Line data Source code
1 : use hyper::Uri;
2 : use std::{
3 : borrow::Cow,
4 : cmp::Ordering,
5 : collections::{BTreeMap, HashMap, HashSet},
6 : ops::Deref,
7 : path::PathBuf,
8 : str::FromStr,
9 : sync::Arc,
10 : time::{Duration, Instant},
11 : };
12 :
13 : use crate::{
14 : background_node_operations::{
15 : Drain, Fill, Operation, OperationError, OperationHandler, MAX_RECONCILES_PER_OPERATION,
16 : },
17 : compute_hook::NotifyError,
18 : drain_utils::{self, TenantShardDrain, TenantShardIterator},
19 : id_lock_map::{trace_exclusive_lock, trace_shared_lock, IdLockMap, TracingExclusiveGuard},
20 : leadership::Leadership,
21 : metrics,
22 : peer_client::GlobalObservedState,
23 : persistence::{
24 : AbortShardSplitStatus, ControllerPersistence, DatabaseResult, MetadataHealthPersistence,
25 : ShardGenerationState, TenantFilter,
26 : },
27 : reconciler::{ReconcileError, ReconcileUnits, ReconcilerConfig, ReconcilerConfigBuilder},
28 : scheduler::{MaySchedule, ScheduleContext, ScheduleMode},
29 : tenant_shard::{
30 : MigrateAttachment, ReconcileNeeded, ReconcilerStatus, ScheduleOptimization,
31 : ScheduleOptimizationAction,
32 : },
33 : };
34 : use anyhow::Context;
35 : use control_plane::storage_controller::{
36 : AttachHookRequest, AttachHookResponse, InspectRequest, InspectResponse,
37 : };
38 : use diesel::result::DatabaseErrorKind;
39 : use futures::{stream::FuturesUnordered, StreamExt};
40 : use itertools::Itertools;
41 : use pageserver_api::{
42 : controller_api::{
43 : MetadataHealthRecord, MetadataHealthUpdateRequest, NodeAvailability, NodeRegisterRequest,
44 : NodeSchedulingPolicy, PlacementPolicy, ShardSchedulingPolicy, TenantCreateRequest,
45 : TenantCreateResponse, TenantCreateResponseShard, TenantDescribeResponse,
46 : TenantDescribeResponseShard, TenantLocateResponse, TenantPolicyRequest,
47 : TenantShardMigrateRequest, TenantShardMigrateResponse,
48 : },
49 : models::{
50 : SecondaryProgress, TenantConfigRequest, TimelineArchivalConfigRequest,
51 : TopTenantShardsRequest,
52 : },
53 : };
54 : use reqwest::StatusCode;
55 : use tracing::{instrument, Instrument};
56 :
57 : use crate::pageserver_client::PageserverClient;
58 : use pageserver_api::{
59 : models::{
60 : self, LocationConfig, LocationConfigListResponse, LocationConfigMode,
61 : PageserverUtilization, ShardParameters, TenantConfig, TenantLocationConfigRequest,
62 : TenantLocationConfigResponse, TenantShardLocation, TenantShardSplitRequest,
63 : TenantShardSplitResponse, TenantTimeTravelRequest, TimelineCreateRequest, TimelineInfo,
64 : },
65 : shard::{ShardCount, ShardIdentity, ShardNumber, ShardStripeSize, TenantShardId},
66 : upcall_api::{
67 : ReAttachRequest, ReAttachResponse, ReAttachResponseTenant, ValidateRequest,
68 : ValidateResponse, ValidateResponseTenant,
69 : },
70 : };
71 : use pageserver_client::mgmt_api;
72 : use tokio::sync::mpsc::error::TrySendError;
73 : use tokio_util::sync::CancellationToken;
74 : use utils::{
75 : completion::Barrier,
76 : failpoint_support,
77 : generation::Generation,
78 : http::error::ApiError,
79 : id::{NodeId, TenantId, TimelineId},
80 : sync::gate::Gate,
81 : };
82 :
83 : use crate::{
84 : compute_hook::ComputeHook,
85 : heartbeater::{Heartbeater, PageserverState},
86 : node::{AvailabilityTransition, Node},
87 : persistence::{split_state::SplitState, DatabaseError, Persistence, TenantShardPersistence},
88 : reconciler::attached_location_conf,
89 : scheduler::Scheduler,
90 : tenant_shard::{
91 : IntentState, ObservedState, ObservedStateLocation, ReconcileResult, ReconcileWaitError,
92 : ReconcilerWaiter, TenantShard,
93 : },
94 : };
95 :
96 : pub mod chaos_injector;
97 :
98 : // For operations that should be quick, like attaching a new tenant
99 : const SHORT_RECONCILE_TIMEOUT: Duration = Duration::from_secs(5);
100 :
101 : // For operations that might be slow, like migrating a tenant with
102 : // some data in it.
103 : pub const RECONCILE_TIMEOUT: Duration = Duration::from_secs(30);
104 :
105 : // If we receive a call using Secondary mode initially, it will omit generation. We will initialize
106 : // tenant shards into this generation, and as long as it remains in this generation, we will accept
107 : // input generation from future requests as authoritative.
108 : const INITIAL_GENERATION: Generation = Generation::new(0);
109 :
110 : /// How long [`Service::startup_reconcile`] is allowed to take before it should give
111 : /// up on unresponsive pageservers and proceed.
112 : pub(crate) const STARTUP_RECONCILE_TIMEOUT: Duration = Duration::from_secs(30);
113 :
114 : /// How long a node may be unresponsive to heartbeats before we declare it offline.
115 : /// This must be long enough to cover node restarts as well as normal operations: in future
116 : pub const MAX_OFFLINE_INTERVAL_DEFAULT: Duration = Duration::from_secs(30);
117 :
118 : /// How long a node may be unresponsive to heartbeats during start up before we declare it
119 : /// offline. This is much more lenient than [`MAX_OFFLINE_INTERVAL_DEFAULT`] since the pageserver's
120 : /// handling of the re-attach response may take a long time and blocks heartbeats from
121 : /// being handled on the pageserver side.
122 : pub const MAX_WARMING_UP_INTERVAL_DEFAULT: Duration = Duration::from_secs(300);
123 :
124 0 : #[derive(Clone, strum_macros::Display)]
125 : enum TenantOperations {
126 : Create,
127 : LocationConfig,
128 : ConfigSet,
129 : TimeTravelRemoteStorage,
130 : Delete,
131 : UpdatePolicy,
132 : ShardSplit,
133 : SecondaryDownload,
134 : TimelineCreate,
135 : TimelineDelete,
136 : AttachHook,
137 : TimelineArchivalConfig,
138 : TimelineDetachAncestor,
139 : }
140 :
141 0 : #[derive(Clone, strum_macros::Display)]
142 : enum NodeOperations {
143 : Register,
144 : Configure,
145 : Delete,
146 : }
147 :
148 : /// The leadership status for the storage controller process.
149 : /// Allowed transitions are:
150 : /// 1. Leader -> SteppedDown
151 : /// 2. Candidate -> Leader
152 : #[derive(
153 : Eq,
154 : PartialEq,
155 : Copy,
156 : Clone,
157 0 : strum_macros::Display,
158 0 : strum_macros::EnumIter,
159 : measured::FixedCardinalityLabel,
160 : )]
161 : #[strum(serialize_all = "snake_case")]
162 : pub(crate) enum LeadershipStatus {
163 : /// This is the steady state where the storage controller can produce
164 : /// side effects in the cluster.
165 : Leader,
166 : /// We've been notified to step down by another candidate. No reconciliations
167 : /// take place in this state.
168 : SteppedDown,
169 : /// Initial state for a new storage controller instance. Will attempt to assume leadership.
170 : #[allow(unused)]
171 : Candidate,
172 : }
173 :
174 : pub const RECONCILER_CONCURRENCY_DEFAULT: usize = 128;
175 :
176 : // Depth of the channel used to enqueue shards for reconciliation when they can't do it immediately.
177 : // This channel is finite-size to avoid using excessive memory if we get into a state where reconciles are finishing more slowly
178 : // than they're being pushed onto the queue.
179 : const MAX_DELAYED_RECONCILES: usize = 10000;
180 :
181 : // Top level state available to all HTTP handlers
182 : struct ServiceState {
183 : leadership_status: LeadershipStatus,
184 :
185 : tenants: BTreeMap<TenantShardId, TenantShard>,
186 :
187 : nodes: Arc<HashMap<NodeId, Node>>,
188 :
189 : scheduler: Scheduler,
190 :
191 : /// Ongoing background operation on the cluster if any is running.
192 : /// Note that only one such operation may run at any given time,
193 : /// hence the type choice.
194 : ongoing_operation: Option<OperationHandler>,
195 :
196 : /// Queue of tenants who are waiting for concurrency limits to permit them to reconcile
197 : delayed_reconcile_rx: tokio::sync::mpsc::Receiver<TenantShardId>,
198 : }
199 :
200 : /// Transform an error from a pageserver into an error to return to callers of a storage
201 : /// controller API.
