 : async fn initial_heartbeat_round<'a>(
662 0 : &self,
663 0 : node_ids: impl Iterator<Item = &'a NodeId>,
664 0 : ) -> HashMap<NodeId, PageserverUtilization> {
665 0 : assert!(!self.startup_complete.is_ready());
666 :
667 0 : let all_nodes = {
668 0 : let locked = self.inner.read().unwrap();
669 0 : locked.nodes.clone()
670 0 : };
671 0 :
672 0 : let mut nodes_to_heartbeat = HashMap::new();
673 0 : for node_id in node_ids {
674 0 : match all_nodes.get(node_id) {
675 0 : Some(node) => {
676 0 : nodes_to_heartbeat.insert(*node_id, node.clone());
677 0 : }
678 : None => {
679 0 : tracing::warn!("Node {node_id} was removed during start-up");
680 : }
681 : }
682 : }
683 :
684 0 : tracing::info!("Sending initial heartbeats...");
685 0 : let res = self
686 0 : .heartbeater
687 0 : .heartbeat(Arc::new(nodes_to_heartbeat))
688 0 : .await;
689 :
690 0 : let mut online_nodes = HashMap::new();
691 0 : if let Ok(deltas) = res {
692 0 : for (node_id, status) in deltas.0 {
693 0 : match status {
694 0 : PageserverState::Available { utilization, .. } => {
695 0 : online_nodes.insert(node_id, utilization);
696 0 : }
697 0 : PageserverState::Offline => {}
698 : PageserverState::WarmingUp { .. } => {
699 0 : unreachable!("Nodes are never marked warming-up during startup reconcile")
700 : }
701 : }
702 : }
703 0 : }
704 :
705 0 : online_nodes
706 0 : }
707 :
708 : /// Used during [`Self::startup_reconcile`]: issue GETs to all nodes concurrently, with a deadline.
709 : ///
710 : /// The result includes only nodes which responded within the deadline
711 0 : async fn scan_node_locations(
712 0 : &self,
713 0 : deadline: Instant,
714 0 : ) -> HashMap<NodeId, LocationConfigListResponse> {
715 0 : let nodes = {
716 0 : let locked = self.inner.read().unwrap();
717 0 : locked.nodes.clone()
718 0 : };
719 0 :
720 0 : let mut node_results = HashMap::new();
721 0 :
722 0 : let mut node_list_futs = FuturesUnordered::new();
723 0 :
724 0 : tracing::info!("Scanning shards on {} nodes...", nodes.len());
725 0 : for node in nodes.values() {
726 0 : node_list_futs.push({
727 0 : async move {
728 0 : tracing::info!("Scanning shards on node {node}...");
729 0 : let timeout = Duration::from_secs(1);
730 0 : let response = node
731 0 : .with_client_retries(
732 0 : |client| async move { client.list_location_config().await },
733 0 : &self.config.jwt_token,
734 0 : 1,
735 0 : 5,
736 0 : timeout,
737 0 : &self.cancel,
738 0 : )
739 0 : .await;
740 0 : (node.get_id(), response)
741 0 : }
742 0 : });
743 0 : }
744 :
745 : loop {
746 0 : let (node_id, result) = tokio::select! {
747 : next = node_list_futs.next() => {
748 : match next {
749 : Some(result) => result,
750 : None =>{
751 : // We got results for all our nodes
752 : break;
753 : }
754 :
755 : }
756 : },
757 : _ = tokio::time::sleep(deadline.duration_since(Instant::now())) => {
758 : // Give up waiting for anyone who hasn't responded: we will yield the results that we have
759 : tracing::info!("Reached deadline while waiting for nodes to respond to location listing requests");
760 : break;
761 : }
762 : };
763 :
764 0 : let Some(list_response) = result else {
765 0 : tracing::info!("Shutdown during startup_reconcile");
766 0 : break;
767 : };
768 :
769 0 : match list_response {
770 0 : Err(e) => {
771 0 : tracing::warn!("Could not scan node {} ({e})", node_id);
772 : }
773 0 : Ok(listing) => {
774 0 : node_results.insert(node_id, listing);
775 0 : }
776 : }
777 : }
778 :
779 0 : node_results
780 0 : }
781 :
782 : /// Used during [`Self::startup_reconcile`]: detach a list of unknown-to-us tenants from pageservers.
783 : ///
784 : /// This is safe to run in the background, because if we don't have this TenantShardId in our map of
785 : /// tenants, then it is probably something incompletely deleted before: we will not fight with any
786 : /// other task trying to attach it.
787 0 : #[instrument(skip_all)]
788 : async fn cleanup_locations(&self, cleanup: Vec<(TenantShardId, NodeId)>) {
789 : let nodes = self.inner.read().unwrap().nodes.clone();
790 :
791 : for (tenant_shard_id, node_id) in cleanup {
792 : // A node reported a tenant_shard_id which is unknown to us: detach it.
793 : let Some(node) = nodes.get(&node_id) else {
794 : // This is legitimate; we run in the background and [`Self::startup_reconcile`] might have identified
795 : // a location to clean up on a node that has since been removed.
796 : tracing::info!(
797 : "Not cleaning up location {node_id}/{tenant_shard_id}: node not found"
798 : );
799 : continue;
800 : };
801 :
802 : if self.cancel.is_cancelled() {
803 : break;
804 : }
805 :
806 : let client = PageserverClient::new(
807 : node.get_id(),
808 : node.base_url(),
809 : self.config.jwt_token.as_deref(),
810 : );
811 : match client
812 : .location_config(
813 : tenant_shard_id,
814 : LocationConfig {
815 : mode: LocationConfigMode::Detached,
816 : generation: None,
817 : secondary_conf: None,
818 : shard_number: tenant_shard_id.shard_number.0,
819 : shard_count: tenant_shard_id.shard_count.literal(),
820 : shard_stripe_size: 0,
821 : tenant_conf: models::TenantConfig::default(),
822 : },
823 : None,
824 : false,
825 : )
826 : .await
827 : {
828 : Ok(()) => {
829 : tracing::info!(
830 : "Detached unknown shard {tenant_shard_id} on pageserver {node_id}"
831 : );
832 : }
833 : Err(e) => {
834 : // Non-fatal error: leaving a tenant shard behind that we are not managing shouldn't
835 : // break anything.
836 : tracing::error!(
837 : "Failed to detach unknkown shard {tenant_shard_id} on pageserver {node_id}: {e}"
838 : );
839 : }
840 : }
841 : }
842 : }
843 :
844 : /// Long running background task that periodically wakes up and looks for shards that need
845 : /// reconciliation. Reconciliation is fallible, so any reconciliation tasks that fail during
846 : /// e.g. a tenant create/attach/migrate must eventually be retried: this task is responsible
847 : /// for those retries.
848 0 : #[instrument(skip_all)]
849 : async fn background_reconcile(self: &Arc<Self>) {
850 : self.startup_complete.clone().wait().await;
851 :
852 : const BACKGROUND_RECONCILE_PERIOD: Duration = Duration::from_secs(20);
853 :
854 : let mut interval = tokio::time::interval(BACKGROUND_RECONCILE_PERIOD);
855 : while !self.reconcilers_cancel.is_cancelled() {
856 : tokio::select! {
857 : _ = interval.tick() => {
858 : let reconciles_spawned = self.reconcile_all();
859 : if reconciles_spawned == 0 {
860 : // Run optimizer only when we didn't find any other work to do
861 : let optimizations = self.optimize_all().await;
862 : if optimizations == 0 {
863 : // Run new splits only when no optimizations are pending
864 : self.autosplit_tenants().await;
865 : }
866 : }
867 : }
868 : _ = self.reconcilers_cancel.cancelled() => return
869 : }
870 : }
871 : }
872 0 : #[instrument(skip_all)]
873 : async fn spawn_heartbeat_driver(&self) {
874 : self.startup_complete.clone().wait().await;
875 :
876 : const HEARTBEAT_INTERVAL: Duration = Duration::from_secs(5);
877 :
878 : let mut interval = tokio::time::interval(HEARTBEAT_INTERVAL);
879 : while !self.cancel.is_cancelled() {
880 : tokio::select! {
881 : _ = interval.tick() => { }
882 : _ = self.cancel.cancelled() => return
883 : };
884 :
885 : let nodes = {
886 : let locked = self.inner.read().unwrap();
887 : locked.nodes.clone()
888 : };
889 :
890 : let res = self.heartbeater.heartbeat(nodes).await;
891 : if let Ok(deltas) = res {
892 : for (node_id, state) in deltas.0 {
893 : let new_availability = match state {
894 : PageserverState::Available { utilization, .. } => NodeAvailability::Active(
895 : UtilizationScore(utilization.utilization_score),
896 : ),
897 : PageserverState::WarmingUp { started_at } => {
898 : NodeAvailability::WarmingUp(started_at)
899 : }
900 : PageserverState::Offline => {
901 : // The node might have been placed in the WarmingUp state
902 : // while the heartbeat round was on-going. Hence, filter out
903 : // offline transitions for WarmingUp nodes that are still within
904 : // their grace period.
905 : if let Ok(NodeAvailability::WarmingUp(started_at)) =
906 0 : self.get_node(node_id).await.map(|n| n.get_availability())
907 : {
908 : let now = Instant::now();
909 : if now - started_at >= self.config.max_warming_up_interval {
910 : NodeAvailability::Offline
911 : } else {
912 : NodeAvailability::WarmingUp(started_at)
913 : }
914 : } else {
915 : NodeAvailability::Offline
916 : }
917 : }
918 : };
919 :
920 : // This is the code path for geniune availability transitions (i.e node
921 : // goes unavailable and/or comes back online).
922 : let res = self
923 : .node_configure(node_id, Some(new_availability), None)
924 : .await;
925 :
926 : match res {
927 : Ok(()) => {}
928 : Err(ApiError::NotFound(_)) => {
929 : // This should be rare, but legitimate since the heartbeats are done
930 : // on a snapshot of the nodes.
931 : tracing::info!("Node {} was not found after heartbeat round", node_id);
932 : }
933 : Err(err) => {
934 : // Transition to active involves reconciling: if a node responds to a heartbeat then
935 : // becomes unavailable again, we may get an error here.
936 : tracing::error!(
937 : "Failed to update node {} after heartbeat round: {}",
938 : node_id,
939 : err
940 : );
941 : }
942 : }
943 : }
944 : }
945 : }
946 : }
947 :
948 : /// Apply the contents of a [`ReconcileResult`] to our in-memory state: if the reconciliation
949 : /// was successful and intent hasn't changed since the Reconciler was spawned, this will update
950 : /// the observed state of the tenant such that subsequent calls to [`TenantShard::get_reconcile_needed`]
951 : /// will indicate that reconciliation is not needed.
952 0 : #[instrument(skip_all, fields(
953 : tenant_id=%result.tenant_shard_id.tenant_id, shard_id=%result.tenant_shard_id.shard_slug(),
954 : sequence=%result.sequence
955 0 : ))]
956 : fn process_result(&self, mut result: ReconcileResult) {
957 : let mut locked = self.inner.write().unwrap();
958 : let (nodes, tenants, _scheduler) = locked.parts_mut();
959 : let Some(tenant) = tenants.get_mut(&result.tenant_shard_id) else {
960 : // A reconciliation result might race with removing a tenant: drop results for
961 : // tenants that aren't in our map.
962 : return;
963 : };
964 :
965 : // Usually generation should only be updated via this path, so the max() isn't
966 : // needed, but it is used to handle out-of-band updates via. e.g. test hook.
967 : tenant.generation = std::cmp::max(tenant.generation, result.generation);
968 :
969 : // If the reconciler signals that it failed to notify compute, set this state on
970 : // the shard so that a future [`TenantShard::maybe_reconcile`] will try again.
971 : tenant.pending_compute_notification = result.pending_compute_notification;
972 :
973 : // Let the TenantShard know it is idle.
974 : tenant.reconcile_complete(result.sequence);
975 :
976 : // In case a node was deleted while this reconcile is in flight, filter it out of the update we will
977 : // make to the tenant
978 : result
979 : .observed
980 : .locations
981 0 : .retain(|node_id, _loc| nodes.contains_key(node_id));
982 :
983 : match result.result {
984 : Ok(()) => {
985 : for (node_id, loc) in &result.observed.locations {
986 : if let Some(conf) = &loc.conf {
987 : tracing::info!("Updating observed location {}: {:?}", node_id, conf);
988 : } else {
989 : tracing::info!("Setting observed location {} to None", node_id,)
990 : }
991 : }
992 :
993 : tenant.observed = result.observed;
994 : tenant.waiter.advance(result.sequence);
995 : }
996 : Err(e) => {
997 : match e {
998 : ReconcileError::Cancel => {
999 : tracing::info!("Reconciler was cancelled");
1000 : }
1001 : ReconcileError::Remote(mgmt_api::Error::Cancelled) => {
1002 : // This might be due to the reconciler getting cancelled, or it might
1003 : // be due to the `Node` being marked offline.
1004 : tracing::info!("Reconciler cancelled during pageserver API call");
1005 : }
1006 : _ => {
1007 : tracing::warn!("Reconcile error: {}", e);
1008 : }
1009 : }
1010 :
1011 : // Ordering: populate last_error before advancing error_seq,
1012 : // so that waiters will see the correct error after waiting.
1013 : tenant.set_last_error(result.sequence, e);
1014 :
1015 : for (node_id, o) in result.observed.locations {
1016 : tenant.observed.locations.insert(node_id, o);
1017 : }
1018 : }
1019 : }
1020 :
1021 : // Maybe some other work can proceed now that this job finished.
1022 : if self.reconciler_concurrency.available_permits() > 0 {
1023 : while let Ok(tenant_shard_id) = locked.delayed_reconcile_rx.try_recv() {
1024 : let (nodes, tenants, _scheduler) = locked.parts_mut();
1025 : if let Some(shard) = tenants.get_mut(&tenant_shard_id) {
1026 : shard.delayed_reconcile = false;
1027 : self.maybe_reconcile_shard(shard, nodes);
1028 : }
1029 :
1030 : if self.reconciler_concurrency.available_permits() == 0 {
1031 : break;
1032 : }
1033 : }
1034 : }
1035 : }
1036 :
1037 0 : async fn process_results(
1038 0 : &self,
1039 0 : mut result_rx: tokio::sync::mpsc::UnboundedReceiver<ReconcileResultRequest>,
1040 0 : mut bg_compute_hook_result_rx: tokio::sync::mpsc::Receiver<
1041 0 : Result<(), (TenantShardId, NotifyError)>,
1042 0 : >,
1043 0 : ) {
1044 0 : loop {
1045 0 : // Wait for the next result, or for cancellation
1046 0 : tokio::select! {
1047 : r = result_rx.recv() => {
1048 : match r {
1049 : Some(ReconcileResultRequest::ReconcileResult(result)) => {self.process_result(result);},
1050 : None | Some(ReconcileResultRequest::Stop) => {break;}
1051 : }
1052 : }
1053 0 : _ = async{
1054 0 : match bg_compute_hook_result_rx.recv().await {
1055 0 : Some(result) => {
1056 0 : if let Err((tenant_shard_id, notify_error)) = result {
1057 0 : tracing::warn!("Marking shard {tenant_shard_id} for notification retry, due to error {notify_error}");
1058 0 : let mut locked = self.inner.write().unwrap();
1059 0 : if let Some(shard) = locked.tenants.get_mut(&tenant_shard_id) {
1060 0 : shard.pending_compute_notification = true;
1061 0 : }
1062 :
1063 0 : }
1064 : },
1065 : None => {
1066 : // This channel is dead, but we don't want to terminate the outer loop{}: just wait for shutdown
1067 0 : self.cancel.cancelled().await;
1068 : }
1069 : }
1070 0 : } => {},
1071 : _ = self.cancel.cancelled() => {
1072 : break;
1073 : }
1074 0 : };
1075 0 : }
1076 0 : }
1077 :
1078 0 : async fn process_aborts(
1079 0 : &self,
1080 0 : mut abort_rx: tokio::sync::mpsc::UnboundedReceiver<TenantShardSplitAbort>,
1081 0 : ) {
1082 : loop {
1083 : // Wait for the next result, or for cancellation
1084 0 : let op = tokio::select! {
1085 : r = abort_rx.recv() => {
1086 : match r {
1087 : Some(op) => {op},
1088 : None => {break;}
1089 : }
1090 : }
1091 : _ = self.cancel.cancelled() => {
1092 : break;
1093 : }
1094 : };
1095 :
1096 : // Retry until shutdown: we must keep this request object alive until it is properly
1097 : // processed, as it holds a lock guard that prevents other operations trying to do things
1098 : // to the tenant while it is in a weird part-split state.
1099 0 : while !self.cancel.is_cancelled() {
1100 0 : match self.abort_tenant_shard_split(&op).await {
1101 0 : Ok(_) => break,
1102 0 : Err(e) => {
1103 0 : tracing::warn!(
1104 0 : "Failed to abort shard split on {}, will retry: {e}",
1105 : op.tenant_id
1106 : );
1107 :
1108 : // If a node is unavailable, we hope that it has been properly marked Offline
1109 : // when we retry, so that the abort op will succeed. If the abort op is failing
1110 : // for some other reason, we will keep retrying forever, or until a human notices
1111 : // and does something about it (either fixing a pageserver or restarting the controller).
1112 0 : tokio::time::timeout(Duration::from_secs(5), self.cancel.cancelled())
1113 0 : .await
1114 0 : .ok();
1115 : }
1116 : }
1117 : }
1118 : }
1119 0 : }
1120 :
1121 0 : pub async fn spawn(config: Config, persistence: Arc<Persistence>) -> anyhow::Result<Arc<Self>> {
1122 0 : let (result_tx, result_rx) = tokio::sync::mpsc::unbounded_channel();
1123 0 : let (abort_tx, abort_rx) = tokio::sync::mpsc::unbounded_channel();
1124 0 :
1125 0 : tracing::info!("Loading nodes from database...");
1126 0 : let nodes = persistence
1127 0 : .list_nodes()
1128 0 : .await?
1129 0 : .into_iter()
1130 0 : .map(Node::from_persistent)
1131 0 : .collect::<Vec<_>>();
1132 0 : let nodes: HashMap<NodeId, Node> = nodes.into_iter().map(|n| (n.get_id(), n)).collect();
1133 0 : tracing::info!("Loaded {} nodes from database.", nodes.len());
1134 :
1135 0 : tracing::info!("Loading shards from database...");
1136 0 : let mut tenant_shard_persistence = persistence.list_tenant_shards().await?;
1137 0 : tracing::info!(
1138 0 : "Loaded {} shards from database.",
1139 0 : tenant_shard_persistence.len()
1140 : );
1141 :
1142 : // If any shard splits were in progress, reset the database state to abort them
1143 0 : let mut tenant_shard_count_min_max: HashMap<TenantId, (ShardCount, ShardCount)> =
1144 0 : HashMap::new();
1145 0 : for tsp in &mut tenant_shard_persistence {
1146 0 : let shard = tsp.get_shard_identity()?;
1147 0 : let tenant_shard_id = tsp.get_tenant_shard_id()?;
1148 0 : let entry = tenant_shard_count_min_max
1149 0 : .entry(tenant_shard_id.tenant_id)
1150 0 : .or_insert_with(|| (shard.count, shard.count));
1151 0 : entry.0 = std::cmp::min(entry.0, shard.count);
1152 0 : entry.1 = std::cmp::max(entry.1, shard.count);
1153 0 : }
1154 :
1155 0 : for (tenant_id, (count_min, count_max)) in tenant_shard_count_min_max {
1156 0 : if count_min != count_max {
1157 : // Aborting the split in the database and dropping the child shards is sufficient: the reconciliation in
1158 : // [`Self::startup_reconcile`] will implicitly drop the child shards on remote pageservers, or they'll
1159 : // be dropped later in [`Self::node_activate_reconcile`] if it isn't available right now.
1160 0 : tracing::info!("Aborting shard split {tenant_id} {count_min:?} -> {count_max:?}");
1161 0 : let abort_status = persistence.abort_shard_split(tenant_id, count_max).await?;
1162 :
1163 : // We may never see the Complete status here: if the split was complete, we wouldn't have
1164 : // identified this tenant has having mismatching min/max counts.
1165 0 : assert!(matches!(abort_status, AbortShardSplitStatus::Aborted));
1166 :
1167 : // Clear the splitting status in-memory, to reflect that we just aborted in the database
1168 0 : tenant_shard_persistence.iter_mut().for_each(|tsp| {
1169 0 : // Set idle split state on those shards that we will retain.
1170 0 : let tsp_tenant_id = TenantId::from_str(tsp.tenant_id.as_str()).unwrap();
1171 0 : if tsp_tenant_id == tenant_id
1172 0 : && tsp.get_shard_identity().unwrap().count == count_min
1173 0 : {
1174 0 : tsp.splitting = SplitState::Idle;
1175 0 : } else if tsp_tenant_id == tenant_id {
1176 : // Leave the splitting state on the child shards: this will be used next to
1177 : // drop them.
1178 0 : tracing::info!(
1179 0 : "Shard {tsp_tenant_id} will be dropped after shard split abort",
1180 : );
1181 0 : }
1182 0 : });
1183 0 :
1184 0 : // Drop shards for this tenant which we didn't just mark idle (i.e. child shards of the aborted split)
1185 0 : tenant_shard_persistence.retain(|tsp| {
1186 0 : TenantId::from_str(tsp.tenant_id.as_str()).unwrap() != tenant_id
1187 0 : || tsp.splitting == SplitState::Idle
1188 0 : });
1189 0 : }
1190 : }
1191 :
1192 0 : let mut tenants = BTreeMap::new();
1193 0 :
1194 0 : let mut scheduler = Scheduler::new(nodes.values());
1195 0 :
1196 0 : #[cfg(feature = "testing")]
1197 0 : {
1198 0 : // Hack: insert scheduler state for all nodes referenced by shards, as compatibility
1199 0 : // tests only store the shards, not the nodes. The nodes will be loaded shortly
1200 0 : // after when pageservers start up and register.
1201 0 : let mut node_ids = HashSet::new();
1202 0 : for tsp in &tenant_shard_persistence {
1203 0 : if let Some(node_id) = tsp.generation_pageserver {
1204 0 : node_ids.insert(node_id);
1205 0 : }
1206 : }
1207 0 : for node_id in node_ids {
1208 0 : tracing::info!("Creating node {} in scheduler for tests", node_id);
1209 0 : let node = Node::new(
1210 0 : NodeId(node_id as u64),
1211 0 : "".to_string(),
1212 0 : 123,
1213 0 : "".to_string(),
1214 0 : 123,
1215 0 : );
1216 0 :
1217 0 : scheduler.node_upsert(&node);
1218 : }
1219 : }
1220 0 : for tsp in tenant_shard_persistence {
1221 0 : let tenant_shard_id = tsp.get_tenant_shard_id()?;
1222 :
1223 : // We will populate intent properly later in [`Self::startup_reconcile`], initially populate
1224 : // it with what we can infer: the node for which a generation was most recently issued.
1225 0 : let mut intent = IntentState::new();
1226 0 : if let Some(generation_pageserver) = tsp.generation_pageserver.map(|n| NodeId(n as u64))
1227 : {
1228 0 : if nodes.contains_key(&generation_pageserver) {
1229 0 : intent.set_attached(&mut scheduler, Some(generation_pageserver));
1230 0 : } else {
1231 : // If a node was removed before being completely drained, it is legal for it to leave behind a `generation_pageserver` referring
1232 : // to a non-existent node, because node deletion doesn't block on completing the reconciliations that will issue new generations
1233 : // on different pageservers.
1234 0 : tracing::warn!("Tenant shard {tenant_shard_id} references non-existent node {generation_pageserver} in database, will be rescheduled");
1235 : }
1236 0 : }
1237 0 : let new_tenant = TenantShard::from_persistent(tsp, intent)?;
1238 :
1239 0 : tenants.insert(tenant_shard_id, new_tenant);
1240 : }
1241 :
1242 0 : let (startup_completion, startup_complete) = utils::completion::channel();
1243 0 :
1244 0 : // This channel is continuously consumed by process_results, so doesn't need to be very large.
1245 0 : let (bg_compute_notify_result_tx, bg_compute_notify_result_rx) =
1246 0 : tokio::sync::mpsc::channel(512);
1247 0 :
1248 0 : let (delayed_reconcile_tx, delayed_reconcile_rx) =
1249 0 : tokio::sync::mpsc::channel(MAX_DELAYED_RECONCILES);
1250 0 :
1251 0 : let cancel = CancellationToken::new();
1252 0 : let reconcilers_cancel = cancel.child_token();
1253 0 :
1254 0 : let heartbeater = Heartbeater::new(
1255 0 : config.jwt_token.clone(),
1256 0 : config.max_offline_interval,
1257 0 : config.max_warming_up_interval,
1258 0 : cancel.clone(),
1259 0 : );
1260 0 : let this = Arc::new(Self {
1261 0 : inner: Arc::new(std::sync::RwLock::new(ServiceState::new(
1262 0 : nodes,
1263 0 : tenants,
1264 0 : scheduler,
1265 0 : delayed_reconcile_rx,
1266 0 : ))),
1267 0 : config: config.clone(),
1268 0 : persistence,
1269 0 : compute_hook: Arc::new(ComputeHook::new(config.clone())),
1270 0 : result_tx,
1271 0 : heartbeater,
1272 0 : reconciler_concurrency: Arc::new(tokio::sync::Semaphore::new(
1273 0 : config.reconciler_concurrency,
1274 0 : )),
1275 0 : delayed_reconcile_tx,
1276 0 : abort_tx,
1277 0 : startup_complete: startup_complete.clone(),
1278 0 : cancel,
1279 0 : reconcilers_cancel,
1280 0 : gate: Gate::default(),
1281 0 : reconcilers_gate: Gate::default(),
1282 0 : tenant_op_locks: Default::default(),
1283 0 : node_op_locks: Default::default(),
1284 0 : });
1285 0 :
1286 0 : let result_task_this = this.clone();
1287 0 : tokio::task::spawn(async move {
1288 : // Block shutdown until we're done (we must respect self.cancel)
1289 0 : if let Ok(_gate) = result_task_this.gate.enter() {
1290 0 : result_task_this
1291 0 : .process_results(result_rx, bg_compute_notify_result_rx)
1292 0 : .await
1293 0 : }
1294 0 : });
1295 0 :
1296 0 : tokio::task::spawn({
1297 0 : let this = this.clone();
1298 0 : async move {
1299 : // Block shutdown until we're done (we must respect self.cancel)
1300 0 : if let Ok(_gate) = this.gate.enter() {
1301 0 : this.process_aborts(abort_rx).await
1302 0 : }
1303 0 : }
1304 0 : });
1305 0 :
1306 0 : tokio::task::spawn({
1307 0 : let this = this.clone();
1308 0 : async move {
1309 0 : if let Ok(_gate) = this.gate.enter() {
1310 0 : loop {
1311 0 : tokio::select! {
1312 : _ = this.cancel.cancelled() => {
1313 : break;
1314 : },
1315 : _ = tokio::time::sleep(Duration::from_secs(60)) => {}
1316 : };
1317 0 : this.tenant_op_locks.housekeeping();
1318 : }
1319 0 : }
1320 0 : }
1321 0 : });
1322 0 :
1323 0 : tokio::task::spawn({
1324 0 : let this = this.clone();
1325 0 : // We will block the [`Service::startup_complete`] barrier until [`Self::startup_reconcile`]
1326 0 : // is done.
1327 0 : let startup_completion = startup_completion.clone();
1328 0 : async move {
1329 : // Block shutdown until we're done (we must respect self.cancel)
1330 0 : let Ok(_gate) = this.gate.enter() else {
1331 0 : return;
1332 : };
1333 :
1334 0 : this.startup_reconcile(bg_compute_notify_result_tx).await;
1335 0 : drop(startup_completion);
1336 0 : }
1337 0 : });
1338 0 :
1339 0 : tokio::task::spawn({
1340 0 : let this = this.clone();
1341 0 : let startup_complete = startup_complete.clone();
1342 0 : async move {
1343 0 : startup_complete.wait().await;
1344 0 : this.background_reconcile().await;
1345 0 : }
1346 0 : });
1347 0 :
1348 0 : tokio::task::spawn({
1349 0 : let this = this.clone();
1350 0 : let startup_complete = startup_complete.clone();
1351 0 : async move {
1352 0 : startup_complete.wait().await;
1353 0 : this.spawn_heartbeat_driver().await;
1354 0 : }
1355 0 : });
1356 0 :
1357 0 : Ok(this)
1358 0 : }
1359 :
1360 0 : pub(crate) async fn attach_hook(
1361 0 : &self,
1362 0 : attach_req: AttachHookRequest,
1363 0 : ) -> anyhow::Result<AttachHookResponse> {
1364 0 : let _tenant_lock = trace_exclusive_lock(
1365 0 : &self.tenant_op_locks,
1366 0 : attach_req.tenant_shard_id.tenant_id,
1367 0 : TenantOperations::AttachHook,
1368 0 : )
1369 0 : .await;
1370 :
1371 : // This is a test hook. To enable using it on tenants that were created directly with
1372 : // the pageserver API (not via this service), we will auto-create any missing tenant
1373 : // shards with default state.
1374 0 : let insert = {
1375 0 : let locked = self.inner.write().unwrap();
1376 0 : !locked.tenants.contains_key(&attach_req.tenant_shard_id)
1377 0 : };
1378 0 :
1379 0 : if insert {
1380 0 : let tsp = TenantShardPersistence {
1381 0 : tenant_id: attach_req.tenant_shard_id.tenant_id.to_string(),
1382 0 : shard_number: attach_req.tenant_shard_id.shard_number.0 as i32,
1383 0 : shard_count: attach_req.tenant_shard_id.shard_count.literal() as i32,
1384 0 : shard_stripe_size: 0,
1385 0 : generation: attach_req.generation_override.or(Some(0)),
1386 0 : generation_pageserver: None,
1387 0 : placement_policy: serde_json::to_string(&PlacementPolicy::Attached(0)).unwrap(),
1388 0 : config: serde_json::to_string(&TenantConfig::default()).unwrap(),
1389 0 : splitting: SplitState::default(),
1390 0 : scheduling_policy: serde_json::to_string(&ShardSchedulingPolicy::default())
1391 0 : .unwrap(),
1392 0 : };
1393 0 :
1394 0 : match self.persistence.insert_tenant_shards(vec![tsp]).await {
1395 0 : Err(e) => match e {
1396 : DatabaseError::Query(diesel::result::Error::DatabaseError(
1397 : DatabaseErrorKind::UniqueViolation,
1398 : _,
1399 : )) => {
1400 0 : tracing::info!(
1401 0 : "Raced with another request to insert tenant {}",
1402 : attach_req.tenant_shard_id
1403 : )
1404 : }
1405 0 : _ => return Err(e.into()),
1406 : },
1407 : Ok(()) => {
1408 0 : tracing::info!("Inserted shard {} in database", attach_req.tenant_shard_id);
1409 :
1410 0 : let mut locked = self.inner.write().unwrap();
1411 0 : locked.tenants.insert(
1412 0 : attach_req.tenant_shard_id,
1413 0 : TenantShard::new(
1414 0 : attach_req.tenant_shard_id,
1415 0 : ShardIdentity::unsharded(),
1416 0 : PlacementPolicy::Attached(0),
1417 0 : ),
1418 0 : );
1419 0 : tracing::info!("Inserted shard {} in memory", attach_req.tenant_shard_id);
1420 : }
1421 : }
1422 0 : }
1423 :
1424 0 : let new_generation = if let Some(req_node_id) = attach_req.node_id {
1425 0 : let maybe_tenant_conf = {
1426 0 : let locked = self.inner.write().unwrap();
1427 0 : locked
1428 0 : .tenants
1429 0 : .get(&attach_req.tenant_shard_id)
1430 0 : .map(|t| t.config.clone())
1431 0 : };
1432 0 :
1433 0 : match maybe_tenant_conf {
1434 0 : Some(conf) => {
1435 0 : let new_generation = self
1436 0 : .persistence
1437 0 : .increment_generation(attach_req.tenant_shard_id, req_node_id)
1438 0 : .await?;
1439 :
1440 : // Persist the placement policy update. This is required
1441 : // when we reattaching a detached tenant.
1442 0 : self.persistence
1443 0 : .update_tenant_shard(
1444 0 : TenantFilter::Shard(attach_req.tenant_shard_id),
1445 0 : Some(PlacementPolicy::Attached(0)),
1446 0 : Some(conf),
1447 0 : None,
1448 0 : None,
1449 0 : )
1450 0 : .await?;
1451 0 : Some(new_generation)
1452 : }
1453 : None => {
1454 0 : anyhow::bail!("Attach hook handling raced with tenant removal")
1455 : }
1456 : }
1457 : } else {
1458 0 : self.persistence.detach(attach_req.tenant_shard_id).await?;
1459 0 : None
1460 : };
1461 :
1462 0 : let mut locked = self.inner.write().unwrap();
1463 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
1464 0 :
1465 0 : let tenant_shard = tenants
1466 0 : .get_mut(&attach_req.tenant_shard_id)
1467 0 : .expect("Checked for existence above");
1468 :
1469 0 : if let Some(new_generation) = new_generation {
1470 0 : tenant_shard.generation = Some(new_generation);
1471 0 : tenant_shard.policy = PlacementPolicy::Attached(0);
1472 0 : } else {
1473 : // This is a detach notification. We must update placement policy to avoid re-attaching
1474 : // during background scheduling/reconciliation, or during storage controller restart.
1475 0 : assert!(attach_req.node_id.is_none());
1476 0 : tenant_shard.policy = PlacementPolicy::Detached;
1477 : }
1478 :
1479 0 : if let Some(attaching_pageserver) = attach_req.node_id.as_ref() {
1480 0 : tracing::info!(
1481 : tenant_id = %attach_req.tenant_shard_id,
1482 : ps_id = %attaching_pageserver,
1483 : generation = ?tenant_shard.generation,
1484 0 : "issuing",
1485 : );
1486 0 : } else if let Some(ps_id) = tenant_shard.intent.get_attached() {
1487 0 : tracing::info!(
1488 : tenant_id = %attach_req.tenant_shard_id,
1489 : %ps_id,
1490 : generation = ?tenant_shard.generation,
1491 0 : "dropping",
1492 : );
1493 : } else {
1494 0 : tracing::info!(
1495 : tenant_id = %attach_req.tenant_shard_id,
1496 0 : "no-op: tenant already has no pageserver");
1497 : }
1498 0 : tenant_shard
1499 0 : .intent
1500 0 : .set_attached(scheduler, attach_req.node_id);
1501 0 :
1502 0 : tracing::info!(
1503 0 : "attach_hook: tenant {} set generation {:?}, pageserver {}",
1504 0 : attach_req.tenant_shard_id,
1505 0 : tenant_shard.generation,
1506 0 : // TODO: this is an odd number of 0xf's
1507 0 : attach_req.node_id.unwrap_or(utils::id::NodeId(0xfffffff))
1508 : );
1509 :
1510 : // Trick the reconciler into not doing anything for this tenant: this helps
1511 : // tests that manually configure a tenant on the pagesrever, and then call this
1512 : // attach hook: they don't want background reconciliation to modify what they
1513 : // did to the pageserver.
1514 : #[cfg(feature = "testing")]
1515 : {
1516 0 : if let Some(node_id) = attach_req.node_id {
1517 0 : tenant_shard.observed.locations = HashMap::from([(
1518 0 : node_id,
1519 0 : ObservedStateLocation {
1520 0 : conf: Some(attached_location_conf(
1521 0 : tenant_shard.generation.unwrap(),
1522 0 : &tenant_shard.shard,
1523 0 : &tenant_shard.config,
1524 0 : &PlacementPolicy::Attached(0),
1525 0 : )),
1526 0 : },
1527 0 : )]);
1528 0 : } else {
1529 0 : tenant_shard.observed.locations.clear();
1530 0 : }
1531 : }
1532 :
1533 0 : Ok(AttachHookResponse {
1534 0 : gen: attach_req
1535 0 : .node_id
1536 0 : .map(|_| tenant_shard.generation.expect("Test hook, not used on tenants that are mid-onboarding with a NULL generation").into().unwrap()),
1537 0 : })
1538 0 : }
1539 :
1540 0 : pub(crate) fn inspect(&self, inspect_req: InspectRequest) -> InspectResponse {
1541 0 : let locked = self.inner.read().unwrap();
1542 0 :
1543 0 : let tenant_shard = locked.tenants.get(&inspect_req.tenant_shard_id);
1544 0 :
1545 0 : InspectResponse {
1546 0 : attachment: tenant_shard.and_then(|s| {
1547 0 : s.intent
1548 0 : .get_attached()
1549 0 : .map(|ps| (s.generation.expect("Test hook, not used on tenants that are mid-onboarding with a NULL generation").into().unwrap(), ps))
1550 0 : }),
1551 0 : }
1552 0 : }
1553 :
1554 : // When the availability state of a node transitions to active, we must do a full reconciliation
1555 : // of LocationConfigs on that node. This is because while a node was offline:
1556 : // - we might have proceeded through startup_reconcile without checking for extraneous LocationConfigs on this node
1557 : // - aborting a tenant shard split might have left rogue child shards behind on this node.
1558 : //
1559 : // This function must complete _before_ setting a `Node` to Active: once it is set to Active, other
1560 : // Reconcilers might communicate with the node, and these must not overlap with the work we do in
1561 : // this function.
1562 : //
1563 : // The reconciliation logic in here is very similar to what [`Self::startup_reconcile`] does, but
1564 : // for written for a single node rather than as a batch job for all nodes.
1565 0 : #[tracing::instrument(skip_all, fields(node_id=%node.get_id()))]
1566 : async fn node_activate_reconcile(
1567 : &self,
1568 : mut node: Node,
1569 : _lock: &TracingExclusiveGuard<NodeOperations>,
1570 : ) -> Result<(), ApiError> {
1571 : // This Node is a mutable local copy: we will set it active so that we can use its
1572 : // API client to reconcile with the node. The Node in [`Self::nodes`] will get updated
1573 : // later.
1574 : node.set_availability(NodeAvailability::Active(UtilizationScore::worst()));
1575 :
1576 : let configs = match node
1577 : .with_client_retries(
1578 0 : |client| async move { client.list_location_config().await },
1579 : &self.config.jwt_token,
1580 : 1,
1581 : 5,
1582 : SHORT_RECONCILE_TIMEOUT,
1583 : &self.cancel,
1584 : )
1585 : .await
1586 : {
1587 : None => {
1588 : // We're shutting down (the Node's cancellation token can't have fired, because
1589 : // we're the only scope that has a reference to it, and we didn't fire it).
1590 : return Err(ApiError::ShuttingDown);
1591 : }
1592 : Some(Err(e)) => {
1593 : // This node didn't succeed listing its locations: it may not proceed to active state
1594 : // as it is apparently unavailable.
1595 : return Err(ApiError::PreconditionFailed(
1596 : format!("Failed to query node location configs, cannot activate ({e})").into(),
1597 : ));
1598 : }
1599 : Some(Ok(configs)) => configs,
1600 : };
1601 : tracing::info!("Loaded {} LocationConfigs", configs.tenant_shards.len());
1602 :
1603 : let mut cleanup = Vec::new();
1604 : {
1605 : let mut locked = self.inner.write().unwrap();
1606 :
1607 : for (tenant_shard_id, observed_loc) in configs.tenant_shards {
1608 : let Some(tenant_shard) = locked.tenants.get_mut(&tenant_shard_id) else {
1609 : cleanup.push(tenant_shard_id);
1610 : continue;
1611 : };
1612 : tenant_shard
1613 : .observed
1614 : .locations
1615 : .insert(node.get_id(), ObservedStateLocation { conf: observed_loc });
1616 : }
1617 : }
1618 :
1619 : for tenant_shard_id in cleanup {
1620 : tracing::info!("Detaching {tenant_shard_id}");
1621 : match node
1622 : .with_client_retries(
1623 0 : |client| async move {
1624 0 : let config = LocationConfig {
1625 0 : mode: LocationConfigMode::Detached,
1626 0 : generation: None,
1627 0 : secondary_conf: None,
1628 0 : shard_number: tenant_shard_id.shard_number.0,
1629 0 : shard_count: tenant_shard_id.shard_count.literal(),
1630 0 : shard_stripe_size: 0,
1631 0 : tenant_conf: models::TenantConfig::default(),
1632 0 : };
1633 0 : client
1634 0 : .location_config(tenant_shard_id, config, None, false)
1635 0 : .await
1636 0 : },
1637 : &self.config.jwt_token,
1638 : 1,
1639 : 5,
1640 : SHORT_RECONCILE_TIMEOUT,
1641 : &self.cancel,
1642 : )
1643 : .await
1644 : {
1645 : None => {
1646 : // We're shutting down (the Node's cancellation token can't have fired, because
1647 : // we're the only scope that has a reference to it, and we didn't fire it).
1648 : return Err(ApiError::ShuttingDown);
1649 : }
1650 : Some(Err(e)) => {
1651 : // Do not let the node proceed to Active state if it is not responsive to requests
1652 : // to detach. This could happen if e.g. a shutdown bug in the pageserver is preventing
1653 : // detach completing: we should not let this node back into the set of nodes considered
1654 : // okay for scheduling.
1655 : return Err(ApiError::Conflict(format!(
1656 : "Node {node} failed to detach {tenant_shard_id}: {e}"
1657 : )));
1658 : }
1659 : Some(Ok(_)) => {}
1660 : };
1661 : }
1662 :
1663 : Ok(())
1664 : }
1665 :
1666 0 : pub(crate) async fn re_attach(
1667 0 : &self,
1668 0 : reattach_req: ReAttachRequest,
1669 0 : ) -> Result<ReAttachResponse, ApiError> {
1670 0 : if let Some(register_req) = reattach_req.register {
1671 0 : self.node_register(register_req).await?;
1672 0 : }
1673 :
1674 : // Ordering: we must persist generation number updates before making them visible in the in-memory state
1675 0 : let incremented_generations = self.persistence.re_attach(reattach_req.node_id).await?;
1676 :
1677 0 : tracing::info!(
1678 : node_id=%reattach_req.node_id,
1679 0 : "Incremented {} tenant shards' generations",
1680 0 : incremented_generations.len()
1681 : );
1682 :
1683 : // Apply the updated generation to our in-memory state, and
1684 : // gather discover secondary locations.
1685 0 : let mut locked = self.inner.write().unwrap();
1686 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
1687 0 :
1688 0 : let mut response = ReAttachResponse {
1689 0 : tenants: Vec::new(),
1690 0 : };
1691 :
1692 : // TODO: cancel/restart any running reconciliation for this tenant, it might be trying
1693 : // to call location_conf API with an old generation. Wait for cancellation to complete
1694 : // before responding to this request. Requires well implemented CancellationToken logic
1695 : // all the way to where we call location_conf. Even then, there can still be a location_conf
1696 : // request in flight over the network: TODO handle that by making location_conf API refuse
1697 : // to go backward in generations.
1698 :
1699 : // Scan through all shards, applying updates for ones where we updated generation
1700 : // and identifying shards that intend to have a secondary location on this node.
1701 0 : for (tenant_shard_id, shard) in tenants {
1702 0 : if let Some(new_gen) = incremented_generations.get(tenant_shard_id) {
1703 0 : let new_gen = *new_gen;
1704 0 : response.tenants.push(ReAttachResponseTenant {
1705 0 : id: *tenant_shard_id,
1706 0 : gen: Some(new_gen.into().unwrap()),
1707 0 : // A tenant is only put into multi or stale modes in the middle of a [`Reconciler::live_migrate`]
1708 0 : // execution. If a pageserver is restarted during that process, then the reconcile pass will
1709 0 : // fail, and start from scratch, so it doesn't make sense for us to try and preserve
1710 0 : // the stale/multi states at this point.
1711 0 : mode: LocationConfigMode::AttachedSingle,
1712 0 : });
1713 0 :
1714 0 : shard.generation = std::cmp::max(shard.generation, Some(new_gen));
1715 0 : if let Some(observed) = shard.observed.locations.get_mut(&reattach_req.node_id) {
1716 : // Why can we update `observed` even though we're not sure our response will be received
1717 : // by the pageserver? Because the pageserver will not proceed with startup until
1718 : // it has processed response: if it loses it, we'll see another request and increment
1719 : // generation again, avoiding any uncertainty about dirtiness of tenant's state.
1720 0 : if let Some(conf) = observed.conf.as_mut() {
1721 0 : conf.generation = new_gen.into();
1722 0 : }
1723 0 : } else {
1724 0 : // This node has no observed state for the shard: perhaps it was offline
1725 0 : // when the pageserver restarted. Insert a None, so that the Reconciler
1726 0 : // will be prompted to learn the location's state before it makes changes.
1727 0 : shard
1728 0 : .observed
1729 0 : .locations
1730 0 : .insert(reattach_req.node_id, ObservedStateLocation { conf: None });
1731 0 : }
1732 0 : } else if shard.intent.get_secondary().contains(&reattach_req.node_id) {
1733 0 : // Ordering: pageserver will not accept /location_config requests until it has
1734 0 : // finished processing the response from re-attach. So we can update our in-memory state
1735 0 : // now, and be confident that we are not stamping on the result of some later location config.
1736 0 : // TODO: however, we are not strictly ordered wrt ReconcileResults queue,
1737 0 : // so we might update observed state here, and then get over-written by some racing
1738 0 : // ReconcileResult. The impact is low however, since we have set state on pageserver something
1739 0 : // that matches intent, so worst case if we race then we end up doing a spurious reconcile.
1740 0 :
1741 0 : response.tenants.push(ReAttachResponseTenant {
1742 0 : id: *tenant_shard_id,
1743 0 : gen: None,
1744 0 : mode: LocationConfigMode::Secondary,
1745 0 : });
1746 0 :
1747 0 : // We must not update observed, because we have no guarantee that our
1748 0 : // response will be received by the pageserver. This could leave it
1749 0 : // falsely dirty, but the resulting reconcile should be idempotent.
1750 0 : }
1751 : }
1752 :
1753 : // We consider a node Active once we have composed a re-attach response, but we
1754 : // do not call [`Self::node_activate_reconcile`]: the handling of the re-attach response
1755 : // implicitly synchronizes the LocationConfigs on the node.
1756 : //
1757 : // Setting a node active unblocks any Reconcilers that might write to the location config API,
1758 : // but those requests will not be accepted by the node until it has finished processing
1759 : // the re-attach response.
1760 : //
1761 : // Additionally, reset the nodes scheduling policy to match the conditional update done
1762 : // in [`Persistence::re_attach`].
1763 0 : if let Some(node) = nodes.get(&reattach_req.node_id) {
1764 0 : let reset_scheduling = matches!(
1765 0 : node.get_scheduling(),
1766 : NodeSchedulingPolicy::PauseForRestart
1767 : | NodeSchedulingPolicy::Draining
1768 : | NodeSchedulingPolicy::Filling
1769 : );
1770 :
1771 0 : let mut new_nodes = (**nodes).clone();
1772 0 : if let Some(node) = new_nodes.get_mut(&reattach_req.node_id) {
1773 0 : if reset_scheduling {
1774 0 : node.set_scheduling(NodeSchedulingPolicy::Active);
1775 0 : }
1776 :
1777 0 : tracing::info!("Marking {} warming-up on reattach", reattach_req.node_id);
1778 0 : node.set_availability(NodeAvailability::WarmingUp(std::time::Instant::now()));
1779 0 :
1780 0 : scheduler.node_upsert(node);
1781 0 : let new_nodes = Arc::new(new_nodes);
1782 0 : *nodes = new_nodes;
1783 : } else {
1784 0 : tracing::error!(
1785 0 : "Reattaching node {} was removed while processing the request",
1786 : reattach_req.node_id
1787 : );
1788 : }
1789 0 : }
1790 :
1791 0 : Ok(response)
1792 0 : }
1793 :
1794 0 : pub(crate) fn validate(&self, validate_req: ValidateRequest) -> ValidateResponse {
1795 0 : let locked = self.inner.read().unwrap();
1796 0 :
1797 0 : let mut response = ValidateResponse {
1798 0 : tenants: Vec::new(),
1799 0 : };
1800 :
1801 0 : for req_tenant in validate_req.tenants {
1802 0 : if let Some(tenant_shard) = locked.tenants.get(&req_tenant.id) {
1803 0 : let valid = tenant_shard.generation == Some(Generation::new(req_tenant.gen));
1804 0 : tracing::info!(
1805 0 : "handle_validate: {}(gen {}): valid={valid} (latest {:?})",
1806 : req_tenant.id,
1807 : req_tenant.gen,
1808 : tenant_shard.generation
1809 : );
1810 0 : response.tenants.push(ValidateResponseTenant {
1811 0 : id: req_tenant.id,
1812 0 : valid,
1813 0 : });
1814 0 : } else {
1815 0 : // After tenant deletion, we may approve any validation. This avoids
1816 0 : // spurious warnings on the pageserver if it has pending LSN updates
1817 0 : // at the point a deletion happens.
1818 0 : response.tenants.push(ValidateResponseTenant {
1819 0 : id: req_tenant.id,
1820 0 : valid: true,
1821 0 : });
1822 0 : }
1823 : }
1824 0 : response
1825 0 : }
1826 :
1827 0 : pub(crate) async fn tenant_create(
1828 0 : &self,
1829 0 : create_req: TenantCreateRequest,
1830 0 : ) -> Result<TenantCreateResponse, ApiError> {
1831 0 : let tenant_id = create_req.new_tenant_id.tenant_id;
1832 :
1833 : // Exclude any concurrent attempts to create/access the same tenant ID
1834 0 : let _tenant_lock = trace_exclusive_lock(
1835 0 : &self.tenant_op_locks,
1836 0 : create_req.new_tenant_id.tenant_id,
1837 0 : TenantOperations::Create,
1838 0 : )
1839 0 : .await;
1840 0 : let (response, waiters) = self.do_tenant_create(create_req).await?;
1841 :
1842 0 : if let Err(e) = self.await_waiters(waiters, RECONCILE_TIMEOUT).await {
1843 : // Avoid deadlock: reconcile may fail while notifying compute, if the cloud control plane refuses to
1844 : // accept compute notifications while it is in the process of creating. Reconciliation will
1845 : // be retried in the background.
1846 0 : tracing::warn!(%tenant_id, "Reconcile not done yet while creating tenant ({e})");
1847 0 : }
1848 0 : Ok(response)
1849 0 : }
1850 :
1851 0 : pub(crate) async fn do_tenant_create(
1852 0 : &self,
1853 0 : create_req: TenantCreateRequest,
1854 0 : ) -> Result<(TenantCreateResponse, Vec<ReconcilerWaiter>), ApiError> {
1855 0 : let placement_policy = create_req
1856 0 : .placement_policy
1857 0 : .clone()
1858 0 : // As a default, zero secondaries is convenient for tests that don't choose a policy.
1859 0 : .unwrap_or(PlacementPolicy::Attached(0));
1860 :
1861 : // This service expects to handle sharding itself: it is an error to try and directly create
1862 : // a particular shard here.
1863 0 : let tenant_id = if !create_req.new_tenant_id.is_unsharded() {
1864 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
1865 0 : "Attempted to create a specific shard, this API is for creating the whole tenant"
1866 0 : )));
1867 : } else {
1868 0 : create_req.new_tenant_id.tenant_id
1869 0 : };
1870 0 :
1871 0 : tracing::info!(
1872 0 : "Creating tenant {}, shard_count={:?}",
1873 : create_req.new_tenant_id,
1874 : create_req.shard_parameters.count,
1875 : );
1876 :
1877 0 : let create_ids = (0..create_req.shard_parameters.count.count())
1878 0 : .map(|i| TenantShardId {
1879 0 : tenant_id,
1880 0 : shard_number: ShardNumber(i),
1881 0 : shard_count: create_req.shard_parameters.count,
1882 0 : })
1883 0 : .collect::<Vec<_>>();
1884 :
1885 : // If the caller specifies a None generation, it means "start from default". This is different
1886 : // to [`Self::tenant_location_config`], where a None generation is used to represent
1887 : // an incompletely-onboarded tenant.
1888 0 : let initial_generation = if matches!(placement_policy, PlacementPolicy::Secondary) {
1889 0 : tracing::info!(
1890 0 : "tenant_create: secondary mode, generation is_some={}",
1891 0 : create_req.generation.is_some()
1892 : );
1893 0 : create_req.generation.map(Generation::new)
1894 : } else {
1895 0 : tracing::info!(
1896 0 : "tenant_create: not secondary mode, generation is_some={}",
1897 0 : create_req.generation.is_some()
1898 : );
1899 0 : Some(
1900 0 : create_req
1901 0 : .generation
1902 0 : .map(Generation::new)
1903 0 : .unwrap_or(INITIAL_GENERATION),
1904 0 : )
1905 : };
1906 :
1907 : // Ordering: we persist tenant shards before creating them on the pageserver. This enables a caller
1908 : // to clean up after themselves by issuing a tenant deletion if something goes wrong and we restart
1909 : // during the creation, rather than risking leaving orphan objects in S3.
1910 0 : let persist_tenant_shards = create_ids
1911 0 : .iter()
1912 0 : .map(|tenant_shard_id| TenantShardPersistence {
1913 0 : tenant_id: tenant_shard_id.tenant_id.to_string(),
1914 0 : shard_number: tenant_shard_id.shard_number.0 as i32,
1915 0 : shard_count: tenant_shard_id.shard_count.literal() as i32,
1916 0 : shard_stripe_size: create_req.shard_parameters.stripe_size.0 as i32,
1917 0 : generation: initial_generation.map(|g| g.into().unwrap() as i32),
1918 0 : // The pageserver is not known until scheduling happens: we will set this column when
1919 0 : // incrementing the generation the first time we attach to a pageserver.
1920 0 : generation_pageserver: None,
1921 0 : placement_policy: serde_json::to_string(&placement_policy).unwrap(),
1922 0 : config: serde_json::to_string(&create_req.config).unwrap(),
1923 0 : splitting: SplitState::default(),
1924 0 : scheduling_policy: serde_json::to_string(&ShardSchedulingPolicy::default())
1925 0 : .unwrap(),
1926 0 : })
1927 0 : .collect();
1928 0 :
1929 0 : match self
1930 0 : .persistence
1931 0 : .insert_tenant_shards(persist_tenant_shards)
1932 0 : .await
1933 : {
1934 0 : Ok(_) => {}
1935 : Err(DatabaseError::Query(diesel::result::Error::DatabaseError(
1936 : DatabaseErrorKind::UniqueViolation,
1937 : _,
1938 : ))) => {
1939 : // Unique key violation: this is probably a retry. Because the shard count is part of the unique key,
1940 : // if we see a unique key violation it means that the creation request's shard count matches the previous
1941 : // creation's shard count.
1942 0 : tracing::info!("Tenant shards already present in database, proceeding with idempotent creation...");
1943 : }
1944 : // Any other database error is unexpected and a bug.
1945 0 : Err(e) => return Err(ApiError::InternalServerError(anyhow::anyhow!(e))),
1946 : };
1947 :
1948 0 : let mut schedule_context = ScheduleContext::default();
1949 :
1950 0 : let (waiters, response_shards) = {
1951 0 : let mut locked = self.inner.write().unwrap();
1952 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
1953 0 :
1954 0 : let mut response_shards = Vec::new();
1955 0 : let mut schcedule_error = None;
1956 :
1957 0 : for tenant_shard_id in create_ids {
1958 0 : tracing::info!("Creating shard {tenant_shard_id}...");
1959 :
1960 : use std::collections::btree_map::Entry;
1961 0 : match tenants.entry(tenant_shard_id) {
1962 0 : Entry::Occupied(mut entry) => {
1963 0 : tracing::info!(
1964 0 : "Tenant shard {tenant_shard_id} already exists while creating"
1965 : );
1966 :
1967 : // TODO: schedule() should take an anti-affinity expression that pushes
1968 : // attached and secondary locations (independently) away frorm those
1969 : // pageservers also holding a shard for this tenant.
1970 :
1971 0 : entry
1972 0 : .get_mut()
1973 0 : .schedule(scheduler, &mut schedule_context)
1974 0 : .map_err(|e| {
1975 0 : ApiError::Conflict(format!(
1976 0 : "Failed to schedule shard {tenant_shard_id}: {e}"
1977 0 : ))
1978 0 : })?;
1979 :
1980 0 : if let Some(node_id) = entry.get().intent.get_attached() {
1981 0 : let generation = entry
1982 0 : .get()
1983 0 : .generation
1984 0 : .expect("Generation is set when in attached mode");
1985 0 : response_shards.push(TenantCreateResponseShard {
1986 0 : shard_id: tenant_shard_id,
1987 0 : node_id: *node_id,
1988 0 : generation: generation.into().unwrap(),
1989 0 : });
1990 0 : }
1991 :
1992 0 : continue;
1993 : }
1994 0 : Entry::Vacant(entry) => {
1995 0 : let state = entry.insert(TenantShard::new(
1996 0 : tenant_shard_id,
1997 0 : ShardIdentity::from_params(
1998 0 : tenant_shard_id.shard_number,
1999 0 : &create_req.shard_parameters,
2000 0 : ),
2001 0 : placement_policy.clone(),
2002 0 : ));
2003 0 :
2004 0 : state.generation = initial_generation;
2005 0 : state.config = create_req.config.clone();
2006 0 : if let Err(e) = state.schedule(scheduler, &mut schedule_context) {
2007 0 : schcedule_error = Some(e);
2008 0 : }
2009 :
2010 : // Only include shards in result if we are attaching: the purpose
2011 : // of the response is to tell the caller where the shards are attached.
2012 0 : if let Some(node_id) = state.intent.get_attached() {
2013 0 : let generation = state
2014 0 : .generation
2015 0 : .expect("Generation is set when in attached mode");
2016 0 : response_shards.push(TenantCreateResponseShard {
2017 0 : shard_id: tenant_shard_id,
2018 0 : node_id: *node_id,
2019 0 : generation: generation.into().unwrap(),
2020 0 : });
2021 0 : }
2022 : }
2023 : };
2024 : }
2025 :
2026 : // If we failed to schedule shards, then they are still created in the controller,
2027 : // but we return an error to the requester to avoid a silent failure when someone
2028 : // tries to e.g. create a tenant whose placement policy requires more nodes than
2029 : // are present in the system. We do this here rather than in the above loop, to
2030 : // avoid situations where we only create a subset of shards in the tenant.