202 0 : fn passthrough_api_error(node: &Node, e: mgmt_api::Error) -> ApiError {
203 0 : match e {
204 0 : mgmt_api::Error::SendRequest(e) => {
205 0 : // Presume errors sending requests are connectivity/availability issues
206 0 : ApiError::ResourceUnavailable(format!("{node} error sending request: {e}").into())
207 : }
208 0 : mgmt_api::Error::ReceiveErrorBody(str) => {
209 0 : // Presume errors receiving body are connectivity/availability issues
210 0 : ApiError::ResourceUnavailable(
211 0 : format!("{node} error receiving error body: {str}").into(),
212 0 : )
213 : }
214 0 : mgmt_api::Error::ReceiveBody(str) => {
215 0 : // Presume errors receiving body are connectivity/availability issues
216 0 : ApiError::ResourceUnavailable(format!("{node} error receiving body: {str}").into())
217 : }
218 0 : mgmt_api::Error::ApiError(StatusCode::NOT_FOUND, msg) => {
219 0 : ApiError::NotFound(anyhow::anyhow!(format!("{node}: {msg}")).into())
220 : }
221 0 : mgmt_api::Error::ApiError(StatusCode::SERVICE_UNAVAILABLE, msg) => {
222 0 : ApiError::ResourceUnavailable(format!("{node}: {msg}").into())
223 : }
224 0 : mgmt_api::Error::ApiError(status @ StatusCode::UNAUTHORIZED, msg)
225 0 : | mgmt_api::Error::ApiError(status @ StatusCode::FORBIDDEN, msg) => {
226 : // Auth errors talking to a pageserver are not auth errors for the caller: they are
227 : // internal server errors, showing that something is wrong with the pageserver or
228 : // storage controller's auth configuration.
229 0 : ApiError::InternalServerError(anyhow::anyhow!("{node} {status}: {msg}"))
230 : }
231 0 : mgmt_api::Error::ApiError(status, msg) => {
232 0 : // Presume general case of pageserver API errors is that we tried to do something
233 0 : // that can't be done right now.
234 0 : ApiError::Conflict(format!("{node} {status}: {status} {msg}"))
235 : }
236 0 : mgmt_api::Error::Cancelled => ApiError::ShuttingDown,
237 : }
238 0 : }
239 :
240 : impl ServiceState {
241 0 : fn new(
242 0 : nodes: HashMap<NodeId, Node>,
243 0 : tenants: BTreeMap<TenantShardId, TenantShard>,
244 0 : scheduler: Scheduler,
245 0 : delayed_reconcile_rx: tokio::sync::mpsc::Receiver<TenantShardId>,
246 0 : initial_leadership_status: LeadershipStatus,
247 0 : ) -> Self {
248 0 : metrics::update_leadership_status(initial_leadership_status);
249 0 :
250 0 : Self {
251 0 : leadership_status: initial_leadership_status,
252 0 : tenants,
253 0 : nodes: Arc::new(nodes),
254 0 : scheduler,
255 0 : ongoing_operation: None,
256 0 : delayed_reconcile_rx,
257 0 : }
258 0 : }
259 :
260 0 : fn parts_mut(
261 0 : &mut self,
262 0 : ) -> (
263 0 : &mut Arc<HashMap<NodeId, Node>>,
264 0 : &mut BTreeMap<TenantShardId, TenantShard>,
265 0 : &mut Scheduler,
266 0 : ) {
267 0 : (&mut self.nodes, &mut self.tenants, &mut self.scheduler)
268 0 : }
269 :
270 0 : fn get_leadership_status(&self) -> LeadershipStatus {
271 0 : self.leadership_status
272 0 : }
273 :
274 0 : fn step_down(&mut self) {
275 0 : self.leadership_status = LeadershipStatus::SteppedDown;
276 0 : metrics::update_leadership_status(self.leadership_status);
277 0 : }
278 :
279 0 : fn become_leader(&mut self) {
280 0 : self.leadership_status = LeadershipStatus::Leader;
281 0 : metrics::update_leadership_status(self.leadership_status);
282 0 : }
283 : }
284 :
285 : #[derive(Clone)]
286 : pub struct Config {
287 : // All pageservers managed by one instance of this service must have
288 : // the same public key. This JWT token will be used to authenticate
289 : // this service to the pageservers it manages.
290 : pub jwt_token: Option<String>,
291 :
292 : // This JWT token will be used to authenticate this service to the control plane.
293 : pub control_plane_jwt_token: Option<String>,
294 :
295 : // This JWT token will be used to authenticate with other storage controller instances
296 : pub peer_jwt_token: Option<String>,
297 :
298 : /// Where the compute hook should send notifications of pageserver attachment locations
299 : /// (this URL points to the control plane in prod). If this is None, the compute hook will
300 : /// assume it is running in a test environment and try to update neon_local.
301 : pub compute_hook_url: Option<String>,
302 :
303 : /// Grace period within which a pageserver does not respond to heartbeats, but is still
304 : /// considered active. Once the grace period elapses, the next heartbeat failure will
305 : /// mark the pagseserver offline.
306 : pub max_offline_interval: Duration,
307 :
308 : /// Extended grace period within which pageserver may not respond to heartbeats.
309 : /// This extended grace period kicks in after the node has been drained for restart
310 : /// and/or upon handling the re-attach request from a node.
311 : pub max_warming_up_interval: Duration,
312 :
313 : /// How many Reconcilers may be spawned concurrently
314 : pub reconciler_concurrency: usize,
315 :
316 : /// How large must a shard grow in bytes before we split it?
317 : /// None disables auto-splitting.
318 : pub split_threshold: Option<u64>,
319 :
320 : // TODO: make this cfg(feature = "testing")
321 : pub neon_local_repo_dir: Option<PathBuf>,
322 :
323 : // Maximum acceptable download lag for the secondary location
324 : // while draining a node. If the secondary location is lagging
325 : // by more than the configured amount, then the secondary is not
326 : // upgraded to primary.
327 : pub max_secondary_lag_bytes: Option<u64>,
328 :
329 : pub address_for_peers: Option<Uri>,
330 :
331 : pub start_as_candidate: bool,
332 :
333 : pub http_service_port: i32,
334 : }
335 :
336 : impl From<DatabaseError> for ApiError {
337 0 : fn from(err: DatabaseError) -> ApiError {
338 0 : match err {
339 0 : DatabaseError::Query(e) => ApiError::InternalServerError(e.into()),
340 : // FIXME: ApiError doesn't have an Unavailable variant, but ShuttingDown maps to 503.
341 : DatabaseError::Connection(_) | DatabaseError::ConnectionPool(_) => {
342 0 : ApiError::ShuttingDown
343 : }
344 0 : DatabaseError::Logical(reason) | DatabaseError::Migration(reason) => {
345 0 : ApiError::InternalServerError(anyhow::anyhow!(reason))
346 : }
347 : }
348 0 : }
349 : }
350 :
351 : pub struct Service {
352 : inner: Arc<std::sync::RwLock<ServiceState>>,
353 : config: Config,
354 : persistence: Arc<Persistence>,
355 : compute_hook: Arc<ComputeHook>,
356 : result_tx: tokio::sync::mpsc::UnboundedSender<ReconcileResultRequest>,
357 :
358 : heartbeater: Heartbeater,
359 :
360 : // Channel for background cleanup from failed operations that require cleanup, such as shard split
361 : abort_tx: tokio::sync::mpsc::UnboundedSender<TenantShardSplitAbort>,
362 :
363 : // Locking on a tenant granularity (covers all shards in the tenant):
364 : // - Take exclusively for rare operations that mutate the tenant's persistent state (e.g. create/delete/split)
365 : // - Take in shared mode for operations that need the set of shards to stay the same to complete reliably (e.g. timeline CRUD)
366 : tenant_op_locks: IdLockMap<TenantId, TenantOperations>,
367 :
368 : // Locking for node-mutating operations: take exclusively for operations that modify the node's persistent state, or
369 : // that transition it to/from Active.
370 : node_op_locks: IdLockMap<NodeId, NodeOperations>,
371 :
372 : // Limit how many Reconcilers we will spawn concurrently
373 : reconciler_concurrency: Arc<tokio::sync::Semaphore>,
374 :
375 : /// Queue of tenants who are waiting for concurrency limits to permit them to reconcile
376 : /// Send into this queue to promptly attempt to reconcile this shard next time units are available.
377 : ///
378 : /// Note that this state logically lives inside ServiceInner, but carrying Sender here makes the code simpler
379 : /// by avoiding needing a &mut ref to something inside the ServiceInner. This could be optimized to
380 : /// use a VecDeque instead of a channel to reduce synchronization overhead, at the cost of some code complexity.