2031 0 : if let Some(e) = schcedule_error {
2032 0 : return Err(ApiError::Conflict(format!(
2033 0 : "Failed to schedule shard(s): {e}"
2034 0 : )));
2035 0 : }
2036 0 :
2037 0 : let waiters = tenants
2038 0 : .range_mut(TenantShardId::tenant_range(tenant_id))
2039 0 : .filter_map(|(_shard_id, shard)| self.maybe_reconcile_shard(shard, nodes))
2040 0 : .collect::<Vec<_>>();
2041 0 : (waiters, response_shards)
2042 0 : };
2043 0 :
2044 0 : Ok((
2045 0 : TenantCreateResponse {
2046 0 : shards: response_shards,
2047 0 : },
2048 0 : waiters,
2049 0 : ))
2050 0 : }
2051 :
2052 : /// Helper for functions that reconcile a number of shards, and would like to do a timeout-bounded
2053 : /// wait for reconciliation to complete before responding.
2054 0 : async fn await_waiters(
2055 0 : &self,
2056 0 : waiters: Vec<ReconcilerWaiter>,
2057 0 : timeout: Duration,
2058 0 : ) -> Result<(), ReconcileWaitError> {
2059 0 : let deadline = Instant::now().checked_add(timeout).unwrap();
2060 0 : for waiter in waiters {
2061 0 : let timeout = deadline.duration_since(Instant::now());
2062 0 : waiter.wait_timeout(timeout).await?;
2063 : }
2064 :
2065 0 : Ok(())
2066 0 : }
2067 :
2068 : /// Same as [`Service::await_waiters`], but returns the waiters which are still
2069 : /// in progress
2070 0 : async fn await_waiters_remainder(
2071 0 : &self,
2072 0 : waiters: Vec<ReconcilerWaiter>,
2073 0 : timeout: Duration,
2074 0 : ) -> Vec<ReconcilerWaiter> {
2075 0 : let deadline = Instant::now().checked_add(timeout).unwrap();
2076 0 : for waiter in waiters.iter() {
2077 0 : let timeout = deadline.duration_since(Instant::now());
2078 0 : let _ = waiter.wait_timeout(timeout).await;
2079 : }
2080 :
2081 0 : waiters
2082 0 : .into_iter()
2083 0 : .filter(|waiter| matches!(waiter.get_status(), ReconcilerStatus::InProgress))
2084 0 : .collect::<Vec<_>>()
2085 0 : }
2086 :
2087 : /// Part of [`Self::tenant_location_config`]: dissect an incoming location config request,
2088 : /// and transform it into either a tenant creation of a series of shard updates.
2089 : ///
2090 : /// If the incoming request makes no changes, a [`TenantCreateOrUpdate::Update`] result will
2091 : /// still be returned.
2092 0 : fn tenant_location_config_prepare(
2093 0 : &self,
2094 0 : tenant_id: TenantId,
2095 0 : req: TenantLocationConfigRequest,
2096 0 : ) -> TenantCreateOrUpdate {
2097 0 : let mut updates = Vec::new();
2098 0 : let mut locked = self.inner.write().unwrap();
2099 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
2100 0 : let tenant_shard_id = TenantShardId::unsharded(tenant_id);
2101 :
2102 : // Use location config mode as an indicator of policy.
2103 0 : let placement_policy = match req.config.mode {
2104 0 : LocationConfigMode::Detached => PlacementPolicy::Detached,
2105 0 : LocationConfigMode::Secondary => PlacementPolicy::Secondary,
2106 : LocationConfigMode::AttachedMulti
2107 : | LocationConfigMode::AttachedSingle
2108 : | LocationConfigMode::AttachedStale => {
2109 0 : if nodes.len() > 1 {
2110 0 : PlacementPolicy::Attached(1)
2111 : } else {
2112 : // Convenience for dev/test: if we just have one pageserver, import
2113 : // tenants into non-HA mode so that scheduling will succeed.
2114 0 : PlacementPolicy::Attached(0)
2115 : }
2116 : }
2117 : };
2118 :
2119 0 : let mut create = true;
2120 0 : for (shard_id, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
2121 : // Saw an existing shard: this is not a creation
2122 0 : create = false;
2123 :
2124 : // Shards may have initially been created by a Secondary request, where we
2125 : // would have left generation as None.
2126 : //
2127 : // We only update generation the first time we see an attached-mode request,
2128 : // and if there is no existing generation set. The caller is responsible for
2129 : // ensuring that no non-storage-controller pageserver ever uses a higher
2130 : // generation than they passed in here.
2131 : use LocationConfigMode::*;
2132 0 : let set_generation = match req.config.mode {
2133 0 : AttachedMulti | AttachedSingle | AttachedStale if shard.generation.is_none() => {
2134 0 : req.config.generation.map(Generation::new)
2135 : }
2136 0 : _ => None,
2137 : };
2138 :
2139 0 : updates.push(ShardUpdate {
2140 0 : tenant_shard_id: *shard_id,
2141 0 : placement_policy: placement_policy.clone(),
2142 0 : tenant_config: req.config.tenant_conf.clone(),
2143 0 : generation: set_generation,
2144 0 : });
2145 : }
2146 :
2147 0 : if create {
2148 : use LocationConfigMode::*;
2149 0 : let generation = match req.config.mode {
2150 0 : AttachedMulti | AttachedSingle | AttachedStale => req.config.generation,
2151 : // If a caller provided a generation in a non-attached request, ignore it
2152 : // and leave our generation as None: this enables a subsequent update to set
2153 : // the generation when setting an attached mode for the first time.
2154 0 : _ => None,
2155 : };
2156 :
2157 0 : TenantCreateOrUpdate::Create(
2158 0 : // Synthesize a creation request
2159 0 : TenantCreateRequest {
2160 0 : new_tenant_id: tenant_shard_id,
2161 0 : generation,
2162 0 : shard_parameters: ShardParameters {
2163 0 : count: tenant_shard_id.shard_count,
2164 0 : // We only import un-sharded or single-sharded tenants, so stripe
2165 0 : // size can be made up arbitrarily here.
2166 0 : stripe_size: ShardParameters::DEFAULT_STRIPE_SIZE,
2167 0 : },
2168 0 : placement_policy: Some(placement_policy),
2169 0 : config: req.config.tenant_conf,
2170 0 : },
2171 0 : )
2172 : } else {
2173 0 : assert!(!updates.is_empty());
2174 0 : TenantCreateOrUpdate::Update(updates)
2175 : }
2176 0 : }
2177 :
2178 : /// This API is used by the cloud control plane to migrate unsharded tenants that it created
2179 : /// directly with pageservers into this service.
2180 : ///
2181 : /// Cloud control plane MUST NOT continue issuing GENERATION NUMBERS for this tenant once it
2182 : /// has attempted to call this API. Failure to oblige to this rule may lead to S3 corruption.
2183 : /// Think of the first attempt to call this API as a transfer of absolute authority over the
2184 : /// tenant's source of generation numbers.
2185 : ///
2186 : /// The mode in this request coarse-grained control of tenants:
2187 : /// - Call with mode Attached* to upsert the tenant.
2188 : /// - Call with mode Secondary to either onboard a tenant without attaching it, or
2189 : /// to set an existing tenant to PolicyMode::Secondary
2190 : /// - Call with mode Detached to switch to PolicyMode::Detached
2191 0 : pub(crate) async fn tenant_location_config(
2192 0 : &self,
2193 0 : tenant_shard_id: TenantShardId,
2194 0 : req: TenantLocationConfigRequest,
2195 0 : ) -> Result<TenantLocationConfigResponse, ApiError> {
2196 : // We require an exclusive lock, because we are updating both persistent and in-memory state
2197 0 : let _tenant_lock = trace_exclusive_lock(
2198 0 : &self.tenant_op_locks,
2199 0 : tenant_shard_id.tenant_id,
2200 0 : TenantOperations::LocationConfig,
2201 0 : )
2202 0 : .await;
2203 :
2204 0 : if !tenant_shard_id.is_unsharded() {
2205 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
2206 0 : "This API is for importing single-sharded or unsharded tenants"
2207 0 : )));
2208 0 : }
2209 0 :
2210 0 : // First check if this is a creation or an update
2211 0 : let create_or_update = self.tenant_location_config_prepare(tenant_shard_id.tenant_id, req);
2212 0 :
2213 0 : let mut result = TenantLocationConfigResponse {
2214 0 : shards: Vec::new(),
2215 0 : stripe_size: None,
2216 0 : };
2217 0 : let waiters = match create_or_update {
2218 0 : TenantCreateOrUpdate::Create(create_req) => {
2219 0 : let (create_resp, waiters) = self.do_tenant_create(create_req).await?;
2220 0 : result.shards = create_resp
2221 0 : .shards
2222 0 : .into_iter()
2223 0 : .map(|s| TenantShardLocation {
2224 0 : node_id: s.node_id,
2225 0 : shard_id: s.shard_id,
2226 0 : })
2227 0 : .collect();
2228 0 : waiters
2229 : }
2230 0 : TenantCreateOrUpdate::Update(updates) => {
2231 0 : // Persist updates
2232 0 : // Ordering: write to the database before applying changes in-memory, so that
2233 0 : // we will not appear time-travel backwards on a restart.
2234 0 : let mut schedule_context = ScheduleContext::default();
2235 : for ShardUpdate {
2236 0 : tenant_shard_id,
2237 0 : placement_policy,
2238 0 : tenant_config,
2239 0 : generation,
2240 0 : } in &updates
2241 : {
2242 0 : self.persistence
2243 0 : .update_tenant_shard(
2244 0 : TenantFilter::Shard(*tenant_shard_id),
2245 0 : Some(placement_policy.clone()),
2246 0 : Some(tenant_config.clone()),
2247 0 : *generation,
2248 0 : None,
2249 0 : )
2250 0 : .await?;
2251 : }
2252 :
2253 : // Apply updates in-memory
2254 0 : let mut waiters = Vec::new();
2255 0 : {
2256 0 : let mut locked = self.inner.write().unwrap();
2257 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
2258 :
2259 : for ShardUpdate {
2260 0 : tenant_shard_id,
2261 0 : placement_policy,
2262 0 : tenant_config,
2263 0 : generation: update_generation,
2264 0 : } in updates
2265 : {
2266 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
2267 0 : tracing::warn!("Shard {tenant_shard_id} removed while updating");
2268 0 : continue;
2269 : };
2270 :
2271 : // Update stripe size
2272 0 : if result.stripe_size.is_none() && shard.shard.count.count() > 1 {
2273 0 : result.stripe_size = Some(shard.shard.stripe_size);
2274 0 : }
2275 :
2276 0 : shard.policy = placement_policy;
2277 0 : shard.config = tenant_config;
2278 0 : if let Some(generation) = update_generation {
2279 0 : shard.generation = Some(generation);
2280 0 : }
2281 :
2282 0 : shard.schedule(scheduler, &mut schedule_context)?;
2283 :
2284 0 : let maybe_waiter = self.maybe_reconcile_shard(shard, nodes);
2285 0 : if let Some(waiter) = maybe_waiter {
2286 0 : waiters.push(waiter);
2287 0 : }
2288 :
2289 0 : if let Some(node_id) = shard.intent.get_attached() {
2290 0 : result.shards.push(TenantShardLocation {
2291 0 : shard_id: tenant_shard_id,
2292 0 : node_id: *node_id,
2293 0 : })
2294 0 : }
2295 : }
2296 : }
2297 0 : waiters
2298 : }
2299 : };
2300 :
2301 0 : if let Err(e) = self.await_waiters(waiters, SHORT_RECONCILE_TIMEOUT).await {
2302 : // Do not treat a reconcile error as fatal: we have already applied any requested
2303 : // Intent changes, and the reconcile can fail for external reasons like unavailable
2304 : // compute notification API. In these cases, it is important that we do not
2305 : // cause the cloud control plane to retry forever on this API.
2306 0 : tracing::warn!(
2307 0 : "Failed to reconcile after /location_config: {e}, returning success anyway"
2308 : );
2309 0 : }
2310 :
2311 : // Logging the full result is useful because it lets us cross-check what the cloud control
2312 : // plane's tenant_shards table should contain.
2313 0 : tracing::info!("Complete, returning {result:?}");
2314 :
2315 0 : Ok(result)
2316 0 : }
2317 :
2318 0 : pub(crate) async fn tenant_config_set(&self, req: TenantConfigRequest) -> Result<(), ApiError> {
2319 : // We require an exclusive lock, because we are updating persistent and in-memory state
2320 0 : let _tenant_lock = trace_exclusive_lock(
2321 0 : &self.tenant_op_locks,
2322 0 : req.tenant_id,
2323 0 : TenantOperations::ConfigSet,
2324 0 : )
2325 0 : .await;
2326 :
2327 0 : let tenant_id = req.tenant_id;
2328 0 : let config = req.config;
2329 0 :
2330 0 : self.persistence
2331 0 : .update_tenant_shard(
2332 0 : TenantFilter::Tenant(req.tenant_id),
2333 0 : None,
2334 0 : Some(config.clone()),
2335 0 : None,
2336 0 : None,
2337 0 : )
2338 0 : .await?;
2339 :
2340 0 : let waiters = {
2341 0 : let mut waiters = Vec::new();
2342 0 : let mut locked = self.inner.write().unwrap();
2343 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
2344 0 : for (_shard_id, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
2345 0 : shard.config = config.clone();
2346 0 : if let Some(waiter) = self.maybe_reconcile_shard(shard, nodes) {
2347 0 : waiters.push(waiter);
2348 0 : }
2349 : }
2350 0 : waiters
2351 : };
2352 :
2353 0 : if let Err(e) = self.await_waiters(waiters, SHORT_RECONCILE_TIMEOUT).await {
2354 : // Treat this as success because we have stored the configuration. If e.g.
2355 : // a node was unavailable at this time, it should not stop us accepting a
2356 : // configuration change.
2357 0 : tracing::warn!(%tenant_id, "Accepted configuration update but reconciliation failed: {e}");
2358 0 : }
2359 :
2360 0 : Ok(())
2361 0 : }
2362 :
2363 0 : pub(crate) fn tenant_config_get(
2364 0 : &self,
2365 0 : tenant_id: TenantId,
2366 0 : ) -> Result<HashMap<&str, serde_json::Value>, ApiError> {
2367 0 : let config = {
2368 0 : let locked = self.inner.read().unwrap();
2369 0 :
2370 0 : match locked
2371 0 : .tenants
2372 0 : .range(TenantShardId::tenant_range(tenant_id))
2373 0 : .next()
2374 : {
2375 0 : Some((_tenant_shard_id, shard)) => shard.config.clone(),
2376 : None => {
2377 0 : return Err(ApiError::NotFound(
2378 0 : anyhow::anyhow!("Tenant not found").into(),
2379 0 : ))
2380 : }
2381 : }
2382 : };
2383 :
2384 : // Unlike the pageserver, we do not have a set of global defaults: the config is
2385 : // entirely per-tenant. Therefore the distinction between `tenant_specific_overrides`
2386 : // and `effective_config` in the response is meaningless, but we retain that syntax
2387 : // in order to remain compatible with the pageserver API.
2388 :
2389 0 : let response = HashMap::from([
2390 : (
2391 : "tenant_specific_overrides",
2392 0 : serde_json::to_value(&config)
2393 0 : .context("serializing tenant specific overrides")
2394 0 : .map_err(ApiError::InternalServerError)?,
2395 : ),
2396 : (
2397 0 : "effective_config",
2398 0 : serde_json::to_value(&config)
2399 0 : .context("serializing effective config")
2400 0 : .map_err(ApiError::InternalServerError)?,
2401 : ),
2402 : ]);
2403 :
2404 0 : Ok(response)
2405 0 : }
2406 :
2407 0 : pub(crate) async fn tenant_time_travel_remote_storage(
2408 0 : &self,
2409 0 : time_travel_req: &TenantTimeTravelRequest,
2410 0 : tenant_id: TenantId,
2411 0 : timestamp: Cow<'_, str>,
2412 0 : done_if_after: Cow<'_, str>,
2413 0 : ) -> Result<(), ApiError> {
2414 0 : let _tenant_lock = trace_exclusive_lock(
2415 0 : &self.tenant_op_locks,
2416 0 : tenant_id,
2417 0 : TenantOperations::TimeTravelRemoteStorage,
2418 0 : )
2419 0 : .await;
2420 :
2421 0 : let node = {
2422 0 : let locked = self.inner.read().unwrap();
2423 : // Just a sanity check to prevent misuse: the API expects that the tenant is fully
2424 : // detached everywhere, and nothing writes to S3 storage. Here, we verify that,
2425 : // but only at the start of the process, so it's really just to prevent operator
2426 : // mistakes.
2427 0 : for (shard_id, shard) in locked.tenants.range(TenantShardId::tenant_range(tenant_id)) {
2428 0 : if shard.intent.get_attached().is_some() || !shard.intent.get_secondary().is_empty()
2429 : {
2430 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
2431 0 : "We want tenant to be attached in shard with tenant_shard_id={shard_id}"
2432 0 : )));
2433 0 : }
2434 0 : let maybe_attached = shard
2435 0 : .observed
2436 0 : .locations
2437 0 : .iter()
2438 0 : .filter_map(|(node_id, observed_location)| {
2439 0 : observed_location
2440 0 : .conf
2441 0 : .as_ref()
2442 0 : .map(|loc| (node_id, observed_location, loc.mode))
2443 0 : })
2444 0 : .find(|(_, _, mode)| *mode != LocationConfigMode::Detached);
2445 0 : if let Some((node_id, _observed_location, mode)) = maybe_attached {
2446 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!("We observed attached={mode:?} tenant in node_id={node_id} shard with tenant_shard_id={shard_id}")));
2447 0 : }
2448 : }
2449 0 : let scheduler = &locked.scheduler;
2450 : // Right now we only perform the operation on a single node without parallelization
2451 : // TODO fan out the operation to multiple nodes for better performance
2452 0 : let node_id = scheduler.schedule_shard(&[], &ScheduleContext::default())?;
2453 0 : let node = locked
2454 0 : .nodes
2455 0 : .get(&node_id)
2456 0 : .expect("Pageservers may not be deleted while lock is active");
2457 0 : node.clone()
2458 0 : };
2459 0 :
2460 0 : // The shard count is encoded in the remote storage's URL, so we need to handle all historically used shard counts
2461 0 : let mut counts = time_travel_req
2462 0 : .shard_counts
2463 0 : .iter()
2464 0 : .copied()
2465 0 : .collect::<HashSet<_>>()
2466 0 : .into_iter()
2467 0 : .collect::<Vec<_>>();
2468 0 : counts.sort_unstable();
2469 :
2470 0 : for count in counts {
2471 0 : let shard_ids = (0..count.count())
2472 0 : .map(|i| TenantShardId {
2473 0 : tenant_id,
2474 0 : shard_number: ShardNumber(i),
2475 0 : shard_count: count,
2476 0 : })
2477 0 : .collect::<Vec<_>>();
2478 0 : for tenant_shard_id in shard_ids {
2479 0 : let client = PageserverClient::new(
2480 0 : node.get_id(),
2481 0 : node.base_url(),
2482 0 : self.config.jwt_token.as_deref(),
2483 0 : );
2484 0 :
2485 0 : tracing::info!("Doing time travel recovery for shard {tenant_shard_id}",);
2486 :
2487 0 : client
2488 0 : .tenant_time_travel_remote_storage(
2489 0 : tenant_shard_id,
2490 0 : ×tamp,
2491 0 : &done_if_after,
2492 0 : )
2493 0 : .await
2494 0 : .map_err(|e| {
2495 0 : ApiError::InternalServerError(anyhow::anyhow!(
2496 0 : "Error doing time travel recovery for shard {tenant_shard_id} on node {}: {e}",
2497 0 : node
2498 0 : ))
2499 0 : })?;
2500 : }
2501 : }
2502 0 : Ok(())
2503 0 : }
2504 :
2505 0 : pub(crate) async fn tenant_secondary_download(
2506 0 : &self,
2507 0 : tenant_id: TenantId,
2508 0 : wait: Option<Duration>,
2509 0 : ) -> Result<(StatusCode, SecondaryProgress), ApiError> {
2510 0 : let _tenant_lock = trace_shared_lock(
2511 0 : &self.tenant_op_locks,
2512 0 : tenant_id,
2513 0 : TenantOperations::SecondaryDownload,
2514 0 : )
2515 0 : .await;
2516 :
2517 : // Acquire lock and yield the collection of shard-node tuples which we will send requests onward to
2518 0 : let targets = {
2519 0 : let locked = self.inner.read().unwrap();
2520 0 : let mut targets = Vec::new();
2521 :
2522 0 : for (tenant_shard_id, shard) in
2523 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
2524 : {
2525 0 : for node_id in shard.intent.get_secondary() {
2526 0 : let node = locked
2527 0 : .nodes
2528 0 : .get(node_id)
2529 0 : .expect("Pageservers may not be deleted while referenced");
2530 0 :
2531 0 : targets.push((*tenant_shard_id, node.clone()));
2532 0 : }
2533 : }
2534 0 : targets
2535 0 : };
2536 0 :
2537 0 : // Issue concurrent requests to all shards' locations
2538 0 : let mut futs = FuturesUnordered::new();
2539 0 : for (tenant_shard_id, node) in targets {
2540 0 : let client = PageserverClient::new(
2541 0 : node.get_id(),
2542 0 : node.base_url(),
2543 0 : self.config.jwt_token.as_deref(),
2544 0 : );
2545 0 : futs.push(async move {
2546 0 : let result = client
2547 0 : .tenant_secondary_download(tenant_shard_id, wait)
2548 0 : .await;
2549 0 : (result, node, tenant_shard_id)
2550 0 : })
2551 : }
2552 :
2553 : // Handle any errors returned by pageservers. This includes cases like this request racing with
2554 : // a scheduling operation, such that the tenant shard we're calling doesn't exist on that pageserver any more, as
2555 : // well as more general cases like 503s, 500s, or timeouts.
2556 0 : let mut aggregate_progress = SecondaryProgress::default();
2557 0 : let mut aggregate_status: Option<StatusCode> = None;
2558 0 : let mut error: Option<mgmt_api::Error> = None;
2559 0 : while let Some((result, node, tenant_shard_id)) = futs.next().await {
2560 0 : match result {
2561 0 : Err(e) => {
2562 0 : // Secondary downloads are always advisory: if something fails, we nevertheless report success, so that whoever
2563 0 : // is calling us will proceed with whatever migration they're doing, albeit with a slightly less warm cache
2564 0 : // than they had hoped for.
2565 0 : tracing::warn!("Secondary download error from pageserver {node}: {e}",);
2566 0 : error = Some(e)
2567 : }
2568 0 : Ok((status_code, progress)) => {
2569 0 : tracing::info!(%tenant_shard_id, "Shard status={status_code} progress: {progress:?}");
2570 0 : aggregate_progress.layers_downloaded += progress.layers_downloaded;
2571 0 : aggregate_progress.layers_total += progress.layers_total;
2572 0 : aggregate_progress.bytes_downloaded += progress.bytes_downloaded;
2573 0 : aggregate_progress.bytes_total += progress.bytes_total;
2574 0 : aggregate_progress.heatmap_mtime =
2575 0 : std::cmp::max(aggregate_progress.heatmap_mtime, progress.heatmap_mtime);
2576 0 : aggregate_status = match aggregate_status {
2577 0 : None => Some(status_code),
2578 0 : Some(StatusCode::OK) => Some(status_code),
2579 0 : Some(cur) => {
2580 0 : // Other status codes (e.g. 202) -- do not overwrite.
2581 0 : Some(cur)
2582 : }
2583 : };
2584 : }
2585 : }
2586 : }
2587 :
2588 : // If any of the shards return 202, indicate our result as 202.
2589 0 : match aggregate_status {
2590 : None => {
2591 0 : match error {
2592 0 : Some(e) => {
2593 0 : // No successes, and an error: surface it
2594 0 : Err(ApiError::Conflict(format!("Error from pageserver: {e}")))
2595 : }
2596 : None => {
2597 : // No shards found
2598 0 : Err(ApiError::NotFound(
2599 0 : anyhow::anyhow!("Tenant {} not found", tenant_id).into(),
2600 0 : ))
2601 : }
2602 : }
2603 : }
2604 0 : Some(aggregate_status) => Ok((aggregate_status, aggregate_progress)),
2605 : }
2606 0 : }
2607 :
2608 0 : pub(crate) async fn tenant_delete(&self, tenant_id: TenantId) -> Result<StatusCode, ApiError> {
2609 0 : let _tenant_lock =
2610 0 : trace_exclusive_lock(&self.tenant_op_locks, tenant_id, TenantOperations::Delete).await;
2611 :
2612 : // Detach all shards
2613 0 : let (detach_waiters, shard_ids, node) = {
2614 0 : let mut shard_ids = Vec::new();
2615 0 : let mut detach_waiters = Vec::new();
2616 0 : let mut locked = self.inner.write().unwrap();
2617 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
2618 0 : for (tenant_shard_id, shard) in
2619 0 : tenants.range_mut(TenantShardId::tenant_range(tenant_id))
2620 : {
2621 0 : shard_ids.push(*tenant_shard_id);
2622 0 :
2623 0 : // Update the tenant's intent to remove all attachments
2624 0 : shard.policy = PlacementPolicy::Detached;
2625 0 : shard
2626 0 : .schedule(scheduler, &mut ScheduleContext::default())
2627 0 : .expect("De-scheduling is infallible");
2628 0 : debug_assert!(shard.intent.get_attached().is_none());
2629 0 : debug_assert!(shard.intent.get_secondary().is_empty());
2630 :
2631 0 : if let Some(waiter) = self.maybe_reconcile_shard(shard, nodes) {
2632 0 : detach_waiters.push(waiter);
2633 0 : }
2634 : }
2635 :
2636 : // Pick an arbitrary node to use for remote deletions (does not have to be where the tenant
2637 : // was attached, just has to be able to see the S3 content)
2638 0 : let node_id = scheduler.schedule_shard(&[], &ScheduleContext::default())?;
2639 0 : let node = nodes
2640 0 : .get(&node_id)
2641 0 : .expect("Pageservers may not be deleted while lock is active");
2642 0 : (detach_waiters, shard_ids, node.clone())
2643 0 : };
2644 0 :
2645 0 : // This reconcile wait can fail in a few ways:
2646 0 : // A there is a very long queue for the reconciler semaphore
2647 0 : // B some pageserver is failing to handle a detach promptly
2648 0 : // C some pageserver goes offline right at the moment we send it a request.
2649 0 : //
2650 0 : // A and C are transient: the semaphore will eventually become available, and once a node is marked offline
2651 0 : // the next attempt to reconcile will silently skip detaches for an offline node and succeed. If B happens,
2652 0 : // it's a bug, and needs resolving at the pageserver level (we shouldn't just leave attachments behind while
2653 0 : // deleting the underlying data).
2654 0 : self.await_waiters(detach_waiters, RECONCILE_TIMEOUT)
2655 0 : .await?;
2656 :
2657 0 : let locations = shard_ids
2658 0 : .into_iter()
2659 0 : .map(|s| (s, node.clone()))
2660 0 : .collect::<Vec<_>>();
2661 0 : let results = self.tenant_for_shards_api(
2662 0 : locations,
2663 0 : |tenant_shard_id, client| async move { client.tenant_delete(tenant_shard_id).await },
2664 0 : 1,
2665 0 : 3,
2666 0 : RECONCILE_TIMEOUT,
2667 0 : &self.cancel,
2668 0 : )
2669 0 : .await;
2670 0 : for result in results {
2671 0 : match result {
2672 : Ok(StatusCode::ACCEPTED) => {
2673 : // This should never happen: we waited for detaches to finish above
2674 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
2675 0 : "Unexpectedly still attached on {}",
2676 0 : node
2677 0 : )));
2678 : }
2679 0 : Ok(_) => {}
2680 : Err(mgmt_api::Error::Cancelled) => {
2681 0 : return Err(ApiError::ShuttingDown);
2682 : }
2683 0 : Err(e) => {
2684 0 : // This is unexpected: remote deletion should be infallible, unless the object store
2685 0 : // at large is unavailable.
2686 0 : tracing::error!("Error deleting via node {}: {e}", node);
2687 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(e)));
2688 : }
2689 : }
2690 : }
2691 :
2692 : // Fall through: deletion of the tenant on pageservers is complete, we may proceed to drop
2693 : // our in-memory state and database state.
2694 :
2695 : // Ordering: we delete persistent state first: if we then
2696 : // crash, we will drop the in-memory state.
2697 :
2698 : // Drop persistent state.
2699 0 : self.persistence.delete_tenant(tenant_id).await?;
2700 :
2701 : // Drop in-memory state
2702 : {
2703 0 : let mut locked = self.inner.write().unwrap();
2704 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
2705 :
2706 : // Dereference Scheduler from shards before dropping them
2707 0 : for (_tenant_shard_id, shard) in
2708 0 : tenants.range_mut(TenantShardId::tenant_range(tenant_id))
2709 0 : {
2710 0 : shard.intent.clear(scheduler);
2711 0 : }
2712 :
2713 0 : tenants.retain(|tenant_shard_id, _shard| tenant_shard_id.tenant_id != tenant_id);
2714 0 : tracing::info!(
2715 0 : "Deleted tenant {tenant_id}, now have {} tenants",
2716 0 : locked.tenants.len()
2717 : );
2718 : };
2719 :
2720 : // Success is represented as 404, to imitate the existing pageserver deletion API
2721 0 : Ok(StatusCode::NOT_FOUND)
2722 0 : }
2723 :
2724 : /// Naming: this configures the storage controller's policies for a tenant, whereas [`Self::tenant_config_set`] is "set the TenantConfig"
2725 : /// for a tenant. The TenantConfig is passed through to pageservers, whereas this function modifies
2726 : /// the tenant's policies (configuration) within the storage controller
2727 0 : pub(crate) async fn tenant_update_policy(
2728 0 : &self,
2729 0 : tenant_id: TenantId,
2730 0 : req: TenantPolicyRequest,
2731 0 : ) -> Result<(), ApiError> {
2732 : // We require an exclusive lock, because we are updating persistent and in-memory state
2733 0 : let _tenant_lock = trace_exclusive_lock(
2734 0 : &self.tenant_op_locks,
2735 0 : tenant_id,
2736 0 : TenantOperations::UpdatePolicy,
2737 0 : )
2738 0 : .await;
2739 :
2740 0 : failpoint_support::sleep_millis_async!("tenant-update-policy-exclusive-lock");
2741 :
2742 : let TenantPolicyRequest {
2743 0 : placement,
2744 0 : scheduling,
2745 0 : } = req;
2746 0 :
2747 0 : self.persistence
2748 0 : .update_tenant_shard(
2749 0 : TenantFilter::Tenant(tenant_id),
2750 0 : placement.clone(),
2751 0 : None,
2752 0 : None,
2753 0 : scheduling,
2754 0 : )
2755 0 : .await?;
2756 :
2757 0 : let mut schedule_context = ScheduleContext::default();
2758 0 : let mut locked = self.inner.write().unwrap();
2759 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
2760 0 : for (shard_id, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
2761 0 : if let Some(placement) = &placement {
2762 0 : shard.policy = placement.clone();
2763 0 :
2764 0 : tracing::info!(tenant_id=%shard_id.tenant_id, shard_id=%shard_id.shard_slug(),
2765 0 : "Updated placement policy to {placement:?}");
2766 0 : }
2767 :
2768 0 : if let Some(scheduling) = &scheduling {
2769 0 : shard.set_scheduling_policy(*scheduling);
2770 0 :
2771 0 : tracing::info!(tenant_id=%shard_id.tenant_id, shard_id=%shard_id.shard_slug(),
2772 0 : "Updated scheduling policy to {scheduling:?}");
2773 0 : }
2774 :
2775 : // In case scheduling is being switched back on, try it now.
2776 0 : shard.schedule(scheduler, &mut schedule_context).ok();
2777 0 : self.maybe_reconcile_shard(shard, nodes);
2778 : }
2779 :
2780 0 : Ok(())
2781 0 : }
2782 :
2783 0 : pub(crate) async fn tenant_timeline_create(
2784 0 : &self,
2785 0 : tenant_id: TenantId,
2786 0 : mut create_req: TimelineCreateRequest,
2787 0 : ) -> Result<TimelineInfo, ApiError> {
2788 0 : tracing::info!(
2789 0 : "Creating timeline {}/{}",
2790 : tenant_id,
2791 : create_req.new_timeline_id,
2792 : );
2793 :
2794 0 : let _tenant_lock = trace_shared_lock(
2795 0 : &self.tenant_op_locks,
2796 0 : tenant_id,
2797 0 : TenantOperations::TimelineCreate,
2798 0 : )
2799 0 : .await;
2800 0 : failpoint_support::sleep_millis_async!("tenant-create-timeline-shared-lock");
2801 :
2802 0 : self.ensure_attached_wait(tenant_id).await?;
2803 :
2804 0 : let mut targets = {
2805 0 : let locked = self.inner.read().unwrap();
2806 0 : let mut targets = Vec::new();
2807 :
2808 0 : for (tenant_shard_id, shard) in
2809 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
2810 0 : {
2811 0 : let node_id = shard.intent.get_attached().ok_or_else(|| {
2812 0 : ApiError::InternalServerError(anyhow::anyhow!("Shard not scheduled"))
2813 0 : })?;
2814 0 : let node = locked
2815 0 : .nodes
2816 0 : .get(&node_id)
2817 0 : .expect("Pageservers may not be deleted while referenced");
2818 0 :
2819 0 : targets.push((*tenant_shard_id, node.clone()));
2820 : }
2821 0 : targets
2822 0 : };
2823 0 :
2824 0 : if targets.is_empty() {
2825 0 : return Err(ApiError::NotFound(
2826 0 : anyhow::anyhow!("Tenant not found").into(),
2827 0 : ));
2828 0 : };
2829 0 : let shard_zero = targets.remove(0);
2830 :
2831 0 : async fn create_one(
2832 0 : tenant_shard_id: TenantShardId,
2833 0 : node: Node,
2834 0 : jwt: Option<String>,
2835 0 : create_req: TimelineCreateRequest,
2836 0 : ) -> Result<TimelineInfo, ApiError> {
2837 0 : tracing::info!(
2838 0 : "Creating timeline on shard {}/{}, attached to node {node}",
2839 : tenant_shard_id,
2840 : create_req.new_timeline_id,
2841 : );
2842 0 : let client = PageserverClient::new(node.get_id(), node.base_url(), jwt.as_deref());
2843 0 :
2844 0 : client
2845 0 : .timeline_create(tenant_shard_id, &create_req)
2846 0 : .await
2847 0 : .map_err(|e| passthrough_api_error(&node, e))
2848 0 : }
2849 :
2850 : // Because the caller might not provide an explicit LSN, we must do the creation first on a single shard, and then
2851 : // use whatever LSN that shard picked when creating on subsequent shards. We arbitrarily use shard zero as the shard
2852 : // that will get the first creation request, and propagate the LSN to all the >0 shards.
2853 0 : let timeline_info = create_one(
2854 0 : shard_zero.0,
2855 0 : shard_zero.1,
2856 0 : self.config.jwt_token.clone(),
2857 0 : create_req.clone(),
2858 0 : )
2859 0 : .await?;
2860 :
2861 : // Propagate the LSN that shard zero picked, if caller didn't provide one
2862 0 : if create_req.ancestor_timeline_id.is_some() && create_req.ancestor_start_lsn.is_none() {
2863 0 : create_req.ancestor_start_lsn = timeline_info.ancestor_lsn;
2864 0 : }
2865 :
2866 : // Create timeline on remaining shards with number >0
2867 0 : if !targets.is_empty() {
2868 : // If we had multiple shards, issue requests for the remainder now.
2869 0 : let jwt = &self.config.jwt_token;
2870 0 : self.tenant_for_shards(targets, |tenant_shard_id: TenantShardId, node: Node| {
2871 0 : let create_req = create_req.clone();
2872 0 : Box::pin(create_one(tenant_shard_id, node, jwt.clone(), create_req))
2873 0 : })
2874 0 : .await?;
2875 0 : }
2876 :
2877 0 : Ok(timeline_info)
2878 0 : }
2879 :
2880 0 : pub(crate) async fn tenant_timeline_detach_ancestor(
2881 0 : &self,
2882 0 : tenant_id: TenantId,
2883 0 : timeline_id: TimelineId,
2884 0 : ) -> Result<models::detach_ancestor::AncestorDetached, ApiError> {
2885 0 : tracing::info!("Detaching timeline {tenant_id}/{timeline_id}",);
2886 :
2887 0 : let _tenant_lock = trace_shared_lock(
2888 0 : &self.tenant_op_locks,
2889 0 : tenant_id,
2890 0 : TenantOperations::TimelineDetachAncestor,
2891 0 : )
2892 0 : .await;
2893 :
2894 0 : self.ensure_attached_wait(tenant_id).await?;
2895 :
2896 0 : let targets = {
2897 0 : let locked = self.inner.read().unwrap();
2898 0 : let mut targets = Vec::new();
2899 :
2900 0 : for (tenant_shard_id, shard) in
2901 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
2902 0 : {
2903 0 : let node_id = shard.intent.get_attached().ok_or_else(|| {
2904 0 : ApiError::InternalServerError(anyhow::anyhow!("Shard not scheduled"))
2905 0 : })?;
2906 0 : let node = locked
2907 0 : .nodes
2908 0 : .get(&node_id)
2909 0 : .expect("Pageservers may not be deleted while referenced");
2910 0 :
2911 0 : targets.push((*tenant_shard_id, node.clone()));
2912 : }
2913 0 : targets
2914 0 : };
2915 0 :
2916 0 : if targets.is_empty() {
2917 0 : return Err(ApiError::NotFound(
2918 0 : anyhow::anyhow!("Tenant not found").into(),
2919 0 : ));
2920 0 : }
2921 :
2922 0 : async fn detach_one(
2923 0 : tenant_shard_id: TenantShardId,
2924 0 : timeline_id: TimelineId,
2925 0 : node: Node,
2926 0 : jwt: Option<String>,
2927 0 : ) -> Result<(ShardNumber, models::detach_ancestor::AncestorDetached), ApiError> {
2928 0 : tracing::info!(
2929 0 : "Detaching timeline on shard {tenant_shard_id}/{timeline_id}, attached to node {node}",
2930 : );
2931 :
2932 0 : let client = PageserverClient::new(node.get_id(), node.base_url(), jwt.as_deref());
2933 0 : client
2934 0 : .timeline_detach_ancestor(tenant_shard_id, timeline_id)
2935 0 : .await
2936 0 : .map_err(|e| {
2937 0 : use mgmt_api::Error;
2938 0 :
2939 0 : match e {
2940 : // no ancestor (ever)
2941 0 : Error::ApiError(StatusCode::CONFLICT, msg) => ApiError::Conflict(format!(
2942 0 : "{node}: {}",
2943 0 : msg.strip_prefix("Conflict: ").unwrap_or(&msg)
2944 0 : )),
2945 : // too many ancestors
2946 0 : Error::ApiError(StatusCode::BAD_REQUEST, msg) => {
2947 0 : ApiError::BadRequest(anyhow::anyhow!("{node}: {msg}"))
2948 : }
2949 : // rest can be mapped
2950 0 : other => passthrough_api_error(&node, other),
2951 : }
2952 0 : })
2953 0 : .map(|res| (tenant_shard_id.shard_number, res))
2954 0 : }
2955 :
2956 : // no shard needs to go first/last; the operation should be idempotent
2957 : // TODO: it would be great to ensure that all shards return the same error
2958 0 : let mut results = self
2959 0 : .tenant_for_shards(targets, |tenant_shard_id, node| {
2960 0 : futures::FutureExt::boxed(detach_one(
2961 0 : tenant_shard_id,
2962 0 : timeline_id,
2963 0 : node,
2964 0 : self.config.jwt_token.clone(),
2965 0 : ))
2966 0 : })
2967 0 : .await?;
2968 :
2969 0 : let any = results.pop().expect("we must have at least one response");
2970 0 :
2971 0 : let mismatching = results
2972 0 : .iter()
2973 0 : .filter(|(_, res)| res != &any.1)
2974 0 : .collect::<Vec<_>>();
2975 0 : if !mismatching.is_empty() {
2976 0 : let matching = results.len() - mismatching.len();
2977 0 : tracing::error!(
2978 : matching,
2979 : compared_against=?any,
2980 : ?mismatching,
2981 0 : "shards returned different results"
2982 : );
2983 0 : }
2984 :
2985 0 : Ok(any.1)
2986 0 : }
2987 :
2988 : /// Helper for concurrently calling a pageserver API on a number of shards, such as timeline creation.
2989 : ///
2990 : /// On success, the returned vector contains exactly the same number of elements as the input `locations`.
2991 0 : async fn tenant_for_shards<F, R>(
2992 0 : &self,
2993 0 : locations: Vec<(TenantShardId, Node)>,
2994 0 : mut req_fn: F,
2995 0 : ) -> Result<Vec<R>, ApiError>
2996 0 : where
2997 0 : F: FnMut(
2998 0 : TenantShardId,
2999 0 : Node,
3000 0 : )
3001 0 : -> std::pin::Pin<Box<dyn futures::Future<Output = Result<R, ApiError>> + Send>>,
3002 0 : {
3003 0 : let mut futs = FuturesUnordered::new();
3004 0 : let mut results = Vec::with_capacity(locations.len());
3005 :
3006 0 : for (tenant_shard_id, node) in locations {
3007 0 : futs.push(req_fn(tenant_shard_id, node));
3008 0 : }
3009 :
3010 0 : while let Some(r) = futs.next().await {
3011 0 : results.push(r?);
3012 : }
3013 :
3014 0 : Ok(results)
3015 0 : }
3016 :
3017 : /// Concurrently invoke a pageserver API call on many shards at once
3018 0 : pub(crate) async fn tenant_for_shards_api<T, O, F>(
3019 0 : &self,
3020 0 : locations: Vec<(TenantShardId, Node)>,
3021 0 : op: O,
3022 0 : warn_threshold: u32,
3023 0 : max_retries: u32,
3024 0 : timeout: Duration,
3025 0 : cancel: &CancellationToken,
3026 0 : ) -> Vec<mgmt_api::Result<T>>
3027 0 : where
3028 0 : O: Fn(TenantShardId, PageserverClient) -> F + Copy,
3029 0 : F: std::future::Future<Output = mgmt_api::Result<T>>,
3030 0 : {
3031 0 : let mut futs = FuturesUnordered::new();
3032 0 : let mut results = Vec::with_capacity(locations.len());
3033 :
3034 0 : for (tenant_shard_id, node) in locations {
3035 0 : futs.push(async move {
3036 0 : node.with_client_retries(
3037 0 : |client| op(tenant_shard_id, client),
3038 0 : &self.config.jwt_token,
3039 0 : warn_threshold,
3040 0 : max_retries,
3041 0 : timeout,
3042 0 : cancel,
3043 0 : )
3044 0 : .await
3045 0 : });
3046 0 : }
3047 :
3048 0 : while let Some(r) = futs.next().await {
3049 0 : let r = r.unwrap_or(Err(mgmt_api::Error::Cancelled));
3050 0 : results.push(r);
3051 0 : }
3052 :
3053 0 : results
3054 0 : }
3055 :
3056 0 : pub(crate) async fn tenant_timeline_delete(
3057 0 : &self,
3058 0 : tenant_id: TenantId,
3059 0 : timeline_id: TimelineId,
3060 0 : ) -> Result<StatusCode, ApiError> {
3061 0 : tracing::info!("Deleting timeline {}/{}", tenant_id, timeline_id,);
3062 0 : let _tenant_lock = trace_shared_lock(
3063 0 : &self.tenant_op_locks,
3064 0 : tenant_id,
3065 0 : TenantOperations::TimelineDelete,
3066 0 : )
3067 0 : .await;
3068 :
3069 0 : self.ensure_attached_wait(tenant_id).await?;
3070 :
3071 0 : let mut targets = {
3072 0 : let locked = self.inner.read().unwrap();
3073 0 : let mut targets = Vec::new();
3074 :
3075 0 : for (tenant_shard_id, shard) in
3076 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
3077 0 : {
3078 0 : let node_id = shard.intent.get_attached().ok_or_else(|| {
3079 0 : ApiError::InternalServerError(anyhow::anyhow!("Shard not scheduled"))
3080 0 : })?;
3081 0 : let node = locked
3082 0 : .nodes
3083 0 : .get(&node_id)
3084 0 : .expect("Pageservers may not be deleted while referenced");
3085 0 :
3086 0 : targets.push((*tenant_shard_id, node.clone()));
3087 : }
3088 0 : targets
3089 0 : };
3090 0 :
3091 0 : if targets.is_empty() {
3092 0 : return Err(ApiError::NotFound(
3093 0 : anyhow::anyhow!("Tenant not found").into(),
3094 0 : ));
3095 0 : }
3096 0 : let shard_zero = targets.remove(0);
3097 :
3098 0 : async fn delete_one(
3099 0 : tenant_shard_id: TenantShardId,
3100 0 : timeline_id: TimelineId,
3101 0 : node: Node,
3102 0 : jwt: Option<String>,
3103 0 : ) -> Result<StatusCode, ApiError> {
3104 0 : tracing::info!(
3105 0 : "Deleting timeline on shard {tenant_shard_id}/{timeline_id}, attached to node {node}",
3106 : );
3107 :
3108 0 : let client = PageserverClient::new(node.get_id(), node.base_url(), jwt.as_deref());
3109 0 : client
3110 0 : .timeline_delete(tenant_shard_id, timeline_id)
3111 0 : .await
3112 0 : .map_err(|e| {
3113 0 : ApiError::InternalServerError(anyhow::anyhow!(
3114 0 : "Error deleting timeline {timeline_id} on {tenant_shard_id} on node {node}: {e}",
3115 0 : ))
3116 0 : })
3117 0 : }
3118 :
3119 0 : let statuses = self
3120 0 : .tenant_for_shards(targets, |tenant_shard_id: TenantShardId, node: Node| {
3121 0 : Box::pin(delete_one(
3122 0 : tenant_shard_id,
3123 0 : timeline_id,
3124 0 : node,
3125 0 : self.config.jwt_token.clone(),
3126 0 : ))
3127 0 : })
3128 0 : .await?;
3129 :
3130 : // If any shards >0 haven't finished deletion yet, don't start deletion on shard zero
3131 0 : if statuses.iter().any(|s| s != &StatusCode::NOT_FOUND) {
3132 0 : return Ok(StatusCode::ACCEPTED);
3133 0 : }
3134 :
3135 : // Delete shard zero last: this is not strictly necessary, but since a caller's GET on a timeline will be routed
3136 : // to shard zero, it gives a more obvious behavior that a GET returns 404 once the deletion is done.
3137 0 : let shard_zero_status = delete_one(
3138 0 : shard_zero.0,
3139 0 : timeline_id,
3140 0 : shard_zero.1,
3141 0 : self.config.jwt_token.clone(),
3142 0 : )
3143 0 : .await?;
3144 :
3145 0 : Ok(shard_zero_status)
3146 0 : }
3147 :
3148 : /// When you need to send an HTTP request to the pageserver that holds shard0 of a tenant, this
3149 : /// function looks up and returns node. If the tenant isn't found, returns Err(ApiError::NotFound)
3150 0 : pub(crate) fn tenant_shard0_node(
3151 0 : &self,
3152 0 : tenant_id: TenantId,
3153 0 : ) -> Result<(Node, TenantShardId), ApiError> {
3154 0 : let locked = self.inner.read().unwrap();
3155 0 : let Some((tenant_shard_id, shard)) = locked
3156 0 : .tenants
3157 0 : .range(TenantShardId::tenant_range(tenant_id))
3158 0 : .next()
3159 : else {
3160 0 : return Err(ApiError::NotFound(
3161 0 : anyhow::anyhow!("Tenant {tenant_id} not found").into(),
3162 0 : ));
3163 : };
3164 :
3165 : // TODO: should use the ID last published to compute_hook, rather than the intent: the intent might
3166 : // point to somewhere we haven't attached yet.
3167 0 : let Some(node_id) = shard.intent.get_attached() else {
3168 0 : tracing::warn!(
3169 0 : tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(),
3170 0 : "Shard not scheduled (policy {:?}), cannot generate pass-through URL",
3171 : shard.policy
3172 : );
3173 0 : return Err(ApiError::Conflict(
3174 0 : "Cannot call timeline API on non-attached tenant".to_string(),
3175 0 : ));
3176 : };
3177 :
3178 0 : let Some(node) = locked.nodes.get(node_id) else {
3179 : // This should never happen
3180 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
3181 0 : "Shard refers to nonexistent node"
3182 0 : )));
3183 : };
3184 :
3185 0 : Ok((node.clone(), *tenant_shard_id))
3186 0 : }
3187 :
3188 0 : pub(crate) fn tenant_locate(
3189 0 : &self,
3190 0 : tenant_id: TenantId,
3191 0 : ) -> Result<TenantLocateResponse, ApiError> {
3192 0 : let locked = self.inner.read().unwrap();
3193 0 : tracing::info!("Locating shards for tenant {tenant_id}");
3194 :
3195 0 : let mut result = Vec::new();
3196 0 : let mut shard_params: Option<ShardParameters> = None;
3197 :
3198 0 : for (tenant_shard_id, shard) in locked.tenants.range(TenantShardId::tenant_range(tenant_id))
3199 : {
3200 0 : let node_id =
3201 0 : shard
3202 0 : .intent
3203 0 : .get_attached()
3204 0 : .ok_or(ApiError::BadRequest(anyhow::anyhow!(
3205 0 : "Cannot locate a tenant that is not attached"
3206 0 : )))?;
3207 :
3208 0 : let node = locked
3209 0 : .nodes
3210 0 : .get(&node_id)
3211 0 : .expect("Pageservers may not be deleted while referenced");
3212 0 :
3213 0 : result.push(node.shard_location(*tenant_shard_id));
3214 0 :
3215 0 : match &shard_params {
3216 0 : None => {
3217 0 : shard_params = Some(ShardParameters {
3218 0 : stripe_size: shard.shard.stripe_size,
3219 0 : count: shard.shard.count,
3220 0 : });
3221 0 : }
3222 0 : Some(params) => {
3223 0 : if params.stripe_size != shard.shard.stripe_size {
3224 : // This should never happen. We enforce at runtime because it's simpler than
3225 : // adding an extra per-tenant data structure to store the things that should be the same
3226 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
3227 0 : "Inconsistent shard stripe size parameters!"
3228 0 : )));
3229 0 : }
3230 : }
3231 : }
3232 : }
3233 :
3234 0 : if result.is_empty() {
3235 0 : return Err(ApiError::NotFound(
3236 0 : anyhow::anyhow!("No shards for this tenant ID found").into(),
3237 0 : ));
3238 0 : }
3239 0 : let shard_params = shard_params.expect("result is non-empty, therefore this is set");
3240 0 : tracing::info!(
3241 0 : "Located tenant {} with params {:?} on shards {}",
3242 0 : tenant_id,
3243 0 : shard_params,
3244 0 : result
3245 0 : .iter()
3246 0 : .map(|s| format!("{:?}", s))
3247 0 : .collect::<Vec<_>>()
3248 0 : .join(",")
3249 : );
3250 :
3251 0 : Ok(TenantLocateResponse {
3252 0 : shards: result,
3253 0 : shard_params,
3254 0 : })
3255 0 : }
3256 :
3257 : /// Returns None if the input iterator of shards does not include a shard with number=0
3258 0 : fn tenant_describe_impl<'a>(
3259 0 : &self,
3260 0 : shards: impl Iterator<Item = &'a TenantShard>,
3261 0 : ) -> Option<TenantDescribeResponse> {
3262 0 : let mut shard_zero = None;
3263 0 : let mut describe_shards = Vec::new();
3264 :
3265 0 : for shard in shards {
3266 0 : if shard.tenant_shard_id.is_shard_zero() {
3267 0 : shard_zero = Some(shard);
3268 0 : }
3269 :
3270 0 : describe_shards.push(TenantDescribeResponseShard {
3271 0 : tenant_shard_id: shard.tenant_shard_id,
3272 0 : node_attached: *shard.intent.get_attached(),
3273 0 : node_secondary: shard.intent.get_secondary().to_vec(),
3274 0 : last_error: shard
3275 0 : .last_error
3276 0 : .lock()
3277 0 : .unwrap()
3278 0 : .as_ref()
3279 0 : .map(|e| format!("{e}"))
3280 0 : .unwrap_or("".to_string())
3281 0 : .clone(),
3282 0 : is_reconciling: shard.reconciler.is_some(),
3283 0 : is_pending_compute_notification: shard.pending_compute_notification,
3284 0 : is_splitting: matches!(shard.splitting, SplitState::Splitting),
3285 0 : scheduling_policy: *shard.get_scheduling_policy(),
3286 : })
3287 : }
3288 :
3289 0 : let shard_zero = shard_zero?;
3290 :
3291 0 : Some(TenantDescribeResponse {
3292 0 : tenant_id: shard_zero.tenant_shard_id.tenant_id,
3293 0 : shards: describe_shards,
3294 0 : stripe_size: shard_zero.shard.stripe_size,
3295 0 : policy: shard_zero.policy.clone(),
3296 0 : config: shard_zero.config.clone(),
3297 0 : })
3298 0 : }
3299 :
3300 0 : pub(crate) fn tenant_describe(
3301 0 : &self,
3302 0 : tenant_id: TenantId,
3303 0 : ) -> Result<TenantDescribeResponse, ApiError> {
3304 0 : let locked = self.inner.read().unwrap();
3305 0 :
3306 0 : self.tenant_describe_impl(
3307 0 : locked
3308 0 : .tenants
3309 0 : .range(TenantShardId::tenant_range(tenant_id))
3310 0 : .map(|(_k, v)| v),
3311 0 : )
3312 0 : .ok_or_else(|| ApiError::NotFound(anyhow::anyhow!("Tenant {tenant_id} not found").into()))
3313 0 : }
3314 :
3315 0 : pub(crate) fn tenant_list(&self) -> Vec<TenantDescribeResponse> {
3316 0 : let locked = self.inner.read().unwrap();
3317 0 :
3318 0 : let mut result = Vec::new();
3319 0 : for (_tenant_id, tenant_shards) in
3320 0 : &locked.tenants.iter().group_by(|(id, _shard)| id.tenant_id)
3321 0 : {
3322 0 : result.push(
3323 0 : self.tenant_describe_impl(tenant_shards.map(|(_k, v)| v))
3324 0 : .expect("Groups are always non-empty"),
3325 0 : );
3326 0 : }
3327 :
3328 0 : result
3329 0 : }
3330 :
3331 0 : #[instrument(skip_all, fields(tenant_id=%op.tenant_id))]
3332 : async fn abort_tenant_shard_split(
3333 : &self,
3334 : op: &TenantShardSplitAbort,
3335 : ) -> Result<(), TenantShardSplitAbortError> {
3336 : // Cleaning up a split:
3337 : // - Parent shards are not destroyed during a split, just detached.
3338 : // - Failed pageserver split API calls can leave the remote node with just the parent attached,
3339 : // just the children attached, or both.
3340 : //
3341 : // Therefore our work to do is to:
3342 : // 1. Clean up storage controller's internal state to just refer to parents, no children
3343 : // 2. Call out to pageservers to ensure that children are detached
3344 : // 3. Call out to pageservers to ensure that parents are attached.
3345 : //
3346 : // Crash safety:
3347 : // - If the storage controller stops running during this cleanup *after* clearing the splitting state
3348 : // from our database, then [`Self::startup_reconcile`] will regard child attachments as garbage
3349 : // and detach them.
3350 : // - TODO: If the storage controller stops running during this cleanup *before* clearing the splitting state
3351 : // from our database, then we will re-enter this cleanup routine on startup.
3352 :
3353 : let TenantShardSplitAbort {
3354 : tenant_id,
3355 : new_shard_count,
3356 : new_stripe_size,
3357 : ..
3358 : } = op;
3359 :
3360 : // First abort persistent state, if any exists.
3361 : match self
3362 : .persistence
3363 : .abort_shard_split(*tenant_id, *new_shard_count)
3364 : .await?
3365 : {
3366 : AbortShardSplitStatus::Aborted => {
3367 : // Proceed to roll back any child shards created on pageservers
3368 : }
3369 : AbortShardSplitStatus::Complete => {
3370 : // The split completed (we might hit that path if e.g. our database transaction
3371 : // to write the completion landed in the database, but we dropped connection
3372 : // before seeing the result).
3373 : //
3374 : // We must update in-memory state to reflect the successful split.
3375 : self.tenant_shard_split_commit_inmem(
3376 : *tenant_id,
3377 : *new_shard_count,
3378 : *new_stripe_size,
3379 : );
3380 : return Ok(());
3381 : }
3382 : }
3383 :
3384 : // Clean up in-memory state, and accumulate the list of child locations that need detaching
3385 : let detach_locations: Vec<(Node, TenantShardId)> = {
3386 : let mut detach_locations = Vec::new();
3387 : let mut locked = self.inner.write().unwrap();
3388 : let (nodes, tenants, scheduler) = locked.parts_mut();
3389 :
3390 : for (tenant_shard_id, shard) in
3391 : tenants.range_mut(TenantShardId::tenant_range(op.tenant_id))
3392 : {
3393 : if shard.shard.count == op.new_shard_count {
3394 : // Surprising: the phase of [`Self::do_tenant_shard_split`] which inserts child shards in-memory
3395 : // is infallible, so if we got an error we shouldn't have got that far.
3396 : tracing::warn!(
3397 : "During split abort, child shard {tenant_shard_id} found in-memory"
3398 : );
3399 : continue;
3400 : }
3401 :
3402 : // Add the children of this shard to this list of things to detach
3403 : if let Some(node_id) = shard.intent.get_attached() {
3404 : for child_id in tenant_shard_id.split(*new_shard_count) {
3405 : detach_locations.push((
3406 : nodes
3407 : .get(node_id)
3408 : .expect("Intent references nonexistent node")
3409 : .clone(),
3410 : child_id,
3411 : ));
3412 : }
3413 : } else {
3414 : tracing::warn!(
3415 : "During split abort, shard {tenant_shard_id} has no attached location"
3416 : );
3417 : }
3418 :
3419 : tracing::info!("Restoring parent shard {tenant_shard_id}");
3420 : shard.splitting = SplitState::Idle;
3421 : if let Err(e) = shard.schedule(scheduler, &mut ScheduleContext::default()) {
3422 : // If this shard can't be scheduled now (perhaps due to offline nodes or
3423 : // capacity issues), that must not prevent us rolling back a split. In this
3424 : // case it should be eventually scheduled in the background.
3425 : tracing::warn!("Failed to schedule {tenant_shard_id} during shard abort: {e}")
3426 : }
3427 :
3428 : self.maybe_reconcile_shard(shard, nodes);
3429 : }
3430 :
3431 : // We don't expect any new_shard_count shards to exist here, but drop them just in case
3432 0 : tenants.retain(|_id, s| s.shard.count != *new_shard_count);
3433 :
3434 : detach_locations
3435 : };
3436 :
3437 : for (node, child_id) in detach_locations {
3438 : if !node.is_available() {
3439 : // An unavailable node cannot be cleaned up now: to avoid blocking forever, we will permit this, and
3440 : // rely on the reconciliation that happens when a node transitions to Active to clean up. Since we have
3441 : // removed child shards from our in-memory state and database, the reconciliation will implicitly remove
3442 : // them from the node.