381 : delayed_reconcile_tx: tokio::sync::mpsc::Sender<TenantShardId>,
382 :
383 : // Process shutdown will fire this token
384 : cancel: CancellationToken,
385 :
386 : // Child token of [`Service::cancel`] used by reconcilers
387 : reconcilers_cancel: CancellationToken,
388 :
389 : // Background tasks will hold this gate
390 : gate: Gate,
391 :
392 : // Reconcilers background tasks will hold this gate
393 : reconcilers_gate: Gate,
394 :
395 : /// This waits for initial reconciliation with pageservers to complete. Until this barrier
396 : /// passes, it isn't safe to do any actions that mutate tenants.
397 : pub(crate) startup_complete: Barrier,
398 : }
399 :
400 : impl From<ReconcileWaitError> for ApiError {
401 0 : fn from(value: ReconcileWaitError) -> Self {
402 0 : match value {
403 0 : ReconcileWaitError::Shutdown => ApiError::ShuttingDown,
404 0 : e @ ReconcileWaitError::Timeout(_) => ApiError::Timeout(format!("{e}").into()),
405 0 : e @ ReconcileWaitError::Failed(..) => ApiError::InternalServerError(anyhow::anyhow!(e)),
406 : }
407 0 : }
408 : }
409 :
410 : impl From<OperationError> for ApiError {
411 0 : fn from(value: OperationError) -> Self {
412 0 : match value {
413 0 : OperationError::NodeStateChanged(err) | OperationError::FinalizeError(err) => {
414 0 : ApiError::InternalServerError(anyhow::anyhow!(err))
415 : }
416 0 : OperationError::Cancelled => ApiError::Conflict("Operation was cancelled".into()),
417 : }
418 0 : }
419 : }
420 :
421 : #[allow(clippy::large_enum_variant)]
422 : enum TenantCreateOrUpdate {
423 : Create(TenantCreateRequest),
424 : Update(Vec<ShardUpdate>),
425 : }
426 :
427 : struct ShardSplitParams {
428 : old_shard_count: ShardCount,
429 : new_shard_count: ShardCount,
430 : new_stripe_size: Option<ShardStripeSize>,
431 : targets: Vec<ShardSplitTarget>,
432 : policy: PlacementPolicy,
433 : config: TenantConfig,
434 : shard_ident: ShardIdentity,
435 : }
436 :
437 : // When preparing for a shard split, we may either choose to proceed with the split,
438 : // or find that the work is already done and return NoOp.
439 : enum ShardSplitAction {
440 : Split(ShardSplitParams),
441 : NoOp(TenantShardSplitResponse),
442 : }
443 :
444 : // A parent shard which will be split
445 : struct ShardSplitTarget {
446 : parent_id: TenantShardId,
447 : node: Node,
448 : child_ids: Vec<TenantShardId>,
449 : }
450 :
451 : /// When we tenant shard split operation fails, we may not be able to clean up immediately, because nodes
452 : /// might not be available. We therefore use a queue of abort operations processed in the background.
453 : struct TenantShardSplitAbort {
454 : tenant_id: TenantId,
455 : /// The target values from the request that failed
456 : new_shard_count: ShardCount,
457 : new_stripe_size: Option<ShardStripeSize>,
458 : /// Until this abort op is complete, no other operations may be done on the tenant
459 : _tenant_lock: TracingExclusiveGuard<TenantOperations>,
460 : }
461 :
462 0 : #[derive(thiserror::Error, Debug)]
463 : enum TenantShardSplitAbortError {
464 : #[error(transparent)]
465 : Database(#[from] DatabaseError),
466 : #[error(transparent)]
467 : Remote(#[from] mgmt_api::Error),
468 : #[error("Unavailable")]
469 : Unavailable,
470 : }
471 :
472 : struct ShardUpdate {
473 : tenant_shard_id: TenantShardId,
474 : placement_policy: PlacementPolicy,
475 : tenant_config: TenantConfig,
476 :
477 : /// If this is None, generation is not updated.
478 : generation: Option<Generation>,
479 : }
480 :
481 : enum StopReconciliationsReason {
482 : ShuttingDown,
483 : SteppingDown,
484 : }
485 :
486 : impl std::fmt::Display for StopReconciliationsReason {
487 0 : fn fmt(&self, writer: &mut std::fmt::Formatter) -> std::fmt::Result {
488 0 : let s = match self {
489 0 : Self::ShuttingDown => "Shutting down",
490 0 : Self::SteppingDown => "Stepping down",
491 : };
492 0 : write!(writer, "{}", s)
493 0 : }
494 : }
495 :
496 : pub(crate) enum ReconcileResultRequest {
497 : ReconcileResult(ReconcileResult),
498 : Stop,
499 : }
500 :
501 : impl Service {
502 0 : pub fn get_config(&self) -> &Config {
503 0 : &self.config
504 0 : }
505 :
506 : /// Called once on startup, this function attempts to contact all pageservers to build an up-to-date
507 : /// view of the world, and determine which pageservers are responsive.
508 0 : #[instrument(skip_all)]
509 : async fn startup_reconcile(
510 : self: &Arc<Service>,
511 : current_leader: Option<ControllerPersistence>,
512 : leader_step_down_state: Option<GlobalObservedState>,
513 : bg_compute_notify_result_tx: tokio::sync::mpsc::Sender<
514 : Result<(), (TenantShardId, NotifyError)>,
515 : >,
516 : ) {
517 : // Startup reconciliation does I/O to other services: whether they
518 : // are responsive or not, we should aim to finish within our deadline, because:
519 : // - If we don't, a k8s readiness hook watching /ready will kill us.
520 : // - While we're waiting for startup reconciliation, we are not fully
521 : // available for end user operations like creating/deleting tenants and timelines.
522 : //
523 : // We set multiple deadlines to break up the time available between the phases of work: this is
524 : // arbitrary, but avoids a situation where the first phase could burn our entire timeout period.
525 : let start_at = Instant::now();
526 : let node_scan_deadline = start_at
527 : .checked_add(STARTUP_RECONCILE_TIMEOUT / 2)
528 : .expect("Reconcile timeout is a modest constant");
529 :
530 : let observed = if let Some(state) = leader_step_down_state {
531 : tracing::info!(
532 : "Using observed state received from leader at {}",
533 : current_leader.as_ref().unwrap().address
534 : );
535 :
536 : state
537 : } else {
538 : self.build_global_observed_state(node_scan_deadline).await
539 : };
540 :
541 : // Accumulate a list of any tenant locations that ought to be detached
542 : let mut cleanup = Vec::new();
543 :
544 : // Send initial heartbeat requests to all nodes loaded from the database
545 : let all_nodes = {
546 : let locked = self.inner.read().unwrap();
547 : locked.nodes.clone()
548 : };
549 : let mut nodes_online = self.initial_heartbeat_round(all_nodes.keys()).await;
550 :
551 : // List of tenants for which we will attempt to notify compute of their location at startup
552 : let mut compute_notifications = Vec::new();
553 :
554 : // Populate intent and observed states for all tenants, based on reported state on pageservers
555 : tracing::info!("Populating tenant shards' states from initial pageserver scan...");
556 : let shard_count = {
557 : let mut locked = self.inner.write().unwrap();
558 : let (nodes, tenants, scheduler) = locked.parts_mut();
559 :
560 : // Mark nodes online if they responded to us: nodes are offline by default after a restart.
561 : let mut new_nodes = (**nodes).clone();
562 : for (node_id, node) in new_nodes.iter_mut() {
563 : if let Some(utilization) = nodes_online.remove(node_id) {
564 : node.set_availability(NodeAvailability::Active(utilization));
565 : scheduler.node_upsert(node);
566 : }
567 : }
568 : *nodes = Arc::new(new_nodes);
569 :
570 : for (tenant_shard_id, observed_state) in observed.0 {
571 : let Some(tenant_shard) = tenants.get_mut(&tenant_shard_id) else {
572 : for node_id in observed_state.locations.keys() {
573 : cleanup.push((tenant_shard_id, *node_id));
574 : }
575 :
576 : continue;
577 : };
578 :
579 : tenant_shard.observed = observed_state;
580 : }
581 :
582 : // Populate each tenant's intent state
583 : let mut schedule_context = ScheduleContext::default();
584 : for (tenant_shard_id, tenant_shard) in tenants.iter_mut() {
585 : if tenant_shard_id.shard_number == ShardNumber(0) {
586 : // Reset scheduling context each time we advance to the next Tenant
587 : schedule_context = ScheduleContext::default();
588 : }
589 :
590 : tenant_shard.intent_from_observed(scheduler);
591 : if let Err(e) = tenant_shard.schedule(scheduler, &mut schedule_context) {
592 : // Non-fatal error: we are unable to properly schedule the tenant, perhaps because
593 : // not enough pageservers are available. The tenant may well still be available
594 : // to clients.