3443 : tracing::warn!("Node {node} unavailable, can't clean up during split abort. It will be cleaned up when it is reactivated.");
3444 : continue;
3445 : }
3446 :
3447 : // Detach the remote child. If the pageserver split API call is still in progress, this call will get
3448 : // a 503 and retry, up to our limit.
3449 : tracing::info!("Detaching {child_id} on {node}...");
3450 : match node
3451 : .with_client_retries(
3452 0 : |client| async move {
3453 0 : let config = LocationConfig {
3454 0 : mode: LocationConfigMode::Detached,
3455 0 : generation: None,
3456 0 : secondary_conf: None,
3457 0 : shard_number: child_id.shard_number.0,
3458 0 : shard_count: child_id.shard_count.literal(),
3459 0 : // Stripe size and tenant config don't matter when detaching
3460 0 : shard_stripe_size: 0,
3461 0 : tenant_conf: TenantConfig::default(),
3462 0 : };
3463 0 :
3464 0 : client.location_config(child_id, config, None, false).await
3465 0 : },
3466 : &self.config.jwt_token,
3467 : 1,
3468 : 10,
3469 : Duration::from_secs(5),
3470 : &self.cancel,
3471 : )
3472 : .await
3473 : {
3474 : Some(Ok(_)) => {}
3475 : Some(Err(e)) => {
3476 : // We failed to communicate with the remote node. This is problematic: we may be
3477 : // leaving it with a rogue child shard.
3478 : tracing::warn!(
3479 : "Failed to detach child {child_id} from node {node} during abort"
3480 : );
3481 : return Err(e.into());
3482 : }
3483 : None => {
3484 : // Cancellation: we were shutdown or the node went offline. Shutdown is fine, we'll
3485 : // clean up on restart. The node going offline requires a retry.
3486 : return Err(TenantShardSplitAbortError::Unavailable);
3487 : }
3488 : };
3489 : }
3490 :
3491 : tracing::info!("Successfully aborted split");
3492 : Ok(())
3493 : }
3494 :
3495 : /// Infallible final stage of [`Self::tenant_shard_split`]: update the contents
3496 : /// of the tenant map to reflect the child shards that exist after the split.
3497 0 : fn tenant_shard_split_commit_inmem(
3498 0 : &self,
3499 0 : tenant_id: TenantId,
3500 0 : new_shard_count: ShardCount,
3501 0 : new_stripe_size: Option<ShardStripeSize>,
3502 0 : ) -> (
3503 0 : TenantShardSplitResponse,
3504 0 : Vec<(TenantShardId, NodeId, ShardStripeSize)>,
3505 0 : Vec<ReconcilerWaiter>,
3506 0 : ) {
3507 0 : let mut response = TenantShardSplitResponse {
3508 0 : new_shards: Vec::new(),
3509 0 : };
3510 0 : let mut child_locations = Vec::new();
3511 0 : let mut waiters = Vec::new();
3512 0 :
3513 0 : {
3514 0 : let mut locked = self.inner.write().unwrap();
3515 0 :
3516 0 : let parent_ids = locked
3517 0 : .tenants
3518 0 : .range(TenantShardId::tenant_range(tenant_id))
3519 0 : .map(|(shard_id, _)| *shard_id)
3520 0 : .collect::<Vec<_>>();
3521 0 :
3522 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
3523 0 : for parent_id in parent_ids {
3524 0 : let child_ids = parent_id.split(new_shard_count);
3525 :
3526 0 : let (pageserver, generation, policy, parent_ident, config) = {
3527 0 : let mut old_state = tenants
3528 0 : .remove(&parent_id)
3529 0 : .expect("It was present, we just split it");
3530 0 :
3531 0 : // A non-splitting state is impossible, because [`Self::tenant_shard_split`] holds
3532 0 : // a TenantId lock and passes it through to [`TenantShardSplitAbort`] in case of cleanup:
3533 0 : // nothing else can clear this.
3534 0 : assert!(matches!(old_state.splitting, SplitState::Splitting));
3535 :
3536 0 : let old_attached = old_state.intent.get_attached().unwrap();
3537 0 : old_state.intent.clear(scheduler);
3538 0 : let generation = old_state.generation.expect("Shard must have been attached");
3539 0 : (
3540 0 : old_attached,
3541 0 : generation,
3542 0 : old_state.policy,
3543 0 : old_state.shard,
3544 0 : old_state.config,
3545 0 : )
3546 0 : };
3547 0 :
3548 0 : let mut schedule_context = ScheduleContext::default();
3549 0 : for child in child_ids {
3550 0 : let mut child_shard = parent_ident;
3551 0 : child_shard.number = child.shard_number;
3552 0 : child_shard.count = child.shard_count;
3553 0 : if let Some(stripe_size) = new_stripe_size {
3554 0 : child_shard.stripe_size = stripe_size;
3555 0 : }
3556 :
3557 0 : let mut child_observed: HashMap<NodeId, ObservedStateLocation> = HashMap::new();
3558 0 : child_observed.insert(
3559 0 : pageserver,
3560 0 : ObservedStateLocation {
3561 0 : conf: Some(attached_location_conf(
3562 0 : generation,
3563 0 : &child_shard,
3564 0 : &config,
3565 0 : &policy,
3566 0 : )),
3567 0 : },
3568 0 : );
3569 0 :
3570 0 : let mut child_state = TenantShard::new(child, child_shard, policy.clone());
3571 0 : child_state.intent = IntentState::single(scheduler, Some(pageserver));
3572 0 : child_state.observed = ObservedState {
3573 0 : locations: child_observed,
3574 0 : };
3575 0 : child_state.generation = Some(generation);
3576 0 : child_state.config = config.clone();
3577 0 :
3578 0 : // The child's TenantShard::splitting is intentionally left at the default value of Idle,
3579 0 : // as at this point in the split process we have succeeded and this part is infallible:
3580 0 : // we will never need to do any special recovery from this state.
3581 0 :
3582 0 : child_locations.push((child, pageserver, child_shard.stripe_size));
3583 :
3584 0 : if let Err(e) = child_state.schedule(scheduler, &mut schedule_context) {
3585 : // This is not fatal, because we've implicitly already got an attached
3586 : // location for the child shard. Failure here just means we couldn't
3587 : // find a secondary (e.g. because cluster is overloaded).
3588 0 : tracing::warn!("Failed to schedule child shard {child}: {e}");
3589 0 : }
3590 : // In the background, attach secondary locations for the new shards
3591 0 : if let Some(waiter) = self.maybe_reconcile_shard(&mut child_state, nodes) {
3592 0 : waiters.push(waiter);
3593 0 : }
3594 :
3595 0 : tenants.insert(child, child_state);
3596 0 : response.new_shards.push(child);
3597 : }
3598 : }
3599 0 : (response, child_locations, waiters)
3600 0 : }
3601 0 : }
3602 :
3603 0 : async fn tenant_shard_split_start_secondaries(
3604 0 : &self,
3605 0 : tenant_id: TenantId,
3606 0 : waiters: Vec<ReconcilerWaiter>,
3607 0 : ) {
3608 : // Wait for initial reconcile of child shards, this creates the secondary locations
3609 0 : if let Err(e) = self.await_waiters(waiters, RECONCILE_TIMEOUT).await {
3610 : // This is not a failure to split: it's some issue reconciling the new child shards, perhaps
3611 : // their secondaries couldn't be attached.
3612 0 : tracing::warn!("Failed to reconcile after split: {e}");
3613 0 : return;
3614 0 : }
3615 :
3616 : // Take the state lock to discover the attached & secondary intents for all shards
3617 0 : let (attached, secondary) = {
3618 0 : let locked = self.inner.read().unwrap();
3619 0 : let mut attached = Vec::new();
3620 0 : let mut secondary = Vec::new();
3621 :
3622 0 : for (tenant_shard_id, shard) in
3623 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
3624 : {
3625 0 : let Some(node_id) = shard.intent.get_attached() else {
3626 : // Unexpected. Race with a PlacementPolicy change?
3627 0 : tracing::warn!(
3628 0 : "No attached node on {tenant_shard_id} immediately after shard split!"
3629 : );
3630 0 : continue;
3631 : };
3632 :
3633 0 : let Some(secondary_node_id) = shard.intent.get_secondary().first() else {
3634 : // No secondary location. Nothing for us to do.
3635 0 : continue;
3636 : };
3637 :
3638 0 : let attached_node = locked
3639 0 : .nodes
3640 0 : .get(node_id)
3641 0 : .expect("Pageservers may not be deleted while referenced");
3642 0 :
3643 0 : let secondary_node = locked
3644 0 : .nodes
3645 0 : .get(secondary_node_id)
3646 0 : .expect("Pageservers may not be deleted while referenced");
3647 0 :
3648 0 : attached.push((*tenant_shard_id, attached_node.clone()));
3649 0 : secondary.push((*tenant_shard_id, secondary_node.clone()));
3650 : }
3651 0 : (attached, secondary)
3652 0 : };
3653 0 :
3654 0 : if secondary.is_empty() {
3655 : // No secondary locations; nothing for us to do
3656 0 : return;
3657 0 : }
3658 :
3659 0 : for result in self
3660 0 : .tenant_for_shards_api(
3661 0 : attached,
3662 0 : |tenant_shard_id, client| async move {
3663 0 : client.tenant_heatmap_upload(tenant_shard_id).await
3664 0 : },
3665 0 : 1,
3666 0 : 1,
3667 0 : SHORT_RECONCILE_TIMEOUT,
3668 0 : &self.cancel,
3669 0 : )
3670 0 : .await
3671 : {
3672 0 : if let Err(e) = result {
3673 0 : tracing::warn!("Error calling heatmap upload after shard split: {e}");
3674 0 : return;
3675 0 : }
3676 : }
3677 :
3678 0 : for result in self
3679 0 : .tenant_for_shards_api(
3680 0 : secondary,
3681 0 : |tenant_shard_id, client| async move {
3682 0 : client
3683 0 : .tenant_secondary_download(tenant_shard_id, Some(Duration::ZERO))
3684 0 : .await
3685 0 : },
3686 0 : 1,
3687 0 : 1,
3688 0 : SHORT_RECONCILE_TIMEOUT,
3689 0 : &self.cancel,
3690 0 : )
3691 0 : .await
3692 : {
3693 0 : if let Err(e) = result {
3694 0 : tracing::warn!("Error calling secondary download after shard split: {e}");
3695 0 : return;
3696 0 : }
3697 : }
3698 0 : }
3699 :
3700 0 : pub(crate) async fn tenant_shard_split(
3701 0 : &self,
3702 0 : tenant_id: TenantId,
3703 0 : split_req: TenantShardSplitRequest,
3704 0 : ) -> Result<TenantShardSplitResponse, ApiError> {
3705 : // TODO: return 503 if we get stuck waiting for this lock
3706 : // (issue https://github.com/neondatabase/neon/issues/7108)
3707 0 : let _tenant_lock = trace_exclusive_lock(
3708 0 : &self.tenant_op_locks,
3709 0 : tenant_id,
3710 0 : TenantOperations::ShardSplit,
3711 0 : )
3712 0 : .await;
3713 :
3714 0 : let new_shard_count = ShardCount::new(split_req.new_shard_count);
3715 0 : let new_stripe_size = split_req.new_stripe_size;
3716 :
3717 : // Validate the request and construct parameters. This phase is fallible, but does not require
3718 : // rollback on errors, as it does no I/O and mutates no state.
3719 0 : let shard_split_params = match self.prepare_tenant_shard_split(tenant_id, split_req)? {
3720 0 : ShardSplitAction::NoOp(resp) => return Ok(resp),
3721 0 : ShardSplitAction::Split(params) => params,
3722 : };
3723 :
3724 : // Execute this split: this phase mutates state and does remote I/O on pageservers. If it fails,
3725 : // we must roll back.
3726 0 : let r = self
3727 0 : .do_tenant_shard_split(tenant_id, shard_split_params)
3728 0 : .await;
3729 :
3730 0 : let (response, waiters) = match r {
3731 0 : Ok(r) => r,
3732 0 : Err(e) => {
3733 0 : // Split might be part-done, we must do work to abort it.
3734 0 : tracing::warn!("Enqueuing background abort of split on {tenant_id}");
3735 0 : self.abort_tx
3736 0 : .send(TenantShardSplitAbort {
3737 0 : tenant_id,
3738 0 : new_shard_count,
3739 0 : new_stripe_size,
3740 0 : _tenant_lock,
3741 0 : })
3742 0 : // Ignore error sending: that just means we're shutting down: aborts are ephemeral so it's fine to drop it.
3743 0 : .ok();
3744 0 : return Err(e);
3745 : }
3746 : };
3747 :
3748 : // The split is now complete. As an optimization, we will trigger all the child shards to upload
3749 : // a heatmap immediately, and all their secondary locations to start downloading: this avoids waiting
3750 : // for the background heatmap/download interval before secondaries get warm enough to migrate shards
3751 : // in [`Self::optimize_all`]
3752 0 : self.tenant_shard_split_start_secondaries(tenant_id, waiters)
3753 0 : .await;
3754 0 : Ok(response)
3755 0 : }
3756 :
3757 0 : fn prepare_tenant_shard_split(
3758 0 : &self,
3759 0 : tenant_id: TenantId,
3760 0 : split_req: TenantShardSplitRequest,
3761 0 : ) -> Result<ShardSplitAction, ApiError> {
3762 0 : fail::fail_point!("shard-split-validation", |_| Err(ApiError::BadRequest(
3763 0 : anyhow::anyhow!("failpoint")
3764 0 : )));
3765 :
3766 0 : let mut policy = None;
3767 0 : let mut config = None;
3768 0 : let mut shard_ident = None;
3769 : // Validate input, and calculate which shards we will create
3770 0 : let (old_shard_count, targets) =
3771 : {
3772 0 : let locked = self.inner.read().unwrap();
3773 0 :
3774 0 : let pageservers = locked.nodes.clone();
3775 0 :
3776 0 : let mut targets = Vec::new();
3777 0 :
3778 0 : // In case this is a retry, count how many already-split shards we found
3779 0 : let mut children_found = Vec::new();
3780 0 : let mut old_shard_count = None;
3781 :
3782 0 : for (tenant_shard_id, shard) in
3783 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
3784 : {
3785 0 : match shard.shard.count.count().cmp(&split_req.new_shard_count) {
3786 : Ordering::Equal => {
3787 : // Already split this
3788 0 : children_found.push(*tenant_shard_id);
3789 0 : continue;
3790 : }
3791 : Ordering::Greater => {
3792 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
3793 0 : "Requested count {} but already have shards at count {}",
3794 0 : split_req.new_shard_count,
3795 0 : shard.shard.count.count()
3796 0 : )));
3797 : }
3798 0 : Ordering::Less => {
3799 0 : // Fall through: this shard has lower count than requested,
3800 0 : // is a candidate for splitting.
3801 0 : }
3802 0 : }
3803 0 :
3804 0 : match old_shard_count {
3805 0 : None => old_shard_count = Some(shard.shard.count),
3806 0 : Some(old_shard_count) => {
3807 0 : if old_shard_count != shard.shard.count {
3808 : // We may hit this case if a caller asked for two splits to
3809 : // different sizes, before the first one is complete.
3810 : // e.g. 1->2, 2->4, where the 4 call comes while we have a mixture
3811 : // of shard_count=1 and shard_count=2 shards in the map.
3812 0 : return Err(ApiError::Conflict(
3813 0 : "Cannot split, currently mid-split".to_string(),
3814 0 : ));
3815 0 : }
3816 : }
3817 : }
3818 0 : if policy.is_none() {
3819 0 : policy = Some(shard.policy.clone());
3820 0 : }
3821 0 : if shard_ident.is_none() {
3822 0 : shard_ident = Some(shard.shard);
3823 0 : }
3824 0 : if config.is_none() {
3825 0 : config = Some(shard.config.clone());
3826 0 : }
3827 :
3828 0 : if tenant_shard_id.shard_count.count() == split_req.new_shard_count {
3829 0 : tracing::info!(
3830 0 : "Tenant shard {} already has shard count {}",
3831 : tenant_shard_id,
3832 : split_req.new_shard_count
3833 : );
3834 0 : continue;
3835 0 : }
3836 :
3837 0 : let node_id = shard.intent.get_attached().ok_or(ApiError::BadRequest(
3838 0 : anyhow::anyhow!("Cannot split a tenant that is not attached"),
3839 0 : ))?;
3840 :
3841 0 : let node = pageservers
3842 0 : .get(&node_id)
3843 0 : .expect("Pageservers may not be deleted while referenced");
3844 0 :
3845 0 : targets.push(ShardSplitTarget {
3846 0 : parent_id: *tenant_shard_id,
3847 0 : node: node.clone(),
3848 0 : child_ids: tenant_shard_id
3849 0 : .split(ShardCount::new(split_req.new_shard_count)),
3850 0 : });
3851 : }
3852 :
3853 0 : if targets.is_empty() {
3854 0 : if children_found.len() == split_req.new_shard_count as usize {
3855 0 : return Ok(ShardSplitAction::NoOp(TenantShardSplitResponse {
3856 0 : new_shards: children_found,
3857 0 : }));
3858 : } else {
3859 : // No shards found to split, and no existing children found: the
3860 : // tenant doesn't exist at all.
3861 0 : return Err(ApiError::NotFound(
3862 0 : anyhow::anyhow!("Tenant {} not found", tenant_id).into(),
3863 0 : ));
3864 : }
3865 0 : }
3866 0 :
3867 0 : (old_shard_count, targets)
3868 0 : };
3869 0 :
3870 0 : // unwrap safety: we would have returned above if we didn't find at least one shard to split
3871 0 : let old_shard_count = old_shard_count.unwrap();
3872 0 : let shard_ident = if let Some(new_stripe_size) = split_req.new_stripe_size {
3873 : // This ShardIdentity will be used as the template for all children, so this implicitly
3874 : // applies the new stripe size to the children.
3875 0 : let mut shard_ident = shard_ident.unwrap();
3876 0 : if shard_ident.count.count() > 1 && shard_ident.stripe_size != new_stripe_size {
3877 0 : return Err(ApiError::BadRequest(anyhow::anyhow!("Attempted to change stripe size ({:?}->{new_stripe_size:?}) on a tenant with multiple shards", shard_ident.stripe_size)));
3878 0 : }
3879 0 :
3880 0 : shard_ident.stripe_size = new_stripe_size;
3881 0 : tracing::info!("applied stripe size {}", shard_ident.stripe_size.0);
3882 0 : shard_ident
3883 : } else {
3884 0 : shard_ident.unwrap()
3885 : };
3886 0 : let policy = policy.unwrap();
3887 0 : let config = config.unwrap();
3888 0 :
3889 0 : Ok(ShardSplitAction::Split(ShardSplitParams {
3890 0 : old_shard_count,
3891 0 : new_shard_count: ShardCount::new(split_req.new_shard_count),
3892 0 : new_stripe_size: split_req.new_stripe_size,
3893 0 : targets,
3894 0 : policy,
3895 0 : config,
3896 0 : shard_ident,
3897 0 : }))
3898 0 : }
3899 :
3900 0 : async fn do_tenant_shard_split(
3901 0 : &self,
3902 0 : tenant_id: TenantId,
3903 0 : params: ShardSplitParams,
3904 0 : ) -> Result<(TenantShardSplitResponse, Vec<ReconcilerWaiter>), ApiError> {
3905 0 : // FIXME: we have dropped self.inner lock, and not yet written anything to the database: another
3906 0 : // request could occur here, deleting or mutating the tenant. begin_shard_split checks that the
3907 0 : // parent shards exist as expected, but it would be neater to do the above pre-checks within the
3908 0 : // same database transaction rather than pre-check in-memory and then maybe-fail the database write.
3909 0 : // (https://github.com/neondatabase/neon/issues/6676)
3910 0 :
3911 0 : let ShardSplitParams {
3912 0 : old_shard_count,
3913 0 : new_shard_count,
3914 0 : new_stripe_size,
3915 0 : mut targets,
3916 0 : policy,
3917 0 : config,
3918 0 : shard_ident,
3919 0 : } = params;
3920 :
3921 : // Drop any secondary locations: pageservers do not support splitting these, and in any case the
3922 : // end-state for a split tenant will usually be to have secondary locations on different nodes.
3923 : // The reconciliation calls in this block also implicitly cancel+barrier wrt any ongoing reconciliation
3924 : // at the time of split.
3925 0 : let waiters = {
3926 0 : let mut locked = self.inner.write().unwrap();
3927 0 : let mut waiters = Vec::new();
3928 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
3929 0 : for target in &mut targets {
3930 0 : let Some(shard) = tenants.get_mut(&target.parent_id) else {
3931 : // Paranoia check: this shouldn't happen: we have the oplock for this tenant ID.
3932 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
3933 0 : "Shard {} not found",
3934 0 : target.parent_id
3935 0 : )));
3936 : };
3937 :
3938 0 : if shard.intent.get_attached() != &Some(target.node.get_id()) {
3939 : // Paranoia check: this shouldn't happen: we have the oplock for this tenant ID.
3940 0 : return Err(ApiError::Conflict(format!(
3941 0 : "Shard {} unexpectedly rescheduled during split",
3942 0 : target.parent_id
3943 0 : )));
3944 0 : }
3945 0 :
3946 0 : // Irrespective of PlacementPolicy, clear secondary locations from intent
3947 0 : shard.intent.clear_secondary(scheduler);
3948 :
3949 : // Run Reconciler to execute detach fo secondary locations.
3950 0 : if let Some(waiter) = self.maybe_reconcile_shard(shard, nodes) {
3951 0 : waiters.push(waiter);
3952 0 : }
3953 : }
3954 0 : waiters
3955 0 : };
3956 0 : self.await_waiters(waiters, RECONCILE_TIMEOUT).await?;
3957 :
3958 : // Before creating any new child shards in memory or on the pageservers, persist them: this
3959 : // enables us to ensure that we will always be able to clean up if something goes wrong. This also
3960 : // acts as the protection against two concurrent attempts to split: one of them will get a database
3961 : // error trying to insert the child shards.
3962 0 : let mut child_tsps = Vec::new();
3963 0 : for target in &targets {
3964 0 : let mut this_child_tsps = Vec::new();
3965 0 : for child in &target.child_ids {
3966 0 : let mut child_shard = shard_ident;
3967 0 : child_shard.number = child.shard_number;
3968 0 : child_shard.count = child.shard_count;
3969 0 :
3970 0 : tracing::info!(
3971 0 : "Create child shard persistence with stripe size {}",
3972 : shard_ident.stripe_size.0
3973 : );
3974 :
3975 0 : this_child_tsps.push(TenantShardPersistence {
3976 0 : tenant_id: child.tenant_id.to_string(),
3977 0 : shard_number: child.shard_number.0 as i32,
3978 0 : shard_count: child.shard_count.literal() as i32,
3979 0 : shard_stripe_size: shard_ident.stripe_size.0 as i32,
3980 0 : // Note: this generation is a placeholder, [`Persistence::begin_shard_split`] will
3981 0 : // populate the correct generation as part of its transaction, to protect us
3982 0 : // against racing with changes in the state of the parent.
3983 0 : generation: None,
3984 0 : generation_pageserver: Some(target.node.get_id().0 as i64),
3985 0 : placement_policy: serde_json::to_string(&policy).unwrap(),
3986 0 : config: serde_json::to_string(&config).unwrap(),
3987 0 : splitting: SplitState::Splitting,
3988 0 :
3989 0 : // Scheduling policies do not carry through to children
3990 0 : scheduling_policy: serde_json::to_string(&ShardSchedulingPolicy::default())
3991 0 : .unwrap(),
3992 0 : });
3993 : }
3994 :
3995 0 : child_tsps.push((target.parent_id, this_child_tsps));
3996 : }
3997 :
3998 0 : if let Err(e) = self
3999 0 : .persistence
4000 0 : .begin_shard_split(old_shard_count, tenant_id, child_tsps)
4001 0 : .await
4002 : {
4003 0 : match e {
4004 : DatabaseError::Query(diesel::result::Error::DatabaseError(
4005 : DatabaseErrorKind::UniqueViolation,
4006 : _,
4007 : )) => {
4008 : // Inserting a child shard violated a unique constraint: we raced with another call to
4009 : // this function
4010 0 : tracing::warn!("Conflicting attempt to split {tenant_id}: {e}");
4011 0 : return Err(ApiError::Conflict("Tenant is already splitting".into()));
4012 : }
4013 0 : _ => return Err(ApiError::InternalServerError(e.into())),
4014 : }
4015 0 : }
4016 0 : fail::fail_point!("shard-split-post-begin", |_| Err(
4017 0 : ApiError::InternalServerError(anyhow::anyhow!("failpoint"))
4018 0 : ));
4019 :
4020 : // Now that I have persisted the splitting state, apply it in-memory. This is infallible, so
4021 : // callers may assume that if splitting is set in memory, then it was persisted, and if splitting
4022 : // is not set in memory, then it was not persisted.
4023 : {
4024 0 : let mut locked = self.inner.write().unwrap();
4025 0 : for target in &targets {
4026 0 : if let Some(parent_shard) = locked.tenants.get_mut(&target.parent_id) {
4027 0 : parent_shard.splitting = SplitState::Splitting;
4028 0 : // Put the observed state to None, to reflect that it is indeterminate once we start the
4029 0 : // split operation.
4030 0 : parent_shard
4031 0 : .observed
4032 0 : .locations
4033 0 : .insert(target.node.get_id(), ObservedStateLocation { conf: None });
4034 0 : }
4035 : }
4036 : }
4037 :
4038 : // TODO: issue split calls concurrently (this only matters once we're splitting
4039 : // N>1 shards into M shards -- initially we're usually splitting 1 shard into N).
4040 :
4041 0 : for target in &targets {
4042 : let ShardSplitTarget {
4043 0 : parent_id,
4044 0 : node,
4045 0 : child_ids,
4046 0 : } = target;
4047 0 : let client = PageserverClient::new(
4048 0 : node.get_id(),
4049 0 : node.base_url(),
4050 0 : self.config.jwt_token.as_deref(),
4051 0 : );
4052 0 : let response = client
4053 0 : .tenant_shard_split(
4054 0 : *parent_id,
4055 0 : TenantShardSplitRequest {
4056 0 : new_shard_count: new_shard_count.literal(),
4057 0 : new_stripe_size,
4058 0 : },
4059 0 : )
4060 0 : .await
4061 0 : .map_err(|e| ApiError::Conflict(format!("Failed to split {}: {}", parent_id, e)))?;
4062 :
4063 0 : fail::fail_point!("shard-split-post-remote", |_| Err(ApiError::Conflict(
4064 0 : "failpoint".to_string()
4065 0 : )));
4066 :
4067 0 : failpoint_support::sleep_millis_async!("shard-split-post-remote-sleep", &self.cancel);
4068 :
4069 0 : tracing::info!(
4070 0 : "Split {} into {}",
4071 0 : parent_id,
4072 0 : response
4073 0 : .new_shards
4074 0 : .iter()
4075 0 : .map(|s| format!("{:?}", s))
4076 0 : .collect::<Vec<_>>()
4077 0 : .join(",")
4078 : );
4079 :
4080 0 : if &response.new_shards != child_ids {
4081 : // This should never happen: the pageserver should agree with us on how shard splits work.
4082 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
4083 0 : "Splitting shard {} resulted in unexpected IDs: {:?} (expected {:?})",
4084 0 : parent_id,
4085 0 : response.new_shards,
4086 0 : child_ids
4087 0 : )));
4088 0 : }
4089 : }
4090 :
4091 : // TODO: if the pageserver restarted concurrently with our split API call,
4092 : // the actual generation of the child shard might differ from the generation
4093 : // we expect it to have. In order for our in-database generation to end up
4094 : // correct, we should carry the child generation back in the response and apply it here
4095 : // in complete_shard_split (and apply the correct generation in memory)
4096 : // (or, we can carry generation in the request and reject the request if
4097 : // it doesn't match, but that requires more retry logic on this side)
4098 :
4099 0 : self.persistence
4100 0 : .complete_shard_split(tenant_id, old_shard_count)
4101 0 : .await?;
4102 :
4103 0 : fail::fail_point!("shard-split-post-complete", |_| Err(
4104 0 : ApiError::InternalServerError(anyhow::anyhow!("failpoint"))
4105 0 : ));
4106 :
4107 : // Replace all the shards we just split with their children: this phase is infallible.
4108 0 : let (response, child_locations, waiters) =
4109 0 : self.tenant_shard_split_commit_inmem(tenant_id, new_shard_count, new_stripe_size);
4110 0 :
4111 0 : // Send compute notifications for all the new shards
4112 0 : let mut failed_notifications = Vec::new();
4113 0 : for (child_id, child_ps, stripe_size) in child_locations {
4114 0 : if let Err(e) = self
4115 0 : .compute_hook
4116 0 : .notify(child_id, child_ps, stripe_size, &self.cancel)
4117 0 : .await
4118 : {
4119 0 : tracing::warn!("Failed to update compute of {}->{} during split, proceeding anyway to complete split ({e})",
4120 : child_id, child_ps);
4121 0 : failed_notifications.push(child_id);
4122 0 : }
4123 : }
4124 :
4125 : // If we failed any compute notifications, make a note to retry later.
4126 0 : if !failed_notifications.is_empty() {
4127 0 : let mut locked = self.inner.write().unwrap();
4128 0 : for failed in failed_notifications {
4129 0 : if let Some(shard) = locked.tenants.get_mut(&failed) {
4130 0 : shard.pending_compute_notification = true;
4131 0 : }
4132 : }
4133 0 : }
4134 :
4135 0 : Ok((response, waiters))
4136 0 : }
4137 :
4138 0 : pub(crate) async fn tenant_shard_migrate(
4139 0 : &self,
4140 0 : tenant_shard_id: TenantShardId,
4141 0 : migrate_req: TenantShardMigrateRequest,
4142 0 : ) -> Result<TenantShardMigrateResponse, ApiError> {
4143 0 : let waiter = {
4144 0 : let mut locked = self.inner.write().unwrap();
4145 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
4146 :
4147 0 : let Some(node) = nodes.get(&migrate_req.node_id) else {
4148 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
4149 0 : "Node {} not found",
4150 0 : migrate_req.node_id
4151 0 : )));
4152 : };
4153 :
4154 0 : if !node.is_available() {
4155 : // Warn but proceed: the caller may intend to manually adjust the placement of
4156 : // a shard even if the node is down, e.g. if intervening during an incident.
4157 0 : tracing::warn!("Migrating to unavailable node {node}");
4158 0 : }
4159 :
4160 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
4161 0 : return Err(ApiError::NotFound(
4162 0 : anyhow::anyhow!("Tenant shard not found").into(),
4163 0 : ));
4164 : };
4165 :
4166 0 : if shard.intent.get_attached() == &Some(migrate_req.node_id) {
4167 : // No-op case: we will still proceed to wait for reconciliation in case it is
4168 : // incomplete from an earlier update to the intent.
4169 0 : tracing::info!("Migrating: intent is unchanged {:?}", shard.intent);
4170 : } else {
4171 0 : let old_attached = *shard.intent.get_attached();
4172 0 :
4173 0 : match shard.policy {
4174 0 : PlacementPolicy::Attached(n) => {
4175 0 : // If our new attached node was a secondary, it no longer should be.
4176 0 : shard.intent.remove_secondary(scheduler, migrate_req.node_id);
4177 :
4178 : // If we were already attached to something, demote that to a secondary
4179 0 : if let Some(old_attached) = old_attached {
4180 0 : if n > 0 {
4181 : // Remove other secondaries to make room for the location we'll demote
4182 0 : while shard.intent.get_secondary().len() >= n {
4183 0 : shard.intent.pop_secondary(scheduler);
4184 0 : }
4185 :
4186 0 : shard.intent.push_secondary(scheduler, old_attached);
4187 0 : }
4188 0 : }
4189 :
4190 0 : shard.intent.set_attached(scheduler, Some(migrate_req.node_id));
4191 : }
4192 0 : PlacementPolicy::Secondary => {
4193 0 : shard.intent.clear(scheduler);
4194 0 : shard.intent.push_secondary(scheduler, migrate_req.node_id);
4195 0 : }
4196 : PlacementPolicy::Detached => {
4197 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
4198 0 : "Cannot migrate a tenant that is PlacementPolicy::Detached: configure it to an attached policy first"
4199 0 : )))
4200 : }
4201 : }
4202 :
4203 0 : tracing::info!("Migrating: new intent {:?}", shard.intent);
4204 0 : shard.sequence = shard.sequence.next();
4205 : }
4206 :
4207 0 : self.maybe_reconcile_shard(shard, nodes)
4208 : };
4209 :
4210 0 : if let Some(waiter) = waiter {
4211 0 : waiter.wait_timeout(RECONCILE_TIMEOUT).await?;
4212 : } else {
4213 0 : tracing::info!("Migration is a no-op");
4214 : }
4215 :
4216 0 : Ok(TenantShardMigrateResponse {})
4217 0 : }
4218 :
4219 : /// This is for debug/support only: we simply drop all state for a tenant, without
4220 : /// detaching or deleting it on pageservers.
4221 0 : pub(crate) async fn tenant_drop(&self, tenant_id: TenantId) -> Result<(), ApiError> {
4222 0 : self.persistence.delete_tenant(tenant_id).await?;
4223 :
4224 0 : let mut locked = self.inner.write().unwrap();
4225 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
4226 0 : let mut shards = Vec::new();
4227 0 : for (tenant_shard_id, _) in tenants.range(TenantShardId::tenant_range(tenant_id)) {
4228 0 : shards.push(*tenant_shard_id);
4229 0 : }
4230 :
4231 0 : for shard_id in shards {
4232 0 : if let Some(mut shard) = tenants.remove(&shard_id) {
4233 0 : shard.intent.clear(scheduler);
4234 0 : }
4235 : }
4236 :
4237 0 : Ok(())
4238 0 : }
4239 :
4240 : /// This is for debug/support only: assuming tenant data is already present in S3, we "create" a
4241 : /// tenant with a very high generation number so that it will see the existing data.
4242 0 : pub(crate) async fn tenant_import(
4243 0 : &self,
4244 0 : tenant_id: TenantId,
4245 0 : ) -> Result<TenantCreateResponse, ApiError> {
4246 0 : // Pick an arbitrary available pageserver to use for scanning the tenant in remote storage
4247 0 : let maybe_node = {
4248 0 : self.inner
4249 0 : .read()
4250 0 : .unwrap()
4251 0 : .nodes
4252 0 : .values()
4253 0 : .find(|n| n.is_available())
4254 0 : .cloned()
4255 : };
4256 0 : let Some(node) = maybe_node else {
4257 0 : return Err(ApiError::BadRequest(anyhow::anyhow!("No nodes available")));
4258 : };
4259 :
4260 0 : let client = PageserverClient::new(
4261 0 : node.get_id(),
4262 0 : node.base_url(),
4263 0 : self.config.jwt_token.as_deref(),
4264 0 : );
4265 :
4266 0 : let scan_result = client
4267 0 : .tenant_scan_remote_storage(tenant_id)
4268 0 : .await
4269 0 : .map_err(|e| passthrough_api_error(&node, e))?;
4270 :
4271 : // A post-split tenant may contain a mixture of shard counts in remote storage: pick the highest count.
4272 0 : let Some(shard_count) = scan_result
4273 0 : .shards
4274 0 : .iter()
4275 0 : .map(|s| s.tenant_shard_id.shard_count)
4276 0 : .max()
4277 : else {
4278 0 : return Err(ApiError::NotFound(
4279 0 : anyhow::anyhow!("No shards found").into(),
4280 0 : ));
4281 : };
4282 :
4283 : // Ideally we would set each newly imported shard's generation independently, but for correctness it is sufficient
4284 : // to
4285 0 : let generation = scan_result
4286 0 : .shards
4287 0 : .iter()
4288 0 : .map(|s| s.generation)
4289 0 : .max()
4290 0 : .expect("We already validated >0 shards");
4291 0 :
4292 0 : // FIXME: we have no way to recover the shard stripe size from contents of remote storage: this will
4293 0 : // only work if they were using the default stripe size.
4294 0 : let stripe_size = ShardParameters::DEFAULT_STRIPE_SIZE;
4295 :
4296 0 : let (response, waiters) = self
4297 0 : .do_tenant_create(TenantCreateRequest {
4298 0 : new_tenant_id: TenantShardId::unsharded(tenant_id),
4299 0 : generation,
4300 0 :
4301 0 : shard_parameters: ShardParameters {
4302 0 : count: shard_count,
4303 0 : stripe_size,
4304 0 : },
4305 0 : placement_policy: Some(PlacementPolicy::Attached(0)), // No secondaries, for convenient debug/hacking
4306 0 :
4307 0 : // There is no way to know what the tenant's config was: revert to defaults
4308 0 : //
4309 0 : // TODO: remove `switch_aux_file_policy` once we finish auxv2 migration
4310 0 : //
4311 0 : // we write to both v1+v2 storage, so that the test case can use either storage format for testing
4312 0 : config: TenantConfig {
4313 0 : switch_aux_file_policy: Some(models::AuxFilePolicy::CrossValidation),
4314 0 : ..TenantConfig::default()
4315 0 : },
4316 0 : })
4317 0 : .await?;
4318 :
4319 0 : if let Err(e) = self.await_waiters(waiters, SHORT_RECONCILE_TIMEOUT).await {
4320 : // Since this is a debug/support operation, all kinds of weird issues are possible (e.g. this
4321 : // tenant doesn't exist in the control plane), so don't fail the request if it can't fully
4322 : // reconcile, as reconciliation includes notifying compute.
4323 0 : tracing::warn!(%tenant_id, "Reconcile not done yet while importing tenant ({e})");
4324 0 : }
4325 :
4326 0 : Ok(response)
4327 0 : }
4328 :
4329 : /// For debug/support: a full JSON dump of TenantShards. Returns a response so that
4330 : /// we don't have to make TenantShard clonable in the return path.
4331 0 : pub(crate) fn tenants_dump(&self) -> Result<hyper::Response<hyper::Body>, ApiError> {
4332 0 : let serialized = {
4333 0 : let locked = self.inner.read().unwrap();
4334 0 : let result = locked.tenants.values().collect::<Vec<_>>();
4335 0 : serde_json::to_string(&result).map_err(|e| ApiError::InternalServerError(e.into()))?
4336 : };
4337 :
4338 0 : hyper::Response::builder()
4339 0 : .status(hyper::StatusCode::OK)
4340 0 : .header(hyper::header::CONTENT_TYPE, "application/json")
4341 0 : .body(hyper::Body::from(serialized))
4342 0 : .map_err(|e| ApiError::InternalServerError(e.into()))
4343 0 : }
4344 :
4345 : /// Check the consistency of in-memory state vs. persistent state, and check that the
4346 : /// scheduler's statistics are up to date.
4347 : ///
4348 : /// These consistency checks expect an **idle** system. If changes are going on while
4349 : /// we run, then we can falsely indicate a consistency issue. This is sufficient for end-of-test
4350 : /// checks, but not suitable for running continuously in the background in the field.
4351 0 : pub(crate) async fn consistency_check(&self) -> Result<(), ApiError> {
4352 0 : let (mut expect_nodes, mut expect_shards) = {
4353 0 : let locked = self.inner.read().unwrap();
4354 0 :
4355 0 : locked
4356 0 : .scheduler
4357 0 : .consistency_check(locked.nodes.values(), locked.tenants.values())
4358 0 : .context("Scheduler checks")
4359 0 : .map_err(ApiError::InternalServerError)?;
4360 :
4361 0 : let expect_nodes = locked
4362 0 : .nodes
4363 0 : .values()
4364 0 : .map(|n| n.to_persistent())
4365 0 : .collect::<Vec<_>>();
4366 0 :
4367 0 : let expect_shards = locked
4368 0 : .tenants
4369 0 : .values()
4370 0 : .map(|t| t.to_persistent())
4371 0 : .collect::<Vec<_>>();
4372 :
4373 : // This method can only validate the state of an idle system: if a reconcile is in
4374 : // progress, fail out early to avoid giving false errors on state that won't match
4375 : // between database and memory under a ReconcileResult is processed.
4376 0 : for t in locked.tenants.values() {
4377 0 : if t.reconciler.is_some() {
4378 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
4379 0 : "Shard {} reconciliation in progress",
4380 0 : t.tenant_shard_id
4381 0 : )));
4382 0 : }
4383 : }
4384 :
4385 0 : (expect_nodes, expect_shards)
4386 : };
4387 :
4388 0 : let mut nodes = self.persistence.list_nodes().await?;
4389 0 : expect_nodes.sort_by_key(|n| n.node_id);
4390 0 : nodes.sort_by_key(|n| n.node_id);
4391 0 :
4392 0 : if nodes != expect_nodes {
4393 0 : tracing::error!("Consistency check failed on nodes.");
4394 0 : tracing::error!(
4395 0 : "Nodes in memory: {}",
4396 0 : serde_json::to_string(&expect_nodes)
4397 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
4398 : );
4399 0 : tracing::error!(
4400 0 : "Nodes in database: {}",
4401 0 : serde_json::to_string(&nodes)
4402 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
4403 : );
4404 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
4405 0 : "Node consistency failure"
4406 0 : )));
4407 0 : }
4408 :
4409 0 : let mut shards = self.persistence.list_tenant_shards().await?;
4410 0 : shards.sort_by_key(|tsp| (tsp.tenant_id.clone(), tsp.shard_number, tsp.shard_count));
4411 0 : expect_shards.sort_by_key(|tsp| (tsp.tenant_id.clone(), tsp.shard_number, tsp.shard_count));
4412 0 :
4413 0 : if shards != expect_shards {
4414 0 : tracing::error!("Consistency check failed on shards.");
4415 0 : tracing::error!(
4416 0 : "Shards in memory: {}",
4417 0 : serde_json::to_string(&expect_shards)
4418 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
4419 : );
4420 0 : tracing::error!(
4421 0 : "Shards in database: {}",
4422 0 : serde_json::to_string(&shards)
4423 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
4424 : );
4425 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
4426 0 : "Shard consistency failure"
4427 0 : )));
4428 0 : }
4429 0 :
4430 0 : Ok(())
4431 0 : }
4432 :
4433 : /// For debug/support: a JSON dump of the [`Scheduler`]. Returns a response so that
4434 : /// we don't have to make TenantShard clonable in the return path.
4435 0 : pub(crate) fn scheduler_dump(&self) -> Result<hyper::Response<hyper::Body>, ApiError> {
4436 0 : let serialized = {
4437 0 : let locked = self.inner.read().unwrap();
4438 0 : serde_json::to_string(&locked.scheduler)
4439 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
4440 : };
4441 :
4442 0 : hyper::Response::builder()
4443 0 : .status(hyper::StatusCode::OK)
4444 0 : .header(hyper::header::CONTENT_TYPE, "application/json")
4445 0 : .body(hyper::Body::from(serialized))
4446 0 : .map_err(|e| ApiError::InternalServerError(e.into()))
4447 0 : }
4448 :
4449 : /// This is for debug/support only: we simply drop all state for a tenant, without
4450 : /// detaching or deleting it on pageservers. We do not try and re-schedule any
4451 : /// tenants that were on this node.
4452 0 : pub(crate) async fn node_drop(&self, node_id: NodeId) -> Result<(), ApiError> {
4453 0 : self.persistence.delete_node(node_id).await?;
4454 :
4455 0 : let mut locked = self.inner.write().unwrap();
4456 :
4457 0 : for shard in locked.tenants.values_mut() {
4458 0 : shard.deref_node(node_id);
4459 0 : shard.observed.locations.remove(&node_id);
4460 0 : }
4461 :
4462 0 : let mut nodes = (*locked.nodes).clone();
4463 0 : nodes.remove(&node_id);
4464 0 : locked.nodes = Arc::new(nodes);
4465 0 :
4466 0 : locked.scheduler.node_remove(node_id);
4467 0 :
4468 0 : Ok(())
4469 0 : }
4470 :
4471 : /// If a node has any work on it, it will be rescheduled: this is "clean" in the sense
4472 : /// that we don't leave any bad state behind in the storage controller, but unclean
4473 : /// in the sense that we are not carefully draining the node.
4474 0 : pub(crate) async fn node_delete(&self, node_id: NodeId) -> Result<(), ApiError> {
4475 0 : let _node_lock =
4476 0 : trace_exclusive_lock(&self.node_op_locks, node_id, NodeOperations::Delete).await;
4477 :
4478 : // 1. Atomically update in-memory state:
4479 : // - set the scheduling state to Pause to make subsequent scheduling ops skip it
4480 : // - update shards' intents to exclude the node, and reschedule any shards whose intents we modified.
4481 : // - drop the node from the main nodes map, so that when running reconciles complete they do not
4482 : // re-insert references to this node into the ObservedState of shards
4483 : // - drop the node from the scheduler
4484 : {
4485 0 : let mut locked = self.inner.write().unwrap();
4486 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
4487 0 :
4488 0 : {
4489 0 : let mut nodes_mut = (*nodes).deref().clone();
4490 0 : match nodes_mut.get_mut(&node_id) {
4491 0 : Some(node) => {
4492 0 : // We do not bother setting this in the database, because we're about to delete the row anyway, and
4493 0 : // if we crash it would not be desirable to leave the node paused after a restart.
4494 0 : node.set_scheduling(NodeSchedulingPolicy::Pause);
4495 0 : }
4496 : None => {
4497 0 : tracing::info!(
4498 0 : "Node not found: presuming this is a retry and returning success"
4499 : );
4500 0 : return Ok(());
4501 : }
4502 : }
4503 :
4504 0 : *nodes = Arc::new(nodes_mut);
4505 : }
4506 :
4507 0 : for (tenant_shard_id, shard) in tenants {
4508 0 : if shard.deref_node(node_id) {
4509 : // FIXME: we need to build a ScheduleContext that reflects this shard's peers, otherwise
4510 : // it won't properly do anti-affinity.
4511 0 : let mut schedule_context = ScheduleContext::default();
4512 :
4513 0 : if let Err(e) = shard.schedule(scheduler, &mut schedule_context) {
4514 : // TODO: implement force flag to remove a node even if we can't reschedule
4515 : // a tenant
4516 0 : tracing::error!("Refusing to delete node, shard {tenant_shard_id} can't be rescheduled: {e}");
4517 0 : return Err(e.into());
4518 : } else {
4519 0 : tracing::info!(
4520 0 : "Rescheduled shard {tenant_shard_id} away from node during deletion"
4521 : )
4522 : }
4523 :
4524 0 : self.maybe_reconcile_shard(shard, nodes);
4525 0 : }
4526 :
4527 : // Here we remove an existing observed location for the node we're removing, and it will
4528 : // not be re-added by a reconciler's completion because we filter out removed nodes in
4529 : // process_result.
4530 : //
4531 : // Note that we update the shard's observed state _after_ calling maybe_reconcile_shard: that
4532 : // means any reconciles we spawned will know about the node we're deleting, enabling them
4533 : // to do live migrations if it's still online.
4534 0 : shard.observed.locations.remove(&node_id);
4535 : }
4536 :
4537 0 : scheduler.node_remove(node_id);
4538 0 :
4539 0 : {
4540 0 : let mut nodes_mut = (**nodes).clone();
4541 0 : nodes_mut.remove(&node_id);
4542 0 : *nodes = Arc::new(nodes_mut);
4543 0 : }
4544 0 : }
4545 0 :
4546 0 : // Note: some `generation_pageserver` columns on tenant shards in the database may still refer to
4547 0 : // the removed node, as this column means "The pageserver to which this generation was issued", and
4548 0 : // their generations won't get updated until the reconcilers moving them away from this node complete.
4549 0 : // That is safe because in Service::spawn we only use generation_pageserver if it refers to a node
4550 0 : // that exists.
4551 0 :
4552 0 : // 2. Actually delete the node from the database and from in-memory state
4553 0 : tracing::info!("Deleting node from database");
4554 0 : self.persistence.delete_node(node_id).await?;
4555 :
4556 0 : Ok(())
4557 0 : }
4558 :
4559 0 : pub(crate) async fn node_list(&self) -> Result<Vec<Node>, ApiError> {
4560 0 : let nodes = {
4561 0 : self.inner
4562 0 : .read()
4563 0 : .unwrap()
4564 0 : .nodes
4565 0 : .values()
4566 0 : .cloned()
4567 0 : .collect::<Vec<_>>()
4568 0 : };
4569 0 :
4570 0 : Ok(nodes)
4571 0 : }
4572 :
4573 0 : pub(crate) async fn get_node(&self, node_id: NodeId) -> Result<Node, ApiError> {
4574 0 : self.inner
4575 0 : .read()
4576 0 : .unwrap()
4577 0 : .nodes
4578 0 : .get(&node_id)
4579 0 : .cloned()
4580 0 : .ok_or(ApiError::NotFound(
4581 0 : format!("Node {node_id} not registered").into(),
4582 0 : ))
4583 0 : }
4584 :
4585 0 : pub(crate) async fn node_register(
4586 0 : &self,
4587 0 : register_req: NodeRegisterRequest,
4588 0 : ) -> Result<(), ApiError> {
4589 0 : let _node_lock = trace_exclusive_lock(
4590 0 : &self.node_op_locks,
4591 0 : register_req.node_id,
4592 0 : NodeOperations::Register,
4593 0 : )
4594 0 : .await;
4595 :
4596 : {
4597 0 : let locked = self.inner.read().unwrap();
4598 0 : if let Some(node) = locked.nodes.get(®ister_req.node_id) {
4599 : // Note that we do not do a total equality of the struct, because we don't require
4600 : // the availability/scheduling states to agree for a POST to be idempotent.
4601 0 : if node.registration_match(®ister_req) {
4602 0 : tracing::info!(
4603 0 : "Node {} re-registered with matching address",
4604 : register_req.node_id
4605 : );
4606 0 : return Ok(());
4607 : } else {
4608 : // TODO: decide if we want to allow modifying node addresses without removing and re-adding
4609 : // the node. Safest/simplest thing is to refuse it, and usually we deploy with
4610 : // a fixed address through the lifetime of a node.
4611 0 : tracing::warn!(
4612 0 : "Node {} tried to register with different address",
4613 : register_req.node_id
4614 : );
4615 0 : return Err(ApiError::Conflict(
4616 0 : "Node is already registered with different address".to_string(),
4617 0 : ));
4618 : }
4619 0 : }
4620 0 : }
4621 0 :
4622 0 : // We do not require that a node is actually online when registered (it will start life
4623 0 : // with it's availability set to Offline), but we _do_ require that its DNS record exists. We're
4624 0 : // therefore not immune to asymmetric L3 connectivity issues, but we are protected against nodes
4625 0 : // that register themselves with a broken DNS config. We check only the HTTP hostname, because
4626 0 : // the postgres hostname might only be resolvable to clients (e.g. if we're on a different VPC than clients).
4627 0 : if tokio::net::lookup_host(format!(
4628 0 : "{}:{}",
4629 0 : register_req.listen_http_addr, register_req.listen_http_port
4630 0 : ))
4631 0 : .await
4632 0 : .is_err()
4633 : {
4634 : // If we have a transient DNS issue, it's up to the caller to retry their registration. Because
4635 : // we can't robustly distinguish between an intermittent issue and a totally bogus DNS situation,
4636 : // we return a soft 503 error, to encourage callers to retry past transient issues.
4637 0 : return Err(ApiError::ResourceUnavailable(
4638 0 : format!(
4639 0 : "Node {} tried to register with unknown DNS name '{}'",
4640 0 : register_req.node_id, register_req.listen_http_addr
4641 0 : )
4642 0 : .into(),
4643 0 : ));
4644 0 : }
4645 0 :
4646 0 : // Ordering: we must persist the new node _before_ adding it to in-memory state.
4647 0 : // This ensures that before we use it for anything or expose it via any external
4648 0 : // API, it is guaranteed to be available after a restart.
4649 0 : let new_node = Node::new(
4650 0 : register_req.node_id,
4651 0 : register_req.listen_http_addr,
4652 0 : register_req.listen_http_port,
4653 0 : register_req.listen_pg_addr,
4654 0 : register_req.listen_pg_port,
4655 0 : );
4656 0 :
4657 0 : // TODO: idempotency if the node already exists in the database
4658 0 : self.persistence.insert_node(&new_node).await?;
4659 :
4660 0 : let mut locked = self.inner.write().unwrap();
4661 0 : let mut new_nodes = (*locked.nodes).clone();
4662 0 :
4663 0 : locked.scheduler.node_upsert(&new_node);
4664 0 : new_nodes.insert(register_req.node_id, new_node);
4665 0 :
4666 0 : locked.nodes = Arc::new(new_nodes);
4667 0 :
4668 0 : tracing::info!(
4669 0 : "Registered pageserver {}, now have {} pageservers",
4670 0 : register_req.node_id,
4671 0 : locked.nodes.len()
4672 : );
4673 0 : Ok(())
4674 0 : }
4675 :
4676 0 : pub(crate) async fn node_configure(
4677 0 : &self,
4678 0 : node_id: NodeId,
4679 0 : availability: Option<NodeAvailability>,
4680 0 : scheduling: Option<NodeSchedulingPolicy>,
4681 0 : ) -> Result<(), ApiError> {
4682 0 : let _node_lock =
4683 0 : trace_exclusive_lock(&self.node_op_locks, node_id, NodeOperations::Configure).await;
4684 :
4685 0 : if let Some(scheduling) = scheduling {
4686 : // Scheduling is a persistent part of Node: we must write updates to the database before
4687 : // applying them in memory
4688 0 : self.persistence.update_node(node_id, scheduling).await?;
4689 0 : }
4690 :
4691 : // If we're activating a node, then before setting it active we must reconcile any shard locations
4692 : // on that node, in case it is out of sync, e.g. due to being unavailable during controller startup,
4693 : // by calling [`Self::node_activate_reconcile`]
4694 : //
4695 : // The transition we calculate here remains valid later in the function because we hold the op lock on the node:
4696 : // nothing else can mutate its availability while we run.
4697 0 : let availability_transition = if let Some(input_availability) = availability {
4698 0 : let (activate_node, availability_transition) = {
4699 0 : let locked = self.inner.read().unwrap();
4700 0 : let Some(node) = locked.nodes.get(&node_id) else {
4701 0 : return Err(ApiError::NotFound(
4702 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
4703 0 : ));
4704 : };
4705 :
4706 0 : (
4707 0 : node.clone(),
4708 0 : node.get_availability_transition(input_availability),
4709 0 : )
4710 : };
4711 :
4712 0 : if matches!(availability_transition, AvailabilityTransition::ToActive) {
4713 0 : self.node_activate_reconcile(activate_node, &_node_lock)
4714 0 : .await?;
4715 0 : }
4716 0 : availability_transition
4717 : } else {
4718 0 : AvailabilityTransition::Unchanged
4719 : };
4720 :
4721 : // Apply changes from the request to our in-memory state for the Node
4722 0 : let mut locked = self.inner.write().unwrap();
4723 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
4724 0 :
4725 0 : let mut new_nodes = (**nodes).clone();
4726 :
4727 0 : let Some(node) = new_nodes.get_mut(&node_id) else {
4728 0 : return Err(ApiError::NotFound(
4729 0 : anyhow::anyhow!("Node not registered").into(),
4730 0 : ));
4731 : };
4732 :
4733 0 : if let Some(availability) = &availability {
4734 0 : node.set_availability(*availability);
4735 0 : }
4736 :
4737 0 : if let Some(scheduling) = scheduling {
4738 0 : node.set_scheduling(scheduling);
4739 0 : }
4740 :
4741 : // Update the scheduler, in case the elegibility of the node for new shards has changed
4742 0 : scheduler.node_upsert(node);
4743 0 :
4744 0 : let new_nodes = Arc::new(new_nodes);
4745 0 :
4746 0 : // Modify scheduling state for any Tenants that are affected by a change in the node's availability state.
4747 0 : match availability_transition {
4748 : AvailabilityTransition::ToOffline => {
4749 0 : tracing::info!("Node {} transition to offline", node_id);
4750 0 : let mut tenants_affected: usize = 0;
4751 :
4752 0 : for (tenant_shard_id, tenant_shard) in tenants {
4753 0 : if let Some(observed_loc) = tenant_shard.observed.locations.get_mut(&node_id) {
4754 0 : // When a node goes offline, we set its observed configuration to None, indicating unknown: we will
4755 0 : // not assume our knowledge of the node's configuration is accurate until it comes back online
4756 0 : observed_loc.conf = None;
4757 0 : }
4758 :
4759 0 : if new_nodes.len() == 1 {
4760 : // Special case for single-node cluster: there is no point trying to reschedule
4761 : // any tenant shards: avoid doing so, in order to avoid spewing warnings about
4762 : // failures to schedule them.
4763 0 : continue;
4764 0 : }
4765 0 :
4766 0 : if !new_nodes
4767 0 : .values()
4768 0 : .any(|n| matches!(n.may_schedule(), MaySchedule::Yes(_)))
4769 : {
4770 : // Special case for when all nodes are unavailable and/or unschedulable: there is no point
4771 : // trying to reschedule since there's nowhere else to go. Without this
4772 : // branch we incorrectly detach tenants in response to node unavailability.
4773 0 : continue;
4774 0 : }
4775 0 :
4776 0 : if tenant_shard.intent.demote_attached(scheduler, node_id) {
4777 0 : tenant_shard.sequence = tenant_shard.sequence.next();
4778 0 :
4779 0 : // TODO: populate a ScheduleContext including all shards in the same tenant_id (only matters
4780 0 : // for tenants without secondary locations: if they have a secondary location, then this
4781 0 : // schedule() call is just promoting an existing secondary)
4782 0 : let mut schedule_context = ScheduleContext::default();
4783 0 :
4784 0 : match tenant_shard.schedule(scheduler, &mut schedule_context) {
4785 0 : Err(e) => {
4786 0 : // It is possible that some tenants will become unschedulable when too many pageservers
4787 0 : // go offline: in this case there isn't much we can do other than make the issue observable.
4788 0 : // TODO: give TenantShard a scheduling error attribute to be queried later.
4789 0 : tracing::warn!(%tenant_shard_id, "Scheduling error when marking pageserver {} offline: {e}", node_id);
4790 : }
4791 : Ok(()) => {
4792 0 : if self
4793 0 : .maybe_reconcile_shard(tenant_shard, &new_nodes)
4794 0 : .is_some()
4795 0 : {
4796 0 : tenants_affected += 1;
4797 0 : };
4798 : }
4799 : }
4800 0 : }
4801 : }
4802 0 : tracing::info!(
4803 0 : "Launched {} reconciler tasks for tenants affected by node {} going offline",
4804 : tenants_affected,
4805 : node_id
4806 : )
4807 : }
4808 : AvailabilityTransition::ToActive => {
4809 0 : tracing::info!("Node {} transition to active", node_id);
4810 : // When a node comes back online, we must reconcile any tenant that has a None observed
4811 : // location on the node.