595 : tracing::error!("Failed to schedule tenant {tenant_shard_id} at startup: {e}");
596 : } else {
597 : // If we're both intending and observed to be attached at a particular node, we will
598 : // emit a compute notification for this. In the case where our observed state does not
599 : // yet match our intent, we will eventually reconcile, and that will emit a compute notification.
600 : if let Some(attached_at) = tenant_shard.stably_attached() {
601 : compute_notifications.push((
602 : *tenant_shard_id,
603 : attached_at,
604 : tenant_shard.shard.stripe_size,
605 : ));
606 : }
607 : }
608 : }
609 :
610 : tenants.len()
611 : };
612 :
613 : // Before making any obeservable changes to the cluster, persist self
614 : // as leader in database and memory.
615 : let leadership = Leadership::new(
616 : self.persistence.clone(),
617 : self.config.clone(),
618 : self.cancel.child_token(),
619 : );
620 :
621 : if let Err(e) = leadership.become_leader(current_leader).await {
622 : tracing::error!("Failed to persist self as leader: {e}. Aborting start-up ...");
623 : std::process::exit(1);
624 : }
625 :
626 : self.inner.write().unwrap().become_leader();
627 :
628 : // TODO: if any tenant's intent now differs from its loaded generation_pageserver, we should clear that
629 : // generation_pageserver in the database.
630 :
631 : // Emit compute hook notifications for all tenants which are already stably attached. Other tenants
632 : // will emit compute hook notifications when they reconcile.
633 : //
634 : // Ordering: our calls to notify_background synchronously establish a relative order for these notifications vs. any later
635 : // calls into the ComputeHook for the same tenant: we can leave these to run to completion in the background and any later
636 : // calls will be correctly ordered wrt these.
637 : //
638 : // Concurrency: we call notify_background for all tenants, which will create O(N) tokio tasks, but almost all of them
639 : // will just wait on the ComputeHook::API_CONCURRENCY semaphore immediately, so very cheap until they get that semaphore
640 : // unit and start doing I/O.
641 : tracing::info!(
642 : "Sending {} compute notifications",
643 : compute_notifications.len()
644 : );
645 : self.compute_hook.notify_background(
646 : compute_notifications,
647 : bg_compute_notify_result_tx.clone(),
648 : &self.cancel,
649 : );
650 :
651 : // Finally, now that the service is up and running, launch reconcile operations for any tenants
652 : // which require it: under normal circumstances this should only include tenants that were in some
653 : // transient state before we restarted, or any tenants whose compute hooks failed above.
654 : tracing::info!("Checking for shards in need of reconciliation...");
655 : let reconcile_tasks = self.reconcile_all();
656 : // We will not wait for these reconciliation tasks to run here: we're now done with startup and
657 : // normal operations may proceed.
658 :
659 : // Clean up any tenants that were found on pageservers but are not known to us. Do this in the
660 : // background because it does not need to complete in order to proceed with other work.
661 : if !cleanup.is_empty() {
662 : tracing::info!("Cleaning up {} locations in the background", cleanup.len());
663 : tokio::task::spawn({
664 : let cleanup_self = self.clone();
665 0 : async move { cleanup_self.cleanup_locations(cleanup).await }
666 : });
667 : }
668 :
669 : tracing::info!("Startup complete, spawned {reconcile_tasks} reconciliation tasks ({shard_count} shards total)");
670 : }
671 :
672 0 : async fn initial_heartbeat_round<'a>(
673 0 : &self,
674 0 : node_ids: impl Iterator<Item = &'a NodeId>,
675 0 : ) -> HashMap<NodeId, PageserverUtilization> {
676 0 : assert!(!self.startup_complete.is_ready());
677 :
678 0 : let all_nodes = {
679 0 : let locked = self.inner.read().unwrap();
680 0 : locked.nodes.clone()
681 0 : };
682 0 :
683 0 : let mut nodes_to_heartbeat = HashMap::new();
684 0 : for node_id in node_ids {
685 0 : match all_nodes.get(node_id) {
686 0 : Some(node) => {
687 0 : nodes_to_heartbeat.insert(*node_id, node.clone());
688 0 : }
689 : None => {
690 0 : tracing::warn!("Node {node_id} was removed during start-up");
691 : }
692 : }
693 : }
694 :
695 0 : tracing::info!("Sending initial heartbeats...");
696 0 : let res = self
697 0 : .heartbeater
698 0 : .heartbeat(Arc::new(nodes_to_heartbeat))
699 0 : .await;
700 :
701 0 : let mut online_nodes = HashMap::new();
702 0 : if let Ok(deltas) = res {
703 0 : for (node_id, status) in deltas.0 {
704 0 : match status {
705 0 : PageserverState::Available { utilization, .. } => {
706 0 : online_nodes.insert(node_id, utilization);
707 0 : }
708 0 : PageserverState::Offline => {}
709 : PageserverState::WarmingUp { .. } => {
710 0 : unreachable!("Nodes are never marked warming-up during startup reconcile")
711 : }
712 : }
713 : }
714 0 : }
715 :
716 0 : online_nodes
717 0 : }
718 :
719 : /// Used during [`Self::startup_reconcile`]: issue GETs to all nodes concurrently, with a deadline.
720 : ///
721 : /// The result includes only nodes which responded within the deadline
722 0 : async fn scan_node_locations(
723 0 : &self,
724 0 : deadline: Instant,
725 0 : ) -> HashMap<NodeId, LocationConfigListResponse> {
726 0 : let nodes = {
727 0 : let locked = self.inner.read().unwrap();
728 0 : locked.nodes.clone()
729 0 : };
730 0 :
731 0 : let mut node_results = HashMap::new();
732 0 :
733 0 : let mut node_list_futs = FuturesUnordered::new();
734 0 :
735 0 : tracing::info!("Scanning shards on {} nodes...", nodes.len());
736 0 : for node in nodes.values() {
737 0 : node_list_futs.push({
738 0 : async move {
739 0 : tracing::info!("Scanning shards on node {node}...");
740 0 : let timeout = Duration::from_secs(1);
741 0 : let response = node
742 0 : .with_client_retries(
743 0 : |client| async move { client.list_location_config().await },
744 0 : &self.config.jwt_token,
745 0 : 1,
746 0 : 5,
747 0 : timeout,
748 0 : &self.cancel,
749 0 : )
750 0 : .await;
751 0 : (node.get_id(), response)
752 0 : }
753 0 : });
754 0 : }
755 :
756 : loop {
757 0 : let (node_id, result) = tokio::select! {
758 : next = node_list_futs.next() => {
759 : match next {
760 : Some(result) => result,
761 : None =>{
762 : // We got results for all our nodes
763 : break;
764 : }
765 :
766 : }
767 : },
768 : _ = tokio::time::sleep(deadline.duration_since(Instant::now())) => {
769 : // Give up waiting for anyone who hasn't responded: we will yield the results that we have
770 : tracing::info!("Reached deadline while waiting for nodes to respond to location listing requests");
771 : break;
772 : }
773 : };
774 :
775 0 : let Some(list_response) = result else {
776 0 : tracing::info!("Shutdown during startup_reconcile");
777 0 : break;
778 : };
779 :
780 0 : match list_response {
781 0 : Err(e) => {
782 0 : tracing::warn!("Could not scan node {} ({e})", node_id);
783 : }
784 0 : Ok(listing) => {
785 0 : node_results.insert(node_id, listing);
786 0 : }
787 : }
788 : }
789 :
790 0 : node_results
791 0 : }
792 :
793 0 : async fn build_global_observed_state(&self, deadline: Instant) -> GlobalObservedState {
794 0 : let node_listings = self.scan_node_locations(deadline).await;
795 0 : let mut observed = GlobalObservedState::default();
796 :
797 0 : for (node_id, location_confs) in node_listings {
798 0 : tracing::info!(
799 0 : "Received {} shard statuses from pageserver {}",
800 0 : location_confs.tenant_shards.len(),
801 : node_id
802 : );
803 :
804 0 : for (tid, location_conf) in location_confs.tenant_shards {
805 0 : let entry = observed.0.entry(tid).or_default();
806 0 : entry.locations.insert(
807 0 : node_id,
808 0 : ObservedStateLocation {
809 0 : conf: location_conf,
810 0 : },
811 0 : );
812 0 : }
813 : }
814 :
815 0 : observed
816 0 : }
817 :
818 : /// Used during [`Self::startup_reconcile`]: detach a list of unknown-to-us tenants from pageservers.
819 : ///
820 : /// This is safe to run in the background, because if we don't have this TenantShardId in our map of
821 : /// tenants, then it is probably something incompletely deleted before: we will not fight with any
822 : /// other task trying to attach it.
823 0 : #[instrument(skip_all)]
824 : async fn cleanup_locations(&self, cleanup: Vec<(TenantShardId, NodeId)>) {
825 : let nodes = self.inner.read().unwrap().nodes.clone();
826 :
827 : for (tenant_shard_id, node_id) in cleanup {
828 : // A node reported a tenant_shard_id which is unknown to us: detach it.