4812 0 : for tenant_shard in locked.tenants.values_mut() {
4813 : // If a reconciliation is already in progress, rely on the previous scheduling
4814 : // decision and skip triggering a new reconciliation.
4815 0 : if tenant_shard.reconciler.is_some() {
4816 0 : continue;
4817 0 : }
4818 :
4819 0 : if let Some(observed_loc) = tenant_shard.observed.locations.get_mut(&node_id) {
4820 0 : if observed_loc.conf.is_none() {
4821 0 : self.maybe_reconcile_shard(tenant_shard, &new_nodes);
4822 0 : }
4823 0 : }
4824 : }
4825 :
4826 : // TODO: in the background, we should balance work back onto this pageserver
4827 : }
4828 : // No action required for the intermediate unavailable state.
4829 : // When we transition into active or offline from the unavailable state,
4830 : // the correct handling above will kick in.
4831 : AvailabilityTransition::ToWarmingUpFromActive => {
4832 0 : tracing::info!("Node {} transition to unavailable from active", node_id);
4833 : }
4834 : AvailabilityTransition::ToWarmingUpFromOffline => {
4835 0 : tracing::info!("Node {} transition to unavailable from offline", node_id);
4836 : }
4837 : AvailabilityTransition::Unchanged => {
4838 0 : tracing::debug!("Node {} no availability change during config", node_id);
4839 : }
4840 : }
4841 :
4842 0 : locked.nodes = new_nodes;
4843 0 :
4844 0 : Ok(())
4845 0 : }
4846 :
4847 0 : pub(crate) async fn start_node_drain(
4848 0 : self: &Arc<Self>,
4849 0 : node_id: NodeId,
4850 0 : ) -> Result<(), ApiError> {
4851 0 : let (ongoing_op, node_available, node_policy, schedulable_nodes_count) = {
4852 0 : let locked = self.inner.read().unwrap();
4853 0 : let nodes = &locked.nodes;
4854 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
4855 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
4856 0 : ))?;
4857 0 : let schedulable_nodes_count = nodes
4858 0 : .iter()
4859 0 : .filter(|(_, n)| matches!(n.may_schedule(), MaySchedule::Yes(_)))
4860 0 : .count();
4861 0 :
4862 0 : (
4863 0 : locked
4864 0 : .ongoing_operation
4865 0 : .as_ref()
4866 0 : .map(|ongoing| ongoing.operation),
4867 0 : node.is_available(),
4868 0 : node.get_scheduling(),
4869 0 : schedulable_nodes_count,
4870 0 : )
4871 0 : };
4872 :
4873 0 : if let Some(ongoing) = ongoing_op {
4874 0 : return Err(ApiError::PreconditionFailed(
4875 0 : format!("Background operation already ongoing for node: {}", ongoing).into(),
4876 0 : ));
4877 0 : }
4878 0 :
4879 0 : if !node_available {
4880 0 : return Err(ApiError::ResourceUnavailable(
4881 0 : format!("Node {node_id} is currently unavailable").into(),
4882 0 : ));
4883 0 : }
4884 0 :
4885 0 : if schedulable_nodes_count == 0 {
4886 0 : return Err(ApiError::PreconditionFailed(
4887 0 : "No other schedulable nodes to drain to".into(),
4888 0 : ));
4889 0 : }
4890 0 :
4891 0 : match node_policy {
4892 : NodeSchedulingPolicy::Active | NodeSchedulingPolicy::Pause => {
4893 0 : self.node_configure(node_id, None, Some(NodeSchedulingPolicy::Draining))
4894 0 : .await?;
4895 :
4896 0 : let cancel = self.cancel.child_token();
4897 0 : let gate_guard = self.gate.enter().map_err(|_| ApiError::ShuttingDown)?;
4898 :
4899 0 : self.inner.write().unwrap().ongoing_operation = Some(OperationHandler {
4900 0 : operation: Operation::Drain(Drain { node_id }),
4901 0 : cancel: cancel.clone(),
4902 0 : });
4903 0 :
4904 0 : tokio::task::spawn({
4905 0 : let service = self.clone();
4906 0 : let cancel = cancel.clone();
4907 0 : async move {
4908 0 : let _gate_guard = gate_guard;
4909 :
4910 : scopeguard::defer! {
4911 : let prev = service.inner.write().unwrap().ongoing_operation.take();
4912 :
4913 0 : if let Some(Operation::Drain(removed_drain)) = prev.map(|h| h.operation) {
4914 : assert_eq!(removed_drain.node_id, node_id, "We always take the same operation");
4915 : } else {
4916 : panic!("We always remove the same operation")
4917 : }
4918 : }
4919 :
4920 0 : tracing::info!(%node_id, "Drain background operation starting");
4921 0 : let res = service.drain_node(node_id, cancel).await;
4922 0 : match res {
4923 : Ok(()) => {
4924 0 : tracing::info!(%node_id, "Drain background operation completed successfully");
4925 : }
4926 : Err(OperationError::Cancelled) => {
4927 0 : tracing::info!(%node_id, "Drain background operation was cancelled");
4928 : }
4929 0 : Err(err) => {
4930 0 : tracing::error!(%node_id, "Drain background operation encountered: {err}")
4931 : }
4932 : }
4933 0 : }
4934 0 : });
4935 0 : }
4936 : NodeSchedulingPolicy::Draining => {
4937 0 : return Err(ApiError::Conflict(format!(
4938 0 : "Node {node_id} has drain in progress"
4939 0 : )));
4940 : }
4941 0 : policy => {
4942 0 : return Err(ApiError::PreconditionFailed(
4943 0 : format!("Node {node_id} cannot be drained due to {policy:?} policy").into(),
4944 0 : ));
4945 : }
4946 : }
4947 :
4948 0 : Ok(())
4949 0 : }
4950 :
4951 0 : pub(crate) async fn cancel_node_drain(&self, node_id: NodeId) -> Result<(), ApiError> {
4952 0 : let (node_available, node_policy) = {
4953 0 : let locked = self.inner.read().unwrap();
4954 0 : let nodes = &locked.nodes;
4955 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
4956 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
4957 0 : ))?;
4958 :
4959 0 : (node.is_available(), node.get_scheduling())
4960 0 : };
4961 0 :
4962 0 : if !node_available {
4963 0 : return Err(ApiError::ResourceUnavailable(
4964 0 : format!("Node {node_id} is currently unavailable").into(),
4965 0 : ));
4966 0 : }
4967 :
4968 0 : if !matches!(node_policy, NodeSchedulingPolicy::Draining) {
4969 0 : return Err(ApiError::PreconditionFailed(
4970 0 : format!("Node {node_id} has no drain in progress").into(),
4971 0 : ));
4972 0 : }
4973 :
4974 0 : if let Some(op_handler) = self.inner.read().unwrap().ongoing_operation.as_ref() {
4975 0 : if let Operation::Drain(drain) = op_handler.operation {
4976 0 : if drain.node_id == node_id {
4977 0 : tracing::info!("Cancelling background drain operation for node {node_id}");
4978 0 : op_handler.cancel.cancel();
4979 0 : return Ok(());
4980 0 : }
4981 0 : }
4982 0 : }
4983 :
4984 0 : Err(ApiError::PreconditionFailed(
4985 0 : format!("Node {node_id} has no drain in progress").into(),
4986 0 : ))
4987 0 : }
4988 :
4989 0 : pub(crate) async fn start_node_fill(self: &Arc<Self>, node_id: NodeId) -> Result<(), ApiError> {
4990 0 : let (ongoing_op, node_available, node_policy, total_nodes_count) = {
4991 0 : let locked = self.inner.read().unwrap();
4992 0 : let nodes = &locked.nodes;
4993 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
4994 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
4995 0 : ))?;
4996 :
4997 0 : (
4998 0 : locked
4999 0 : .ongoing_operation
5000 0 : .as_ref()
5001 0 : .map(|ongoing| ongoing.operation),
5002 0 : node.is_available(),
5003 0 : node.get_scheduling(),
5004 0 : nodes.len(),
5005 0 : )
5006 0 : };
5007 :
5008 0 : if let Some(ongoing) = ongoing_op {
5009 0 : return Err(ApiError::PreconditionFailed(
5010 0 : format!("Background operation already ongoing for node: {}", ongoing).into(),
5011 0 : ));
5012 0 : }
5013 0 :
5014 0 : if !node_available {
5015 0 : return Err(ApiError::ResourceUnavailable(
5016 0 : format!("Node {node_id} is currently unavailable").into(),
5017 0 : ));
5018 0 : }
5019 0 :
5020 0 : if total_nodes_count <= 1 {
5021 0 : return Err(ApiError::PreconditionFailed(
5022 0 : "No other nodes to fill from".into(),
5023 0 : ));
5024 0 : }
5025 0 :
5026 0 : match node_policy {
5027 : NodeSchedulingPolicy::Active => {
5028 0 : self.node_configure(node_id, None, Some(NodeSchedulingPolicy::Filling))
5029 0 : .await?;
5030 :
5031 0 : let cancel = self.cancel.child_token();
5032 0 : let gate_guard = self.gate.enter().map_err(|_| ApiError::ShuttingDown)?;
5033 :
5034 0 : self.inner.write().unwrap().ongoing_operation = Some(OperationHandler {
5035 0 : operation: Operation::Fill(Fill { node_id }),
5036 0 : cancel: cancel.clone(),
5037 0 : });
5038 0 :
5039 0 : tokio::task::spawn({
5040 0 : let service = self.clone();
5041 0 : let cancel = cancel.clone();
5042 0 : async move {
5043 0 : let _gate_guard = gate_guard;
5044 :
5045 : scopeguard::defer! {
5046 : let prev = service.inner.write().unwrap().ongoing_operation.take();
5047 :
5048 0 : if let Some(Operation::Fill(removed_fill)) = prev.map(|h| h.operation) {
5049 : assert_eq!(removed_fill.node_id, node_id, "We always take the same operation");
5050 : } else {
5051 : panic!("We always remove the same operation")
5052 : }
5053 : }
5054 :
5055 0 : tracing::info!(%node_id, "Fill background operation starting");
5056 0 : let res = service.fill_node(node_id, cancel).await;
5057 0 : match res {
5058 : Ok(()) => {
5059 0 : tracing::info!(%node_id, "Fill background operation completed successfully");
5060 : }
5061 : Err(OperationError::Cancelled) => {
5062 0 : tracing::info!(%node_id, "Fill background operation was cancelled");
5063 : }
5064 0 : Err(err) => {
5065 0 : tracing::error!(%node_id, "Fill background operation encountered: {err}")
5066 : }
5067 : }
5068 0 : }
5069 0 : });
5070 0 : }
5071 : NodeSchedulingPolicy::Filling => {
5072 0 : return Err(ApiError::Conflict(format!(
5073 0 : "Node {node_id} has fill in progress"
5074 0 : )));
5075 : }
5076 0 : policy => {
5077 0 : return Err(ApiError::PreconditionFailed(
5078 0 : format!("Node {node_id} cannot be filled due to {policy:?} policy").into(),
5079 0 : ));
5080 : }
5081 : }
5082 :
5083 0 : Ok(())
5084 0 : }
5085 :
5086 0 : pub(crate) async fn cancel_node_fill(&self, node_id: NodeId) -> Result<(), ApiError> {
5087 0 : let (node_available, node_policy) = {
5088 0 : let locked = self.inner.read().unwrap();
5089 0 : let nodes = &locked.nodes;
5090 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
5091 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
5092 0 : ))?;
5093 :
5094 0 : (node.is_available(), node.get_scheduling())
5095 0 : };
5096 0 :
5097 0 : if !node_available {
5098 0 : return Err(ApiError::ResourceUnavailable(
5099 0 : format!("Node {node_id} is currently unavailable").into(),
5100 0 : ));
5101 0 : }
5102 :
5103 0 : if !matches!(node_policy, NodeSchedulingPolicy::Filling) {
5104 0 : return Err(ApiError::PreconditionFailed(
5105 0 : format!("Node {node_id} has no fill in progress").into(),
5106 0 : ));
5107 0 : }
5108 :
5109 0 : if let Some(op_handler) = self.inner.read().unwrap().ongoing_operation.as_ref() {
5110 0 : if let Operation::Fill(fill) = op_handler.operation {
5111 0 : if fill.node_id == node_id {
5112 0 : tracing::info!("Cancelling background drain operation for node {node_id}");
5113 0 : op_handler.cancel.cancel();
5114 0 : return Ok(());
5115 0 : }
5116 0 : }
5117 0 : }
5118 :
5119 0 : Err(ApiError::PreconditionFailed(
5120 0 : format!("Node {node_id} has no fill in progress").into(),
5121 0 : ))
5122 0 : }
5123 :
5124 : /// Helper for methods that will try and call pageserver APIs for
5125 : /// a tenant, such as timeline CRUD: they cannot proceed unless the tenant
5126 : /// is attached somewhere.
5127 0 : fn ensure_attached_schedule(
5128 0 : &self,
5129 0 : mut locked: std::sync::RwLockWriteGuard<'_, ServiceState>,
5130 0 : tenant_id: TenantId,
5131 0 : ) -> Result<Vec<ReconcilerWaiter>, anyhow::Error> {
5132 0 : let mut waiters = Vec::new();
5133 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
5134 0 :
5135 0 : let mut schedule_context = ScheduleContext::default();
5136 0 : for (tenant_shard_id, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
5137 0 : shard.schedule(scheduler, &mut schedule_context)?;
5138 :
5139 : // The shard's policies may not result in an attached location being scheduled: this
5140 : // is an error because our caller needs it attached somewhere.
5141 0 : if shard.intent.get_attached().is_none() {
5142 0 : return Err(anyhow::anyhow!(
5143 0 : "Tenant {tenant_id} not scheduled to be attached"
5144 0 : ));
5145 0 : };
5146 0 :
5147 0 : if shard.stably_attached().is_some() {
5148 : // We do not require the shard to be totally up to date on reconciliation: we just require
5149 : // that it has been attached on the intended node. Other dirty state such as unattached secondary
5150 : // locations, or compute hook notifications can be ignored.
5151 0 : continue;
5152 0 : }
5153 :
5154 0 : if let Some(waiter) = self.maybe_reconcile_shard(shard, nodes) {
5155 0 : tracing::info!("Waiting for shard {tenant_shard_id} to reconcile, in order to ensure it is attached");
5156 0 : waiters.push(waiter);
5157 0 : }
5158 : }
5159 0 : Ok(waiters)
5160 0 : }
5161 :
5162 0 : async fn ensure_attached_wait(&self, tenant_id: TenantId) -> Result<(), ApiError> {
5163 0 : let ensure_waiters = {
5164 0 : let locked = self.inner.write().unwrap();
5165 :
5166 : // Check if the tenant is splitting: in this case, even if it is attached,
5167 : // we must act as if it is not: this blocks e.g. timeline creation/deletion
5168 : // operations during the split.
5169 0 : for (_shard_id, shard) in locked.tenants.range(TenantShardId::tenant_range(tenant_id)) {
5170 0 : if !matches!(shard.splitting, SplitState::Idle) {
5171 0 : return Err(ApiError::ResourceUnavailable(
5172 0 : "Tenant shards are currently splitting".into(),
5173 0 : ));
5174 0 : }
5175 : }
5176 :
5177 0 : self.ensure_attached_schedule(locked, tenant_id)
5178 0 : .map_err(ApiError::InternalServerError)?
5179 : };
5180 :
5181 0 : let deadline = Instant::now().checked_add(Duration::from_secs(5)).unwrap();
5182 0 : for waiter in ensure_waiters {
5183 0 : let timeout = deadline.duration_since(Instant::now());
5184 0 : waiter.wait_timeout(timeout).await?;
5185 : }
5186 :
5187 0 : Ok(())
5188 0 : }
5189 :
5190 : /// Wrap [`TenantShard`] reconciliation methods with acquisition of [`Gate`] and [`ReconcileUnits`],
5191 0 : fn maybe_reconcile_shard(
5192 0 : &self,
5193 0 : shard: &mut TenantShard,
5194 0 : nodes: &Arc<HashMap<NodeId, Node>>,
5195 0 : ) -> Option<ReconcilerWaiter> {
5196 0 : let reconcile_needed = shard.get_reconcile_needed(nodes);
5197 0 :
5198 0 : match reconcile_needed {
5199 0 : ReconcileNeeded::No => return None,
5200 0 : ReconcileNeeded::WaitExisting(waiter) => return Some(waiter),
5201 0 : ReconcileNeeded::Yes => {
5202 0 : // Fall through to try and acquire units for spawning reconciler
5203 0 : }
5204 : };
5205 :
5206 0 : let units = match self.reconciler_concurrency.clone().try_acquire_owned() {
5207 0 : Ok(u) => ReconcileUnits::new(u),
5208 : Err(_) => {
5209 0 : tracing::info!(tenant_id=%shard.tenant_shard_id.tenant_id, shard_id=%shard.tenant_shard_id.shard_slug(),
5210 0 : "Concurrency limited: enqueued for reconcile later");
5211 0 : if !shard.delayed_reconcile {
5212 0 : match self.delayed_reconcile_tx.try_send(shard.tenant_shard_id) {
5213 0 : Err(TrySendError::Closed(_)) => {
5214 0 : // Weird mid-shutdown case?
5215 0 : }
5216 : Err(TrySendError::Full(_)) => {
5217 : // It is safe to skip sending our ID in the channel: we will eventually get retried by the background reconcile task.
5218 0 : tracing::warn!(
5219 0 : "Many shards are waiting to reconcile: delayed_reconcile queue is full"
5220 : );
5221 : }
5222 0 : Ok(()) => {
5223 0 : shard.delayed_reconcile = true;
5224 0 : }
5225 : }
5226 0 : }
5227 :
5228 : // We won't spawn a reconciler, but we will construct a waiter that waits for the shard's sequence
5229 : // number to advance. When this function is eventually called again and succeeds in getting units,
5230 : // it will spawn a reconciler that makes this waiter complete.
5231 0 : return Some(shard.future_reconcile_waiter());
5232 : }
5233 : };
5234 :
5235 0 : let Ok(gate_guard) = self.reconcilers_gate.enter() else {
5236 : // Gate closed: we're shutting down, drop out.
5237 0 : return None;
5238 : };
5239 :
5240 0 : shard.spawn_reconciler(
5241 0 : &self.result_tx,
5242 0 : nodes,
5243 0 : &self.compute_hook,
5244 0 : &self.config,
5245 0 : &self.persistence,
5246 0 : units,
5247 0 : gate_guard,
5248 0 : &self.reconcilers_cancel,
5249 0 : )
5250 0 : }
5251 :
5252 : /// Check all tenants for pending reconciliation work, and reconcile those in need.
5253 : /// Additionally, reschedule tenants that require it.
5254 : ///
5255 : /// Returns how many reconciliation tasks were started, or `1` if no reconciles were
5256 : /// spawned but some _would_ have been spawned if `reconciler_concurrency` units where
5257 : /// available. A return value of 0 indicates that everything is fully reconciled already.
5258 0 : fn reconcile_all(&self) -> usize {
5259 0 : let mut locked = self.inner.write().unwrap();
5260 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
5261 0 : let pageservers = nodes.clone();
5262 0 :
5263 0 : let mut schedule_context = ScheduleContext::default();
5264 0 :
5265 0 : let mut reconciles_spawned = 0;
5266 0 : for (tenant_shard_id, shard) in tenants.iter_mut() {
5267 0 : if tenant_shard_id.is_shard_zero() {
5268 0 : schedule_context = ScheduleContext::default();
5269 0 : }
5270 :
5271 : // Skip checking if this shard is already enqueued for reconciliation
5272 0 : if shard.delayed_reconcile && self.reconciler_concurrency.available_permits() == 0 {
5273 : // If there is something delayed, then return a nonzero count so that
5274 : // callers like reconcile_all_now do not incorrectly get the impression
5275 : // that the system is in a quiescent state.
5276 0 : reconciles_spawned = std::cmp::max(1, reconciles_spawned);
5277 0 : continue;
5278 0 : }
5279 0 :
5280 0 : // Eventual consistency: if an earlier reconcile job failed, and the shard is still
5281 0 : // dirty, spawn another rone
5282 0 : if self.maybe_reconcile_shard(shard, &pageservers).is_some() {
5283 0 : reconciles_spawned += 1;
5284 0 : }
5285 :
5286 0 : schedule_context.avoid(&shard.intent.all_pageservers());
5287 : }
5288 :
5289 0 : reconciles_spawned
5290 0 : }
5291 :
5292 : /// `optimize` in this context means identifying shards which have valid scheduled locations, but
5293 : /// could be scheduled somewhere better:
5294 : /// - Cutting over to a secondary if the node with the secondary is more lightly loaded
5295 : /// * e.g. after a node fails then recovers, to move some work back to it
5296 : /// - Cutting over to a secondary if it improves the spread of shard attachments within a tenant
5297 : /// * e.g. after a shard split, the initial attached locations will all be on the node where
5298 : /// we did the split, but are probably better placed elsewhere.
5299 : /// - Creating new secondary locations if it improves the spreading of a sharded tenant
5300 : /// * e.g. after a shard split, some locations will be on the same node (where the split
5301 : /// happened), and will probably be better placed elsewhere.
5302 : ///
5303 : /// To put it more briefly: whereas the scheduler respects soft constraints in a ScheduleContext at
5304 : /// the time of scheduling, this function looks for cases where a better-scoring location is available
5305 : /// according to those same soft constraints.
5306 0 : async fn optimize_all(&self) -> usize {
5307 0 : // Limit on how many shards' optmizations each call to this function will execute. Combined
5308 0 : // with the frequency of background calls, this acts as an implicit rate limit that runs a small
5309 0 : // trickle of optimizations in the background, rather than executing a large number in parallel
5310 0 : // when a change occurs.
5311 0 : const MAX_OPTIMIZATIONS_EXEC_PER_PASS: usize = 2;
5312 0 :
5313 0 : // Synchronous prepare: scan shards for possible scheduling optimizations
5314 0 : let candidate_work = self.optimize_all_plan();
5315 0 : let candidate_work_len = candidate_work.len();
5316 :
5317 : // Asynchronous validate: I/O to pageservers to make sure shards are in a good state to apply validation
5318 0 : let validated_work = self.optimize_all_validate(candidate_work).await;
5319 :
5320 0 : let was_work_filtered = validated_work.len() != candidate_work_len;
5321 0 :
5322 0 : // Synchronous apply: update the shards' intent states according to validated optimisations
5323 0 : let mut reconciles_spawned = 0;
5324 0 : let mut optimizations_applied = 0;
5325 0 : let mut locked = self.inner.write().unwrap();
5326 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
5327 0 : for (tenant_shard_id, optimization) in validated_work {
5328 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
5329 : // Shard was dropped between planning and execution;
5330 0 : continue;
5331 : };
5332 0 : if shard.apply_optimization(scheduler, optimization) {
5333 0 : optimizations_applied += 1;
5334 0 : if self.maybe_reconcile_shard(shard, nodes).is_some() {
5335 0 : reconciles_spawned += 1;
5336 0 : }
5337 0 : }
5338 :
5339 0 : if optimizations_applied >= MAX_OPTIMIZATIONS_EXEC_PER_PASS {
5340 0 : break;
5341 0 : }
5342 : }
5343 :
5344 0 : if was_work_filtered {
5345 0 : // If we filtered any work out during validation, ensure we return a nonzero value to indicate
5346 0 : // to callers that the system is not in a truly quiet state, it's going to do some work as soon
5347 0 : // as these validations start passing.
5348 0 : reconciles_spawned = std::cmp::max(reconciles_spawned, 1);
5349 0 : }
5350 :
5351 0 : reconciles_spawned
5352 0 : }
5353 :
5354 0 : fn optimize_all_plan(&self) -> Vec<(TenantShardId, ScheduleOptimization)> {
5355 0 : let mut schedule_context = ScheduleContext::default();
5356 0 :
5357 0 : let mut tenant_shards: Vec<&TenantShard> = Vec::new();
5358 0 :
5359 0 : // How many candidate optimizations we will generate, before evaluating them for readniess: setting
5360 0 : // this higher than the execution limit gives us a chance to execute some work even if the first
5361 0 : // few optimizations we find are not ready.
5362 0 : const MAX_OPTIMIZATIONS_PLAN_PER_PASS: usize = 8;
5363 0 :
5364 0 : let mut work = Vec::new();
5365 0 :
5366 0 : let mut locked = self.inner.write().unwrap();
5367 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
5368 0 : for (tenant_shard_id, shard) in tenants.iter() {
5369 0 : if tenant_shard_id.is_shard_zero() {
5370 0 : // Reset accumulators on the first shard in a tenant
5371 0 : schedule_context = ScheduleContext::default();
5372 0 : schedule_context.mode = ScheduleMode::Speculative;
5373 0 : tenant_shards.clear();
5374 0 : }
5375 :
5376 0 : if work.len() >= MAX_OPTIMIZATIONS_PLAN_PER_PASS {
5377 0 : break;
5378 0 : }
5379 0 :
5380 0 : match shard.get_scheduling_policy() {
5381 0 : ShardSchedulingPolicy::Active => {
5382 0 : // Ok to do optimization
5383 0 : }
5384 : ShardSchedulingPolicy::Essential
5385 : | ShardSchedulingPolicy::Pause
5386 : | ShardSchedulingPolicy::Stop => {
5387 : // Policy prevents optimizing this shard.
5388 0 : continue;
5389 : }
5390 : }
5391 :
5392 : // Accumulate the schedule context for all the shards in a tenant: we must have
5393 : // the total view of all shards before we can try to optimize any of them.
5394 0 : schedule_context.avoid(&shard.intent.all_pageservers());
5395 0 : if let Some(attached) = shard.intent.get_attached() {
5396 0 : schedule_context.push_attached(*attached);
5397 0 : }
5398 0 : tenant_shards.push(shard);
5399 0 :
5400 0 : // Once we have seen the last shard in the tenant, proceed to search across all shards
5401 0 : // in the tenant for optimizations
5402 0 : if shard.shard.number.0 == shard.shard.count.count() - 1 {
5403 0 : if tenant_shards.iter().any(|s| s.reconciler.is_some()) {
5404 : // Do not start any optimizations while another change to the tenant is ongoing: this
5405 : // is not necessary for correctness, but simplifies operations and implicitly throttles
5406 : // optimization changes to happen in a "trickle" over time.
5407 0 : continue;
5408 0 : }
5409 0 :
5410 0 : if tenant_shards.iter().any(|s| {
5411 0 : !matches!(s.splitting, SplitState::Idle)
5412 0 : || matches!(s.policy, PlacementPolicy::Detached)
5413 0 : }) {
5414 : // Never attempt to optimize a tenant that is currently being split, or
5415 : // a tenant that is meant to be detached
5416 0 : continue;
5417 0 : }
5418 :
5419 : // TODO: optimization calculations are relatively expensive: create some fast-path for
5420 : // the common idle case (avoiding the search on tenants that we have recently checked)
5421 :
5422 0 : for shard in &tenant_shards {
5423 0 : if let Some(optimization) =
5424 : // If idle, maybe ptimize attachments: if a shard has a secondary location that is preferable to
5425 : // its primary location based on soft constraints, cut it over.
5426 0 : shard.optimize_attachment(nodes, &schedule_context)
5427 : {
5428 0 : work.push((shard.tenant_shard_id, optimization));
5429 0 : break;
5430 0 : } else if let Some(optimization) =
5431 : // If idle, maybe optimize secondary locations: if a shard has a secondary location that would be
5432 : // better placed on another node, based on ScheduleContext, then adjust it. This
5433 : // covers cases like after a shard split, where we might have too many shards
5434 : // in the same tenant with secondary locations on the node where they originally split.
5435 0 : shard.optimize_secondary(scheduler, &schedule_context)
5436 : {
5437 0 : work.push((shard.tenant_shard_id, optimization));
5438 0 : break;
5439 0 : }
5440 :
5441 : // TODO: extend this mechanism to prefer attaching on nodes with fewer attached
5442 : // tenants (i.e. extend schedule state to distinguish attached from secondary counts),
5443 : // for the total number of attachments on a node (not just within a tenant.)
5444 : }
5445 0 : }
5446 : }
5447 :
5448 0 : work
5449 0 : }
5450 :
5451 0 : async fn optimize_all_validate(
5452 0 : &self,
5453 0 : candidate_work: Vec<(TenantShardId, ScheduleOptimization)>,
5454 0 : ) -> Vec<(TenantShardId, ScheduleOptimization)> {
5455 0 : // Take a clone of the node map to use outside the lock in async validation phase
5456 0 : let validation_nodes = { self.inner.read().unwrap().nodes.clone() };
5457 0 :
5458 0 : let mut want_secondary_status = Vec::new();
5459 0 :
5460 0 : // Validate our plans: this is an async phase where we may do I/O to pageservers to
5461 0 : // check that the state of locations is acceptable to run the optimization, such as
5462 0 : // checking that a secondary location is sufficiently warmed-up to cleanly cut over
5463 0 : // in a live migration.