829 : let Some(node) = nodes.get(&node_id) else {
830 : // This is legitimate; we run in the background and [`Self::startup_reconcile`] might have identified
831 : // a location to clean up on a node that has since been removed.
832 : tracing::info!(
833 : "Not cleaning up location {node_id}/{tenant_shard_id}: node not found"
834 : );
835 : continue;
836 : };
837 :
838 : if self.cancel.is_cancelled() {
839 : break;
840 : }
841 :
842 : let client = PageserverClient::new(
843 : node.get_id(),
844 : node.base_url(),
845 : self.config.jwt_token.as_deref(),
846 : );
847 : match client
848 : .location_config(
849 : tenant_shard_id,
850 : LocationConfig {
851 : mode: LocationConfigMode::Detached,
852 : generation: None,
853 : secondary_conf: None,
854 : shard_number: tenant_shard_id.shard_number.0,
855 : shard_count: tenant_shard_id.shard_count.literal(),
856 : shard_stripe_size: 0,
857 : tenant_conf: models::TenantConfig::default(),
858 : },
859 : None,
860 : false,
861 : )
862 : .await
863 : {
864 : Ok(()) => {
865 : tracing::info!(
866 : "Detached unknown shard {tenant_shard_id} on pageserver {node_id}"
867 : );
868 : }
869 : Err(e) => {
870 : // Non-fatal error: leaving a tenant shard behind that we are not managing shouldn't
871 : // break anything.
872 : tracing::error!(
873 : "Failed to detach unknkown shard {tenant_shard_id} on pageserver {node_id}: {e}"
874 : );
875 : }
876 : }
877 : }
878 : }
879 :
880 : /// Long running background task that periodically wakes up and looks for shards that need
881 : /// reconciliation. Reconciliation is fallible, so any reconciliation tasks that fail during
882 : /// e.g. a tenant create/attach/migrate must eventually be retried: this task is responsible
883 : /// for those retries.
884 0 : #[instrument(skip_all)]
885 : async fn background_reconcile(self: &Arc<Self>) {
886 : self.startup_complete.clone().wait().await;
887 :
888 : const BACKGROUND_RECONCILE_PERIOD: Duration = Duration::from_secs(20);
889 :
890 : let mut interval = tokio::time::interval(BACKGROUND_RECONCILE_PERIOD);
891 : while !self.reconcilers_cancel.is_cancelled() {
892 : tokio::select! {
893 : _ = interval.tick() => {
894 : let reconciles_spawned = self.reconcile_all();
895 : if reconciles_spawned == 0 {
896 : // Run optimizer only when we didn't find any other work to do
897 : let optimizations = self.optimize_all().await;
898 : if optimizations == 0 {
899 : // Run new splits only when no optimizations are pending
900 : self.autosplit_tenants().await;
901 : }
902 : }
903 : }
904 : _ = self.reconcilers_cancel.cancelled() => return
905 : }
906 : }
907 : }
908 0 : #[instrument(skip_all)]
909 : async fn spawn_heartbeat_driver(&self) {
910 : self.startup_complete.clone().wait().await;
911 :
912 : const HEARTBEAT_INTERVAL: Duration = Duration::from_secs(5);
913 :
914 : let mut interval = tokio::time::interval(HEARTBEAT_INTERVAL);
915 : while !self.cancel.is_cancelled() {
916 : tokio::select! {
917 : _ = interval.tick() => { }
918 : _ = self.cancel.cancelled() => return
919 : };
920 :
921 : let nodes = {
922 : let locked = self.inner.read().unwrap();
923 : locked.nodes.clone()
924 : };
925 :
926 : let res = self.heartbeater.heartbeat(nodes).await;
927 : if let Ok(deltas) = res {
928 : for (node_id, state) in deltas.0 {
929 : let new_availability = match state {
930 : PageserverState::Available { utilization, .. } => {
931 : NodeAvailability::Active(utilization)
932 : }
933 : PageserverState::WarmingUp { started_at } => {
934 : NodeAvailability::WarmingUp(started_at)
935 : }
936 : PageserverState::Offline => {
937 : // The node might have been placed in the WarmingUp state
938 : // while the heartbeat round was on-going. Hence, filter out
939 : // offline transitions for WarmingUp nodes that are still within
940 : // their grace period.
941 : if let Ok(NodeAvailability::WarmingUp(started_at)) = self
942 : .get_node(node_id)
943 : .await
944 : .as_ref()
945 0 : .map(|n| n.get_availability())
946 : {
947 : let now = Instant::now();
948 : if now - *started_at >= self.config.max_warming_up_interval {
949 : NodeAvailability::Offline
950 : } else {
951 : NodeAvailability::WarmingUp(*started_at)
952 : }
953 : } else {
954 : NodeAvailability::Offline
955 : }
956 : }
957 : };
958 :
959 : // This is the code path for geniune availability transitions (i.e node
960 : // goes unavailable and/or comes back online).
961 : let res = self
962 : .node_configure(node_id, Some(new_availability), None)
963 : .await;
964 :
965 : match res {
966 : Ok(()) => {}
967 : Err(ApiError::NotFound(_)) => {
968 : // This should be rare, but legitimate since the heartbeats are done
969 : // on a snapshot of the nodes.
970 : tracing::info!("Node {} was not found after heartbeat round", node_id);
971 : }
972 : Err(err) => {
973 : // Transition to active involves reconciling: if a node responds to a heartbeat then
974 : // becomes unavailable again, we may get an error here.
975 : tracing::error!(
976 : "Failed to update node {} after heartbeat round: {}",
977 : node_id,
978 : err
979 : );
980 : }
981 : }
982 : }
983 : }
984 : }
985 : }
986 :
987 : /// Apply the contents of a [`ReconcileResult`] to our in-memory state: if the reconciliation
988 : /// was successful and intent hasn't changed since the Reconciler was spawned, this will update
989 : /// the observed state of the tenant such that subsequent calls to [`TenantShard::get_reconcile_needed`]
990 : /// will indicate that reconciliation is not needed.
991 0 : #[instrument(skip_all, fields(
992 : tenant_id=%result.tenant_shard_id.tenant_id, shard_id=%result.tenant_shard_id.shard_slug(),
993 : sequence=%result.sequence
994 0 : ))]
995 : fn process_result(&self, mut result: ReconcileResult) {
996 : let mut locked = self.inner.write().unwrap();
997 : let (nodes, tenants, _scheduler) = locked.parts_mut();
998 : let Some(tenant) = tenants.get_mut(&result.tenant_shard_id) else {
999 : // A reconciliation result might race with removing a tenant: drop results for
1000 : // tenants that aren't in our map.
1001 : return;
1002 : };
1003 :
1004 : // Usually generation should only be updated via this path, so the max() isn't
1005 : // needed, but it is used to handle out-of-band updates via. e.g. test hook.
1006 : tenant.generation = std::cmp::max(tenant.generation, result.generation);
1007 :
1008 : // If the reconciler signals that it failed to notify compute, set this state on
1009 : // the shard so that a future [`TenantShard::maybe_reconcile`] will try again.
1010 : tenant.pending_compute_notification = result.pending_compute_notification;
1011 :
1012 : // Let the TenantShard know it is idle.
1013 : tenant.reconcile_complete(result.sequence);
1014 :
1015 : // In case a node was deleted while this reconcile is in flight, filter it out of the update we will
1016 : // make to the tenant
1017 : result
1018 : .observed
1019 : .locations
1020 0 : .retain(|node_id, _loc| nodes.contains_key(node_id));
1021 :
1022 : match result.result {
1023 : Ok(()) => {
1024 : for (node_id, loc) in &result.observed.locations {
1025 : if let Some(conf) = &loc.conf {
1026 : tracing::info!("Updating observed location {}: {:?}", node_id, conf);
1027 : } else {
1028 : tracing::info!("Setting observed location {} to None", node_id,)
1029 : }
1030 : }
1031 :
1032 : tenant.observed = result.observed;
1033 : tenant.waiter.advance(result.sequence);
1034 : }
1035 : Err(e) => {
1036 : match e {
1037 : ReconcileError::Cancel => {
1038 : tracing::info!("Reconciler was cancelled");
1039 : }
1040 : ReconcileError::Remote(mgmt_api::Error::Cancelled) => {
1041 : // This might be due to the reconciler getting cancelled, or it might
1042 : // be due to the `Node` being marked offline.
1043 : tracing::info!("Reconciler cancelled during pageserver API call");
1044 : }
1045 : _ => {
1046 : tracing::warn!("Reconcile error: {}", e);
1047 : }
1048 : }
1049 :
1050 : // Ordering: populate last_error before advancing error_seq,
1051 : // so that waiters will see the correct error after waiting.