5464 0 : let mut validated_work = Vec::new();
5465 0 : for (tenant_shard_id, optimization) in candidate_work {
5466 0 : match optimization.action {
5467 : ScheduleOptimizationAction::MigrateAttachment(MigrateAttachment {
5468 : old_attached_node_id: _,
5469 0 : new_attached_node_id,
5470 0 : }) => {
5471 0 : match validation_nodes.get(&new_attached_node_id) {
5472 0 : None => {
5473 0 : // Node was dropped between planning and validation
5474 0 : }
5475 0 : Some(node) => {
5476 0 : if !node.is_available() {
5477 0 : tracing::info!("Skipping optimization migration of {tenant_shard_id} to {new_attached_node_id} because node unavailable");
5478 0 : } else {
5479 0 : // Accumulate optimizations that require fetching secondary status, so that we can execute these
5480 0 : // remote API requests concurrently.
5481 0 : want_secondary_status.push((
5482 0 : tenant_shard_id,
5483 0 : node.clone(),
5484 0 : optimization,
5485 0 : ));
5486 0 : }
5487 : }
5488 : }
5489 : }
5490 : ScheduleOptimizationAction::ReplaceSecondary(_) => {
5491 : // No extra checks needed to replace a secondary: this does not interrupt client access
5492 0 : validated_work.push((tenant_shard_id, optimization))
5493 : }
5494 : };
5495 : }
5496 :
5497 : // Call into pageserver API to find out if the destination secondary location is warm enough for a reasonably smooth migration: we
5498 : // do this so that we avoid spawning a Reconciler that would have to wait minutes/hours for a destination to warm up: that reconciler
5499 : // would hold a precious reconcile semaphore unit the whole time it was waiting for the destination to warm up.
5500 0 : let results = self
5501 0 : .tenant_for_shards_api(
5502 0 : want_secondary_status
5503 0 : .iter()
5504 0 : .map(|i| (i.0, i.1.clone()))
5505 0 : .collect(),
5506 0 : |tenant_shard_id, client| async move {
5507 0 : client.tenant_secondary_status(tenant_shard_id).await
5508 0 : },
5509 0 : 1,
5510 0 : 1,
5511 0 : SHORT_RECONCILE_TIMEOUT,
5512 0 : &self.cancel,
5513 0 : )
5514 0 : .await;
5515 :
5516 0 : for ((tenant_shard_id, node, optimization), secondary_status) in
5517 0 : want_secondary_status.into_iter().zip(results.into_iter())
5518 : {
5519 0 : match secondary_status {
5520 0 : Err(e) => {
5521 0 : tracing::info!("Skipping migration of {tenant_shard_id} to {node}, error querying secondary: {e}");
5522 : }
5523 0 : Ok(progress) => {
5524 0 : // We require secondary locations to have less than 10GiB of downloads pending before we will use
5525 0 : // them in an optimization
5526 0 : const DOWNLOAD_FRESHNESS_THRESHOLD: u64 = 10 * 1024 * 1024 * 1024;
5527 0 :
5528 0 : if progress.heatmap_mtime.is_none()
5529 0 : || progress.bytes_total < DOWNLOAD_FRESHNESS_THRESHOLD
5530 0 : && progress.bytes_downloaded != progress.bytes_total
5531 0 : || progress.bytes_total - progress.bytes_downloaded
5532 0 : > DOWNLOAD_FRESHNESS_THRESHOLD
5533 : {
5534 0 : tracing::info!("Skipping migration of {tenant_shard_id} to {node} because secondary isn't ready: {progress:?}");
5535 : } else {
5536 : // Location looks ready: proceed
5537 0 : tracing::info!(
5538 0 : "{tenant_shard_id} secondary on {node} is warm enough for migration: {progress:?}"
5539 : );
5540 0 : validated_work.push((tenant_shard_id, optimization))
5541 : }
5542 : }
5543 : }
5544 : }
5545 :
5546 0 : validated_work
5547 0 : }
5548 :
5549 : /// Look for shards which are oversized and in need of splitting
5550 0 : async fn autosplit_tenants(self: &Arc<Self>) {
5551 0 : let Some(split_threshold) = self.config.split_threshold else {
5552 : // Auto-splitting is disabled
5553 0 : return;
5554 : };
5555 :
5556 0 : let nodes = self.inner.read().unwrap().nodes.clone();
5557 0 :
5558 0 : const SPLIT_TO_MAX: ShardCount = ShardCount::new(8);
5559 0 :
5560 0 : let mut top_n = Vec::new();
5561 0 :
5562 0 : // Call into each node to look for big tenants
5563 0 : let top_n_request = TopTenantShardsRequest {
5564 0 : // We currently split based on logical size, for simplicity: logical size is a signal of
5565 0 : // the user's intent to run a large database, whereas physical/resident size can be symptoms
5566 0 : // of compaction issues. Eventually we should switch to using resident size to bound the
5567 0 : // disk space impact of one shard.
5568 0 : order_by: models::TenantSorting::MaxLogicalSize,
5569 0 : limit: 10,
5570 0 : where_shards_lt: Some(SPLIT_TO_MAX),
5571 0 : where_gt: Some(split_threshold),
5572 0 : };
5573 0 : for node in nodes.values() {
5574 0 : let request_ref = &top_n_request;
5575 0 : match node
5576 0 : .with_client_retries(
5577 0 : |client| async move {
5578 0 : let request = request_ref.clone();
5579 0 : client.top_tenant_shards(request.clone()).await
5580 0 : },
5581 0 : &self.config.jwt_token,
5582 0 : 3,
5583 0 : 3,
5584 0 : Duration::from_secs(5),
5585 0 : &self.cancel,
5586 0 : )
5587 0 : .await
5588 : {
5589 0 : Some(Ok(node_top_n)) => {
5590 0 : top_n.extend(node_top_n.shards.into_iter());
5591 0 : }
5592 : Some(Err(mgmt_api::Error::Cancelled)) => {
5593 0 : continue;
5594 : }
5595 0 : Some(Err(e)) => {
5596 0 : tracing::warn!("Failed to fetch top N tenants from {node}: {e}");
5597 0 : continue;
5598 : }
5599 : None => {
5600 : // Node is shutting down
5601 0 : continue;
5602 : }
5603 : };
5604 : }
5605 :
5606 : // Pick the biggest tenant to split first
5607 0 : top_n.sort_by_key(|i| i.resident_size);
5608 0 : let Some(split_candidate) = top_n.into_iter().next() else {
5609 0 : tracing::debug!("No split-elegible shards found");
5610 0 : return;
5611 : };
5612 :
5613 : // We spawn a task to run this, so it's exactly like some external API client requesting it. We don't
5614 : // want to block the background reconcile loop on this.
5615 0 : tracing::info!("Auto-splitting tenant for size threshold {split_threshold}: current size {split_candidate:?}");
5616 :
5617 0 : let this = self.clone();
5618 0 : tokio::spawn(
5619 0 : async move {
5620 0 : match this
5621 0 : .tenant_shard_split(
5622 0 : split_candidate.id.tenant_id,
5623 0 : TenantShardSplitRequest {
5624 0 : // Always split to the max number of shards: this avoids stepping through
5625 0 : // intervening shard counts and encountering the overrhead of a split+cleanup
5626 0 : // each time as a tenant grows, and is not too expensive because our max shard
5627 0 : // count is relatively low anyway.
5628 0 : // This policy will be adjusted in future once we support higher shard count.
5629 0 : new_shard_count: SPLIT_TO_MAX.literal(),
5630 0 : new_stripe_size: Some(ShardParameters::DEFAULT_STRIPE_SIZE),
5631 0 : },
5632 0 : )
5633 0 : .await
5634 : {
5635 : Ok(_) => {
5636 0 : tracing::info!("Successful auto-split");
5637 : }
5638 0 : Err(e) => {
5639 0 : tracing::error!("Auto-split failed: {e}");
5640 : }
5641 : }
5642 0 : }
5643 0 : .instrument(tracing::info_span!("auto_split", tenant_id=%split_candidate.id.tenant_id)),
5644 : );
5645 0 : }
5646 :
5647 : /// Useful for tests: run whatever work a background [`Self::reconcile_all`] would have done, but
5648 : /// also wait for any generated Reconcilers to complete. Calling this until it returns zero should
5649 : /// put the system into a quiescent state where future background reconciliations won't do anything.
5650 0 : pub(crate) async fn reconcile_all_now(&self) -> Result<usize, ReconcileWaitError> {
5651 0 : let reconciles_spawned = self.reconcile_all();
5652 0 : let reconciles_spawned = if reconciles_spawned == 0 {
5653 : // Only optimize when we are otherwise idle
5654 0 : self.optimize_all().await
5655 : } else {
5656 0 : reconciles_spawned
5657 : };
5658 :
5659 0 : let waiters = {
5660 0 : let mut waiters = Vec::new();
5661 0 : let locked = self.inner.read().unwrap();
5662 0 : for (_tenant_shard_id, shard) in locked.tenants.iter() {
5663 0 : if let Some(waiter) = shard.get_waiter() {
5664 0 : waiters.push(waiter);
5665 0 : }
5666 : }
5667 0 : waiters
5668 0 : };
5669 0 :
5670 0 : let waiter_count = waiters.len();
5671 0 : match self.await_waiters(waiters, RECONCILE_TIMEOUT).await {
5672 0 : Ok(()) => {}
5673 0 : Err(ReconcileWaitError::Failed(_, reconcile_error))
5674 0 : if matches!(*reconcile_error, ReconcileError::Cancel) =>
5675 0 : {
5676 0 : // Ignore reconciler cancel errors: this reconciler might have shut down
5677 0 : // because some other change superceded it. We will return a nonzero number,
5678 0 : // so the caller knows they might have to call again to quiesce the system.
5679 0 : }
5680 0 : Err(e) => {
5681 0 : return Err(e);
5682 : }
5683 : };
5684 :
5685 0 : tracing::info!(
5686 0 : "{} reconciles in reconcile_all, {} waiters",
5687 : reconciles_spawned,
5688 : waiter_count
5689 : );
5690 :
5691 0 : Ok(std::cmp::max(waiter_count, reconciles_spawned))
5692 0 : }
5693 :
5694 0 : async fn stop_reconciliations(&self, reason: StopReconciliationsReason) {
5695 0 : // Cancel all on-going reconciles and wait for them to exit the gate.
5696 0 : tracing::info!("{reason}: cancelling and waiting for in-flight reconciles");
5697 0 : self.reconcilers_cancel.cancel();
5698 0 : self.reconcilers_gate.close().await;
5699 :
5700 : // Signal the background loop in [`Service::process_results`] to exit once
5701 : // it has proccessed the results from all the reconciles we cancelled earlier.
5702 0 : tracing::info!("{reason}: processing results from previously in-flight reconciles");
5703 0 : self.result_tx.send(ReconcileResultRequest::Stop).ok();
5704 0 : self.result_tx.closed().await;
5705 0 : }
5706 :
5707 0 : pub async fn shutdown(&self) {
5708 0 : self.stop_reconciliations(StopReconciliationsReason::ShuttingDown)
5709 0 : .await;
5710 :
5711 : // Background tasks hold gate guards: this notifies them of the cancellation and
5712 : // waits for them all to complete.
5713 0 : tracing::info!("Shutting down: cancelling and waiting for background tasks to exit");
5714 0 : self.cancel.cancel();
5715 0 : self.gate.close().await;
5716 0 : }
5717 :
5718 : /// Drain a node by moving the shards attached to it as primaries.
5719 : /// This is a long running operation and it should run as a separate Tokio task.
5720 0 : pub(crate) async fn drain_node(
5721 0 : &self,
5722 0 : node_id: NodeId,
5723 0 : cancel: CancellationToken,
5724 0 : ) -> Result<(), OperationError> {
5725 0 : let mut last_inspected_shard: Option<TenantShardId> = None;
5726 0 : let mut inspected_all_shards = false;
5727 0 : let mut waiters = Vec::new();
5728 :
5729 0 : while !inspected_all_shards {
5730 0 : if cancel.is_cancelled() {
5731 0 : match self
5732 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
5733 0 : .await
5734 : {
5735 0 : Ok(()) => return Err(OperationError::Cancelled),
5736 0 : Err(err) => {
5737 0 : return Err(OperationError::FinalizeError(
5738 0 : format!(
5739 0 : "Failed to finalise drain cancel of {} by setting scheduling policy to Active: {}",
5740 0 : node_id, err
5741 0 : )
5742 0 : .into(),
5743 0 : ));
5744 : }
5745 : }
5746 0 : }
5747 0 :
5748 0 : {
5749 0 : let mut locked = self.inner.write().unwrap();
5750 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
5751 :
5752 0 : let node = nodes.get(&node_id).ok_or(OperationError::NodeStateChanged(
5753 0 : format!("node {node_id} was removed").into(),
5754 0 : ))?;
5755 :
5756 0 : let current_policy = node.get_scheduling();
5757 0 : if !matches!(current_policy, NodeSchedulingPolicy::Draining) {
5758 : // TODO(vlad): maybe cancel pending reconciles before erroring out. need to think
5759 : // about it
5760 0 : return Err(OperationError::NodeStateChanged(
5761 0 : format!("node {node_id} changed state to {current_policy:?}").into(),
5762 0 : ));
5763 0 : }
5764 0 :
5765 0 : let mut cursor = tenants.iter_mut().skip_while({
5766 0 : let skip_past = last_inspected_shard;
5767 0 : move |(tid, _)| match skip_past {
5768 0 : Some(last) => **tid != last,
5769 0 : None => false,
5770 0 : }
5771 0 : });
5772 :
5773 0 : while waiters.len() < MAX_RECONCILES_PER_OPERATION {
5774 0 : let (tid, tenant_shard) = match cursor.next() {
5775 0 : Some(some) => some,
5776 : None => {
5777 0 : inspected_all_shards = true;
5778 0 : break;
5779 : }
5780 : };
5781 :
5782 : // If the shard is not attached to the node being drained, skip it.
5783 0 : if *tenant_shard.intent.get_attached() != Some(node_id) {
5784 0 : last_inspected_shard = Some(*tid);
5785 0 : continue;
5786 0 : }
5787 0 :
5788 0 : match tenant_shard.reschedule_to_secondary(None, scheduler) {
5789 0 : Err(e) => {
5790 0 : tracing::warn!(
5791 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
5792 0 : "Scheduling error when draining pageserver {} : {e}", node_id
5793 : );
5794 : }
5795 : Ok(()) => {
5796 0 : let scheduled_to = tenant_shard.intent.get_attached();
5797 0 : tracing::info!(
5798 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
5799 0 : "Rescheduled shard while draining node {}: {} -> {:?}",
5800 : node_id,
5801 : node_id,
5802 : scheduled_to
5803 : );
5804 :
5805 0 : let waiter = self.maybe_reconcile_shard(tenant_shard, nodes);
5806 0 : if let Some(some) = waiter {
5807 0 : waiters.push(some);
5808 0 : }
5809 : }
5810 : }
5811 :
5812 0 : last_inspected_shard = Some(*tid);
5813 : }
5814 : }
5815 :
5816 0 : waiters = self
5817 0 : .await_waiters_remainder(waiters, SHORT_RECONCILE_TIMEOUT)
5818 0 : .await;
5819 :
5820 0 : failpoint_support::sleep_millis_async!("sleepy-drain-loop");
5821 : }
5822 :
5823 0 : while !waiters.is_empty() {
5824 0 : if cancel.is_cancelled() {
5825 0 : match self
5826 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
5827 0 : .await
5828 : {
5829 0 : Ok(()) => return Err(OperationError::Cancelled),
5830 0 : Err(err) => {
5831 0 : return Err(OperationError::FinalizeError(
5832 0 : format!(
5833 0 : "Failed to finalise drain cancel of {} by setting scheduling policy to Active: {}",
5834 0 : node_id, err
5835 0 : )
5836 0 : .into(),
5837 0 : ));
5838 : }
5839 : }
5840 0 : }
5841 0 :
5842 0 : tracing::info!("Awaiting {} pending drain reconciliations", waiters.len());
5843 :
5844 0 : waiters = self
5845 0 : .await_waiters_remainder(waiters, SHORT_RECONCILE_TIMEOUT)
5846 0 : .await;
5847 : }
5848 :
5849 : // At this point we have done the best we could to drain shards from this node.
5850 : // Set the node scheduling policy to `[NodeSchedulingPolicy::PauseForRestart]`
5851 : // to complete the drain.
5852 0 : if let Err(err) = self
5853 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::PauseForRestart))
5854 0 : .await
5855 : {
5856 : // This is not fatal. Anything that is polling the node scheduling policy to detect
5857 : // the end of the drain operations will hang, but all such places should enforce an
5858 : // overall timeout. The scheduling policy will be updated upon node re-attach and/or
5859 : // by the counterpart fill operation.
5860 0 : return Err(OperationError::FinalizeError(
5861 0 : format!(
5862 0 : "Failed to finalise drain of {node_id} by setting scheduling policy to PauseForRestart: {err}"
5863 0 : )
5864 0 : .into(),
5865 0 : ));
5866 0 : }
5867 0 :
5868 0 : Ok(())
5869 0 : }
5870 :
5871 : /// Create a node fill plan (pick secondaries to promote) that meets the following requirements:
5872 : /// 1. The node should be filled until it reaches the expected cluster average of
5873 : /// attached shards. If there are not enough secondaries on the node, the plan stops early.
5874 : /// 2. Select tenant shards to promote such that the number of attached shards is balanced
5875 : /// throughout the cluster. We achieve this by picking tenant shards from each node,
5876 : /// starting from the ones with the largest number of attached shards, until the node
5877 : /// reaches the expected cluster average.
5878 : /// 3. Avoid promoting more shards of the same tenant than required. The upper bound
5879 : /// for the number of tenants from the same shard promoted to the node being filled is:
5880 : /// shard count for the tenant divided by the number of nodes in the cluster.
5881 0 : fn fill_node_plan(&self, node_id: NodeId) -> Vec<TenantShardId> {
5882 0 : let mut locked = self.inner.write().unwrap();
5883 0 : let fill_requirement = locked.scheduler.compute_fill_requirement(node_id);
5884 0 :
5885 0 : let mut tids_by_node = locked
5886 0 : .tenants
5887 0 : .iter_mut()
5888 0 : .filter_map(|(tid, tenant_shard)| {
5889 0 : if tenant_shard.intent.get_secondary().contains(&node_id) {
5890 0 : if let Some(primary) = tenant_shard.intent.get_attached() {
5891 0 : return Some((*primary, *tid));
5892 0 : }
5893 0 : }
5894 :
5895 0 : None
5896 0 : })
5897 0 : .into_group_map();
5898 0 :
5899 0 : let expected_attached = locked.scheduler.expected_attached_shard_count();
5900 0 : let nodes_by_load = locked.scheduler.nodes_by_attached_shard_count();
5901 0 :
5902 0 : let mut promoted_per_tenant: HashMap<TenantId, usize> = HashMap::new();
5903 0 : let mut plan = Vec::new();
5904 :
5905 0 : for (node_id, attached) in nodes_by_load {
5906 0 : let available = locked
5907 0 : .nodes
5908 0 : .get(&node_id)
5909 0 : .map_or(false, |n| n.is_available());
5910 0 : if !available {
5911 0 : continue;
5912 0 : }
5913 0 :
5914 0 : if plan.len() >= fill_requirement
5915 0 : || tids_by_node.is_empty()
5916 0 : || attached <= expected_attached
5917 : {
5918 0 : break;
5919 0 : }
5920 0 :
5921 0 : let can_take = attached - expected_attached;
5922 0 : let needed = fill_requirement - plan.len();
5923 0 : let mut take = std::cmp::min(can_take, needed);
5924 0 :
5925 0 : let mut remove_node = false;
5926 0 : while take > 0 {
5927 0 : match tids_by_node.get_mut(&node_id) {
5928 0 : Some(tids) => match tids.pop() {
5929 0 : Some(tid) => {
5930 0 : let max_promote_for_tenant = std::cmp::max(
5931 0 : tid.shard_count.count() as usize / locked.nodes.len(),
5932 0 : 1,
5933 0 : );
5934 0 : let promoted = promoted_per_tenant.entry(tid.tenant_id).or_default();
5935 0 : if *promoted < max_promote_for_tenant {
5936 0 : plan.push(tid);
5937 0 : *promoted += 1;
5938 0 : take -= 1;
5939 0 : }
5940 : }
5941 : None => {
5942 0 : remove_node = true;
5943 0 : break;
5944 : }
5945 : },
5946 : None => {
5947 0 : break;
5948 : }
5949 : }
5950 : }
5951 :
5952 0 : if remove_node {
5953 0 : tids_by_node.remove(&node_id);
5954 0 : }
5955 : }
5956 :
5957 0 : plan
5958 0 : }
5959 :
5960 : /// Fill a node by promoting its secondaries until the cluster is balanced
5961 : /// with regards to attached shard counts. Note that this operation only
5962 : /// makes sense as a counterpart to the drain implemented in [`Service::drain_node`].
5963 : /// This is a long running operation and it should run as a separate Tokio task.
5964 0 : pub(crate) async fn fill_node(
5965 0 : &self,
5966 0 : node_id: NodeId,
5967 0 : cancel: CancellationToken,
5968 0 : ) -> Result<(), OperationError> {
5969 0 : // TODO(vlad): Currently this operates on the assumption that all
5970 0 : // secondaries are warm. This is not always true (e.g. we just migrated the
5971 0 : // tenant). Take that into consideration by checking the secondary status.
5972 0 : let mut tids_to_promote = self.fill_node_plan(node_id);
5973 0 : let mut waiters = Vec::new();
5974 :
5975 : // Execute the plan we've composed above. Before aplying each move from the plan,
5976 : // we validate to ensure that it has not gone stale in the meantime.
5977 0 : while !tids_to_promote.is_empty() {
5978 0 : if cancel.is_cancelled() {
5979 0 : match self
5980 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
5981 0 : .await
5982 : {
5983 0 : Ok(()) => return Err(OperationError::Cancelled),
5984 0 : Err(err) => {
5985 0 : return Err(OperationError::FinalizeError(
5986 0 : format!(
5987 0 : "Failed to finalise drain cancel of {} by setting scheduling policy to Active: {}",
5988 0 : node_id, err
5989 0 : )
5990 0 : .into(),
5991 0 : ));
5992 : }
5993 : }
5994 0 : }
5995 0 :
5996 0 : {
5997 0 : let mut locked = self.inner.write().unwrap();
5998 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
5999 :
6000 0 : let node = nodes.get(&node_id).ok_or(OperationError::NodeStateChanged(
6001 0 : format!("node {node_id} was removed").into(),
6002 0 : ))?;
6003 :
6004 0 : let current_policy = node.get_scheduling();
6005 0 : if !matches!(current_policy, NodeSchedulingPolicy::Filling) {
6006 : // TODO(vlad): maybe cancel pending reconciles before erroring out. need to think
6007 : // about it
6008 0 : return Err(OperationError::NodeStateChanged(
6009 0 : format!("node {node_id} changed state to {current_policy:?}").into(),
6010 0 : ));
6011 0 : }
6012 :
6013 0 : while waiters.len() < MAX_RECONCILES_PER_OPERATION {
6014 0 : if let Some(tid) = tids_to_promote.pop() {
6015 0 : if let Some(tenant_shard) = tenants.get_mut(&tid) {
6016 : // If the node being filled is not a secondary anymore,
6017 : // skip the promotion.
6018 0 : if !tenant_shard.intent.get_secondary().contains(&node_id) {
6019 0 : continue;
6020 0 : }
6021 0 :
6022 0 : let previously_attached_to = *tenant_shard.intent.get_attached();
6023 0 : match tenant_shard.reschedule_to_secondary(Some(node_id), scheduler) {
6024 0 : Err(e) => {
6025 0 : tracing::warn!(
6026 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
6027 0 : "Scheduling error when filling pageserver {} : {e}", node_id
6028 : );
6029 : }
6030 : Ok(()) => {
6031 0 : tracing::info!(
6032 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
6033 0 : "Rescheduled shard while filling node {}: {:?} -> {}",
6034 : node_id,
6035 : previously_attached_to,
6036 : node_id
6037 : );
6038 :
6039 0 : if let Some(waiter) =
6040 0 : self.maybe_reconcile_shard(tenant_shard, nodes)
6041 0 : {
6042 0 : waiters.push(waiter);
6043 0 : }
6044 : }
6045 : }
6046 0 : }
6047 : } else {
6048 0 : break;
6049 : }
6050 : }
6051 : }
6052 :
6053 0 : waiters = self
6054 0 : .await_waiters_remainder(waiters, SHORT_RECONCILE_TIMEOUT)
6055 0 : .await;
6056 : }
6057 :
6058 0 : while !waiters.is_empty() {
6059 0 : if cancel.is_cancelled() {
6060 0 : match self
6061 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
6062 0 : .await
6063 : {
6064 0 : Ok(()) => return Err(OperationError::Cancelled),
6065 0 : Err(err) => {
6066 0 : return Err(OperationError::FinalizeError(
6067 0 : format!(
6068 0 : "Failed to finalise drain cancel of {} by setting scheduling policy to Active: {}",
6069 0 : node_id, err
6070 0 : )
6071 0 : .into(),
6072 0 : ));
6073 : }
6074 : }
6075 0 : }
6076 0 :
6077 0 : tracing::info!("Awaiting {} pending fill reconciliations", waiters.len());
6078 :
6079 0 : waiters = self
6080 0 : .await_waiters_remainder(waiters, SHORT_RECONCILE_TIMEOUT)
6081 0 : .await;
6082 : }
6083 :
6084 0 : if let Err(err) = self
6085 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
6086 0 : .await
6087 : {
6088 : // This isn't a huge issue since the filling process starts upon request. However, it
6089 : // will prevent the next drain from starting. The only case in which this can fail
6090 : // is database unavailability. Such a case will require manual intervention.
6091 0 : return Err(OperationError::FinalizeError(
6092 0 : format!("Failed to finalise fill of {node_id} by setting scheduling policy to Active: {err}")
6093 0 : .into(),
6094 0 : ));
6095 0 : }
6096 0 :
6097 0 : Ok(())
6098 0 : }
6099 :
6100 : /// Updates scrubber metadata health check results.
6101 0 : pub(crate) async fn metadata_health_update(
6102 0 : &self,
6103 0 : update_req: MetadataHealthUpdateRequest,
6104 0 : ) -> Result<(), ApiError> {
6105 0 : let now = chrono::offset::Utc::now();
6106 0 : let (healthy_records, unhealthy_records) = {
6107 0 : let locked = self.inner.read().unwrap();
6108 0 : let healthy_records = update_req
6109 0 : .healthy_tenant_shards
6110 0 : .into_iter()
6111 0 : // Retain only health records associated with tenant shards managed by storage controller.
6112 0 : .filter(|tenant_shard_id| locked.tenants.contains_key(tenant_shard_id))
6113 0 : .map(|tenant_shard_id| MetadataHealthPersistence::new(tenant_shard_id, true, now))
6114 0 : .collect();
6115 0 : let unhealthy_records = update_req
6116 0 : .unhealthy_tenant_shards
6117 0 : .into_iter()
6118 0 : .filter(|tenant_shard_id| locked.tenants.contains_key(tenant_shard_id))
6119 0 : .map(|tenant_shard_id| MetadataHealthPersistence::new(tenant_shard_id, false, now))
6120 0 : .collect();
6121 0 :
6122 0 : (healthy_records, unhealthy_records)
6123 0 : };
6124 0 :
6125 0 : self.persistence
6126 0 : .update_metadata_health_records(healthy_records, unhealthy_records, now)
6127 0 : .await?;
6128 0 : Ok(())
6129 0 : }
6130 :
6131 : /// Lists the tenant shards that has unhealthy metadata status.
6132 0 : pub(crate) async fn metadata_health_list_unhealthy(
6133 0 : &self,
6134 0 : ) -> Result<Vec<TenantShardId>, ApiError> {
6135 0 : let result = self
6136 0 : .persistence
6137 0 : .list_unhealthy_metadata_health_records()
6138 0 : .await?
6139 0 : .iter()
6140 0 : .map(|p| p.get_tenant_shard_id().unwrap())
6141 0 : .collect();
6142 0 :
6143 0 : Ok(result)
6144 0 : }
6145 :
6146 : /// Lists the tenant shards that have not been scrubbed for some duration.
6147 0 : pub(crate) async fn metadata_health_list_outdated(
6148 0 : &self,
6149 0 : not_scrubbed_for: Duration,
6150 0 : ) -> Result<Vec<MetadataHealthRecord>, ApiError> {
6151 0 : let earlier = chrono::offset::Utc::now() - not_scrubbed_for;
6152 0 : let result = self
6153 0 : .persistence
6154 0 : .list_outdated_metadata_health_records(earlier)
6155 0 : .await?
6156 0 : .into_iter()
6157 0 : .map(|record| record.into())
6158 0 : .collect();
6159 0 : Ok(result)
6160 0 : }
6161 :
6162 0 : pub(crate) fn get_leadership_status(&self) -> LeadershipStatus {
6163 0 : self.inner.read().unwrap().get_leadership_status()
6164 0 : }
6165 :
6166 0 : pub(crate) async fn step_down(&self) -> GlobalObservedState {
6167 0 : tracing::info!("Received step down request from peer");
6168 :
6169 0 : self.inner.write().unwrap().step_down();
6170 0 : // TODO: would it make sense to have a time-out for this?
6171 0 : self.stop_reconciliations(StopReconciliationsReason::SteppingDown)
6172 0 : .await;
6173 :
6174 0 : let mut global_observed = GlobalObservedState::default();
6175 0 : let locked = self.inner.read().unwrap();
6176 0 : for (tid, tenant_shard) in locked.tenants.iter() {
6177 0 : global_observed
6178 0 : .0
6179 0 : .insert(*tid, tenant_shard.observed.clone());
6180 0 : }
6181 :
6182 0 : global_observed
6183 0 : }
6184 : }
|