1052 : tenant.set_last_error(result.sequence, e);
1053 :
1054 : for (node_id, o) in result.observed.locations {
1055 : tenant.observed.locations.insert(node_id, o);
1056 : }
1057 : }
1058 : }
1059 :
1060 : // Maybe some other work can proceed now that this job finished.
1061 : if self.reconciler_concurrency.available_permits() > 0 {
1062 : while let Ok(tenant_shard_id) = locked.delayed_reconcile_rx.try_recv() {
1063 : let (nodes, tenants, _scheduler) = locked.parts_mut();
1064 : if let Some(shard) = tenants.get_mut(&tenant_shard_id) {
1065 : shard.delayed_reconcile = false;
1066 : self.maybe_reconcile_shard(shard, nodes);
1067 : }
1068 :
1069 : if self.reconciler_concurrency.available_permits() == 0 {
1070 : break;
1071 : }
1072 : }
1073 : }
1074 : }
1075 :
1076 0 : async fn process_results(
1077 0 : &self,
1078 0 : mut result_rx: tokio::sync::mpsc::UnboundedReceiver<ReconcileResultRequest>,
1079 0 : mut bg_compute_hook_result_rx: tokio::sync::mpsc::Receiver<
1080 0 : Result<(), (TenantShardId, NotifyError)>,
1081 0 : >,
1082 0 : ) {
1083 0 : loop {
1084 0 : // Wait for the next result, or for cancellation
1085 0 : tokio::select! {
1086 : r = result_rx.recv() => {
1087 : match r {
1088 : Some(ReconcileResultRequest::ReconcileResult(result)) => {self.process_result(result);},
1089 : None | Some(ReconcileResultRequest::Stop) => {break;}
1090 : }
1091 : }
1092 0 : _ = async{
1093 0 : match bg_compute_hook_result_rx.recv().await {
1094 0 : Some(result) => {
1095 0 : if let Err((tenant_shard_id, notify_error)) = result {
1096 0 : tracing::warn!("Marking shard {tenant_shard_id} for notification retry, due to error {notify_error}");
1097 0 : let mut locked = self.inner.write().unwrap();
1098 0 : if let Some(shard) = locked.tenants.get_mut(&tenant_shard_id) {
1099 0 : shard.pending_compute_notification = true;
1100 0 : }
1101 :
1102 0 : }
1103 : },
1104 : None => {
1105 : // This channel is dead, but we don't want to terminate the outer loop{}: just wait for shutdown
1106 0 : self.cancel.cancelled().await;
1107 : }
1108 : }
1109 0 : } => {},
1110 : _ = self.cancel.cancelled() => {
1111 : break;
1112 : }
1113 0 : };
1114 0 : }
1115 0 : }
1116 :
1117 0 : async fn process_aborts(
1118 0 : &self,
1119 0 : mut abort_rx: tokio::sync::mpsc::UnboundedReceiver<TenantShardSplitAbort>,
1120 0 : ) {
1121 : loop {
1122 : // Wait for the next result, or for cancellation
1123 0 : let op = tokio::select! {
1124 : r = abort_rx.recv() => {
1125 : match r {
1126 : Some(op) => {op},
1127 : None => {break;}
1128 : }
1129 : }
1130 : _ = self.cancel.cancelled() => {
1131 : break;
1132 : }
1133 : };
1134 :
1135 : // Retry until shutdown: we must keep this request object alive until it is properly
1136 : // processed, as it holds a lock guard that prevents other operations trying to do things
1137 : // to the tenant while it is in a weird part-split state.
1138 0 : while !self.cancel.is_cancelled() {
1139 0 : match self.abort_tenant_shard_split(&op).await {
1140 0 : Ok(_) => break,
1141 0 : Err(e) => {
1142 0 : tracing::warn!(
1143 0 : "Failed to abort shard split on {}, will retry: {e}",
1144 : op.tenant_id
1145 : );
1146 :
1147 : // If a node is unavailable, we hope that it has been properly marked Offline
1148 : // when we retry, so that the abort op will succeed. If the abort op is failing
1149 : // for some other reason, we will keep retrying forever, or until a human notices
1150 : // and does something about it (either fixing a pageserver or restarting the controller).
1151 0 : tokio::time::timeout(Duration::from_secs(5), self.cancel.cancelled())
1152 0 : .await
1153 0 : .ok();
1154 : }
1155 : }
1156 : }
1157 : }
1158 0 : }
1159 :
1160 0 : pub async fn spawn(config: Config, persistence: Arc<Persistence>) -> anyhow::Result<Arc<Self>> {
1161 0 : let (result_tx, result_rx) = tokio::sync::mpsc::unbounded_channel();
1162 0 : let (abort_tx, abort_rx) = tokio::sync::mpsc::unbounded_channel();
1163 0 :
1164 0 : let leadership_cancel = CancellationToken::new();
1165 0 : let leadership = Leadership::new(persistence.clone(), config.clone(), leadership_cancel);
1166 0 : let (leader, leader_step_down_state) = leadership.step_down_current_leader().await?;
1167 :
1168 : // Apply the migrations **after** the current leader has stepped down
1169 : // (or we've given up waiting for it), but **before** reading from the
1170 : // database. The only exception is reading the current leader before
1171 : // migrating.
1172 0 : persistence.migration_run().await?;
1173 :
1174 0 : tracing::info!("Loading nodes from database...");
1175 0 : let nodes = persistence
1176 0 : .list_nodes()
1177 0 : .await?
1178 0 : .into_iter()
1179 0 : .map(Node::from_persistent)
1180 0 : .collect::<Vec<_>>();
1181 0 : let nodes: HashMap<NodeId, Node> = nodes.into_iter().map(|n| (n.get_id(), n)).collect();
1182 0 : tracing::info!("Loaded {} nodes from database.", nodes.len());
1183 :
1184 0 : tracing::info!("Loading shards from database...");
1185 0 : let mut tenant_shard_persistence = persistence.list_tenant_shards().await?;
1186 0 : tracing::info!(
1187 0 : "Loaded {} shards from database.",
1188 0 : tenant_shard_persistence.len()
1189 : );
1190 :
1191 : // If any shard splits were in progress, reset the database state to abort them
1192 0 : let mut tenant_shard_count_min_max: HashMap<TenantId, (ShardCount, ShardCount)> =
1193 0 : HashMap::new();
1194 0 : for tsp in &mut tenant_shard_persistence {
1195 0 : let shard = tsp.get_shard_identity()?;
1196 0 : let tenant_shard_id = tsp.get_tenant_shard_id()?;
1197 0 : let entry = tenant_shard_count_min_max
1198 0 : .entry(tenant_shard_id.tenant_id)
1199 0 : .or_insert_with(|| (shard.count, shard.count));
1200 0 : entry.0 = std::cmp::min(entry.0, shard.count);
1201 0 : entry.1 = std::cmp::max(entry.1, shard.count);
1202 0 : }
1203 :
1204 0 : for (tenant_id, (count_min, count_max)) in tenant_shard_count_min_max {
1205 0 : if count_min != count_max {
1206 : // Aborting the split in the database and dropping the child shards is sufficient: the reconciliation in
1207 : // [`Self::startup_reconcile`] will implicitly drop the child shards on remote pageservers, or they'll
1208 : // be dropped later in [`Self::node_activate_reconcile`] if it isn't available right now.
1209 0 : tracing::info!("Aborting shard split {tenant_id} {count_min:?} -> {count_max:?}");
1210 0 : let abort_status = persistence.abort_shard_split(tenant_id, count_max).await?;
1211 :
1212 : // We may never see the Complete status here: if the split was complete, we wouldn't have
1213 : // identified this tenant has having mismatching min/max counts.
1214 0 : assert!(matches!(abort_status, AbortShardSplitStatus::Aborted));
1215 :
1216 : // Clear the splitting status in-memory, to reflect that we just aborted in the database
1217 0 : tenant_shard_persistence.iter_mut().for_each(|tsp| {
1218 0 : // Set idle split state on those shards that we will retain.
1219 0 : let tsp_tenant_id = TenantId::from_str(tsp.tenant_id.as_str()).unwrap();
1220 0 : if tsp_tenant_id == tenant_id
1221 0 : && tsp.get_shard_identity().unwrap().count == count_min
1222 0 : {
1223 0 : tsp.splitting = SplitState::Idle;
1224 0 : } else if tsp_tenant_id == tenant_id {
1225 : // Leave the splitting state on the child shards: this will be used next to
1226 : // drop them.
1227 0 : tracing::info!(
1228 0 : "Shard {tsp_tenant_id} will be dropped after shard split abort",
1229 : );
1230 0 : }
1231 0 : });
1232 0 :
1233 0 : // Drop shards for this tenant which we didn't just mark idle (i.e. child shards of the aborted split)
1234 0 : tenant_shard_persistence.retain(|tsp| {
1235 0 : TenantId::from_str(tsp.tenant_id.as_str()).unwrap() != tenant_id
1236 0 : || tsp.splitting == SplitState::Idle
1237 0 : });
1238 0 : }
1239 : }
1240 :
1241 0 : let mut tenants = BTreeMap::new();
1242 0 :
1243 0 : let mut scheduler = Scheduler::new(nodes.values());
1244 0 :
1245 0 : #[cfg(feature = "testing")]
1246 0 : {
1247 0 : // Hack: insert scheduler state for all nodes referenced by shards, as compatibility
1248 0 : // tests only store the shards, not the nodes. The nodes will be loaded shortly
1249 0 : // after when pageservers start up and register.
1250 0 : let mut node_ids = HashSet::new();
1251 0 : for tsp in &tenant_shard_persistence {
1252 0 : if let Some(node_id) = tsp.generation_pageserver {
1253 0 : node_ids.insert(node_id);
1254 0 : }
1255 : }
1256 0 : for node_id in node_ids {
1257 0 : tracing::info!("Creating node {} in scheduler for tests", node_id);
1258 0 : let node = Node::new(
1259 0 : NodeId(node_id as u64),
1260 0 : "".to_string(),
1261 0 : 123,
1262 0 : "".to_string(),
1263 0 : 123,
1264 0 : None,
1265 0 : );
1266 0 :
1267 0 : scheduler.node_upsert(&node);
1268 : }
1269 : }
1270 0 : for tsp in tenant_shard_persistence {
1271 0 : let tenant_shard_id = tsp.get_tenant_shard_id()?;
1272 :
1273 : // We will populate intent properly later in [`Self::startup_reconcile`], initially populate
1274 : // it with what we can infer: the node for which a generation was most recently issued.
1275 0 : let mut intent = IntentState::new();
1276 0 : if let Some(generation_pageserver) = tsp.generation_pageserver.map(|n| NodeId(n as u64))
1277 : {
1278 0 : if nodes.contains_key(&generation_pageserver) {
1279 0 : intent.set_attached(&mut scheduler, Some(generation_pageserver));
1280 0 : } else {
1281 : // If a node was removed before being completely drained, it is legal for it to leave behind a `generation_pageserver` referring
1282 : // to a non-existent node, because node deletion doesn't block on completing the reconciliations that will issue new generations
1283 : // on different pageservers.
1284 0 : tracing::warn!("Tenant shard {tenant_shard_id} references non-existent node {generation_pageserver} in database, will be rescheduled");
1285 : }
1286 0 : }
1287 0 : let new_tenant = TenantShard::from_persistent(tsp, intent)?;
1288 :
1289 0 : tenants.insert(tenant_shard_id, new_tenant);
1290 : }
1291 :
1292 0 : let (startup_completion, startup_complete) = utils::completion::channel();
1293 0 :
1294 0 : // This channel is continuously consumed by process_results, so doesn't need to be very large.
1295 0 : let (bg_compute_notify_result_tx, bg_compute_notify_result_rx) =
1296 0 : tokio::sync::mpsc::channel(512);
1297 0 :
1298 0 : let (delayed_reconcile_tx, delayed_reconcile_rx) =
1299 0 : tokio::sync::mpsc::channel(MAX_DELAYED_RECONCILES);
1300 0 :
1301 0 : let cancel = CancellationToken::new();
1302 0 : let reconcilers_cancel = cancel.child_token();
1303 0 :
1304 0 : let heartbeater = Heartbeater::new(
1305 0 : config.jwt_token.clone(),
1306 0 : config.max_offline_interval,
1307 0 : config.max_warming_up_interval,
1308 0 : cancel.clone(),
1309 0 : );
1310 :
1311 0 : let initial_leadership_status = if config.start_as_candidate {
1312 0 : LeadershipStatus::Candidate
1313 : } else {
1314 0 : LeadershipStatus::Leader
1315 : };
1316 :
1317 0 : let this = Arc::new(Self {
1318 0 : inner: Arc::new(std::sync::RwLock::new(ServiceState::new(
1319 0 : nodes,
1320 0 : tenants,
1321 0 : scheduler,
1322 0 : delayed_reconcile_rx,
1323 0 : initial_leadership_status,
1324 0 : ))),
1325 0 : config: config.clone(),
1326 0 : persistence,
1327 0 : compute_hook: Arc::new(ComputeHook::new(config.clone())),
1328 0 : result_tx,
1329 0 : heartbeater,
1330 0 : reconciler_concurrency: Arc::new(tokio::sync::Semaphore::new(
1331 0 : config.reconciler_concurrency,
1332 0 : )),
1333 0 : delayed_reconcile_tx,
1334 0 : abort_tx,
1335 0 : startup_complete: startup_complete.clone(),
1336 0 : cancel,
1337 0 : reconcilers_cancel,
1338 0 : gate: Gate::default(),
1339 0 : reconcilers_gate: Gate::default(),
1340 0 : tenant_op_locks: Default::default(),
1341 0 : node_op_locks: Default::default(),
1342 0 : });
1343 0 :
1344 0 : let result_task_this = this.clone();
1345 0 : tokio::task::spawn(async move {
1346 : // Block shutdown until we're done (we must respect self.cancel)
1347 0 : if let Ok(_gate) = result_task_this.gate.enter() {
1348 0 : result_task_this
1349 0 : .process_results(result_rx, bg_compute_notify_result_rx)
1350 0 : .await
1351 0 : }
1352 0 : });
1353 0 :
1354 0 : tokio::task::spawn({
1355 0 : let this = this.clone();
1356 0 : async move {
1357 : // Block shutdown until we're done (we must respect self.cancel)
1358 0 : if let Ok(_gate) = this.gate.enter() {
1359 0 : this.process_aborts(abort_rx).await
1360 0 : }
1361 0 : }
1362 0 : });
1363 0 :
1364 0 : tokio::task::spawn({
1365 0 : let this = this.clone();
1366 0 : async move {
1367 0 : if let Ok(_gate) = this.gate.enter() {
1368 0 : loop {
1369 0 : tokio::select! {
1370 : _ = this.cancel.cancelled() => {
1371 : break;
1372 : },
1373 : _ = tokio::time::sleep(Duration::from_secs(60)) => {}
1374 : };
1375 0 : this.tenant_op_locks.housekeeping();
1376 : }
1377 0 : }
1378 0 : }
1379 0 : });
1380 0 :
1381 0 : tokio::task::spawn({
1382 0 : let this = this.clone();
1383 0 : // We will block the [`Service::startup_complete`] barrier until [`Self::startup_reconcile`]
1384 0 : // is done.
1385 0 : let startup_completion = startup_completion.clone();
1386 0 : async move {
1387 : // Block shutdown until we're done (we must respect self.cancel)
1388 0 : let Ok(_gate) = this.gate.enter() else {
1389 0 : return;
1390 : };
1391 :
1392 0 : this.startup_reconcile(leader, leader_step_down_state, bg_compute_notify_result_tx)
1393 0 : .await;
1394 :
1395 0 : drop(startup_completion);
1396 0 : }
1397 0 : });
1398 0 :
1399 0 : tokio::task::spawn({
1400 0 : let this = this.clone();
1401 0 : let startup_complete = startup_complete.clone();
1402 0 : async move {
1403 0 : startup_complete.wait().await;
1404 0 : this.background_reconcile().await;
1405 0 : }
1406 0 : });
1407 0 :
1408 0 : tokio::task::spawn({
1409 0 : let this = this.clone();
1410 0 : let startup_complete = startup_complete.clone();
1411 0 : async move {
1412 0 : startup_complete.wait().await;
1413 0 : this.spawn_heartbeat_driver().await;
1414 0 : }
1415 0 : });
1416 0 :
1417 0 : Ok(this)
1418 0 : }
1419 :
1420 0 : pub(crate) async fn attach_hook(
1421 0 : &self,
1422 0 : attach_req: AttachHookRequest,
1423 0 : ) -> anyhow::Result<AttachHookResponse> {
1424 0 : let _tenant_lock = trace_exclusive_lock(
1425 0 : &self.tenant_op_locks,
1426 0 : attach_req.tenant_shard_id.tenant_id,
1427 0 : TenantOperations::AttachHook,
1428 0 : )
1429 0 : .await;
1430 :
1431 : // This is a test hook. To enable using it on tenants that were created directly with
1432 : // the pageserver API (not via this service), we will auto-create any missing tenant
1433 : // shards with default state.
1434 0 : let insert = {
1435 0 : let locked = self.inner.write().unwrap();
1436 0 : !locked.tenants.contains_key(&attach_req.tenant_shard_id)
1437 0 : };
1438 0 :
1439 0 : if insert {
1440 0 : let tsp = TenantShardPersistence {
1441 0 : tenant_id: attach_req.tenant_shard_id.tenant_id.to_string(),
1442 0 : shard_number: attach_req.tenant_shard_id.shard_number.0 as i32,
1443 0 : shard_count: attach_req.tenant_shard_id.shard_count.literal() as i32,
1444 0 : shard_stripe_size: 0,
1445 0 : generation: attach_req.generation_override.or(Some(0)),
1446 0 : generation_pageserver: None,
1447 0 : placement_policy: serde_json::to_string(&PlacementPolicy::Attached(0)).unwrap(),
1448 0 : config: serde_json::to_string(&TenantConfig::default()).unwrap(),
1449 0 : splitting: SplitState::default(),
1450 0 : scheduling_policy: serde_json::to_string(&ShardSchedulingPolicy::default())
1451 0 : .unwrap(),
1452 0 : };
1453 0 :
1454 0 : match self.persistence.insert_tenant_shards(vec![tsp]).await {
1455 0 : Err(e) => match e {
1456 : DatabaseError::Query(diesel::result::Error::DatabaseError(
1457 : DatabaseErrorKind::UniqueViolation,
1458 : _,
1459 : )) => {
1460 0 : tracing::info!(
1461 0 : "Raced with another request to insert tenant {}",
1462 : attach_req.tenant_shard_id
1463 : )
1464 : }
1465 0 : _ => return Err(e.into()),
1466 : },
1467 : Ok(()) => {
1468 0 : tracing::info!("Inserted shard {} in database", attach_req.tenant_shard_id);
1469 :
1470 0 : let mut locked = self.inner.write().unwrap();
1471 0 : locked.tenants.insert(
1472 0 : attach_req.tenant_shard_id,
1473 0 : TenantShard::new(
1474 0 : attach_req.tenant_shard_id,
1475 0 : ShardIdentity::unsharded(),
1476 0 : PlacementPolicy::Attached(0),
1477 0 : ),
1478 0 : );
1479 0 : tracing::info!("Inserted shard {} in memory", attach_req.tenant_shard_id);
1480 : }
1481 : }
1482 0 : }
1483 :
1484 0 : let new_generation = if let Some(req_node_id) = attach_req.node_id {
1485 0 : let maybe_tenant_conf = {
1486 0 : let locked = self.inner.write().unwrap();
1487 0 : locked
1488 0 : .tenants
1489 0 : .get(&attach_req.tenant_shard_id)
1490 0 : .map(|t| t.config.clone())
1491 0 : };
1492 0 :
1493 0 : match maybe_tenant_conf {
1494 0 : Some(conf) => {
1495 0 : let new_generation = self
1496 0 : .persistence
1497 0 : .increment_generation(attach_req.tenant_shard_id, req_node_id)
1498 0 : .await?;
1499 :
1500 : // Persist the placement policy update. This is required
1501 : // when we reattaching a detached tenant.
1502 0 : self.persistence
1503 0 : .update_tenant_shard(
1504 0 : TenantFilter::Shard(attach_req.tenant_shard_id),
1505 0 : Some(PlacementPolicy::Attached(0)),
1506 0 : Some(conf),
1507 0 : None,
1508 0 : None,
1509 0 : )
1510 0 : .await?;
1511 0 : Some(new_generation)
1512 : }
1513 : None => {
1514 0 : anyhow::bail!("Attach hook handling raced with tenant removal")
1515 : }
1516 : }
1517 : } else {
1518 0 : self.persistence.detach(attach_req.tenant_shard_id).await?;
1519 0 : None
1520 : };
1521 :
1522 0 : let mut locked = self.inner.write().unwrap();
1523 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
1524 0 :
1525 0 : let tenant_shard = tenants
1526 0 : .get_mut(&attach_req.tenant_shard_id)
1527 0 : .expect("Checked for existence above");
1528 :
1529 0 : if let Some(new_generation) = new_generation {
1530 0 : tenant_shard.generation = Some(new_generation);
1531 0 : tenant_shard.policy = PlacementPolicy::Attached(0);
1532 0 : } else {
1533 : // This is a detach notification. We must update placement policy to avoid re-attaching
1534 : // during background scheduling/reconciliation, or during storage controller restart.
1535 0 : assert!(attach_req.node_id.is_none());
1536 0 : tenant_shard.policy = PlacementPolicy::Detached;
1537 : }
1538 :
1539 0 : if let Some(attaching_pageserver) = attach_req.node_id.as_ref() {
1540 0 : tracing::info!(
1541 : tenant_id = %attach_req.tenant_shard_id,
1542 : ps_id = %attaching_pageserver,
1543 : generation = ?tenant_shard.generation,
1544 0 : "issuing",
1545 : );
1546 0 : } else if let Some(ps_id) = tenant_shard.intent.get_attached() {
1547 0 : tracing::info!(
1548 : tenant_id = %attach_req.tenant_shard_id,
1549 : %ps_id,
1550 : generation = ?tenant_shard.generation,
1551 0 : "dropping",
1552 : );
1553 : } else {
1554 0 : tracing::info!(
1555 : tenant_id = %attach_req.tenant_shard_id,
1556 0 : "no-op: tenant already has no pageserver");
1557 : }
1558 0 : tenant_shard
1559 0 : .intent
1560 0 : .set_attached(scheduler, attach_req.node_id);
1561 0 :
1562 0 : tracing::info!(
1563 0 : "attach_hook: tenant {} set generation {:?}, pageserver {}",
1564 0 : attach_req.tenant_shard_id,
1565 0 : tenant_shard.generation,
1566 0 : // TODO: this is an odd number of 0xf's
1567 0 : attach_req.node_id.unwrap_or(utils::id::NodeId(0xfffffff))
1568 : );
1569 :
1570 : // Trick the reconciler into not doing anything for this tenant: this helps
1571 : // tests that manually configure a tenant on the pagesrever, and then call this
1572 : // attach hook: they don't want background reconciliation to modify what they
1573 : // did to the pageserver.
1574 : #[cfg(feature = "testing")]
1575 : {
1576 0 : if let Some(node_id) = attach_req.node_id {
1577 0 : tenant_shard.observed.locations = HashMap::from([(
1578 0 : node_id,
1579 0 : ObservedStateLocation {
1580 0 : conf: Some(attached_location_conf(
1581 0 : tenant_shard.generation.unwrap(),
1582 0 : &tenant_shard.shard,
1583 0 : &tenant_shard.config,
1584 0 : &PlacementPolicy::Attached(0),
1585 0 : )),
1586 0 : },
1587 0 : )]);
1588 0 : } else {
1589 0 : tenant_shard.observed.locations.clear();
1590 0 : }
1591 : }
1592 :
1593 0 : Ok(AttachHookResponse {
1594 0 : gen: attach_req
1595 0 : .node_id
1596 0 : .map(|_| tenant_shard.generation.expect("Test hook, not used on tenants that are mid-onboarding with a NULL generation").into().unwrap()),
1597 0 : })
1598 0 : }
1599 :
1600 0 : pub(crate) fn inspect(&self, inspect_req: InspectRequest) -> InspectResponse {
1601 0 : let locked = self.inner.read().unwrap();
1602 0 :
1603 0 : let tenant_shard = locked.tenants.get(&inspect_req.tenant_shard_id);
1604 0 :
1605 0 : InspectResponse {
1606 0 : attachment: tenant_shard.and_then(|s| {
1607 0 : s.intent
1608 0 : .get_attached()
1609 0 : .map(|ps| (s.generation.expect("Test hook, not used on tenants that are mid-onboarding with a NULL generation").into().unwrap(), ps))
1610 0 : }),
1611 0 : }
1612 0 : }
1613 :
1614 : // When the availability state of a node transitions to active, we must do a full reconciliation
1615 : // of LocationConfigs on that node. This is because while a node was offline:
1616 : // - we might have proceeded through startup_reconcile without checking for extraneous LocationConfigs on this node
1617 : // - aborting a tenant shard split might have left rogue child shards behind on this node.
1618 : //
1619 : // This function must complete _before_ setting a `Node` to Active: once it is set to Active, other
1620 : // Reconcilers might communicate with the node, and these must not overlap with the work we do in
1621 : // this function.
1622 : //
1623 : // The reconciliation logic in here is very similar to what [`Self::startup_reconcile`] does, but
1624 : // for written for a single node rather than as a batch job for all nodes.
1625 0 : #[tracing::instrument(skip_all, fields(node_id=%node.get_id()))]
1626 : async fn node_activate_reconcile(
1627 : &self,
1628 : mut node: Node,
1629 : _lock: &TracingExclusiveGuard<NodeOperations>,
1630 : ) -> Result<(), ApiError> {
1631 : // This Node is a mutable local copy: we will set it active so that we can use its
1632 : // API client to reconcile with the node. The Node in [`Self::nodes`] will get updated
1633 : // later.
1634 : node.set_availability(NodeAvailability::Active(PageserverUtilization::full()));
1635 :
1636 : let configs = match node
1637 : .with_client_retries(
1638 0 : |client| async move { client.list_location_config().await },
1639 : &self.config.jwt_token,
1640 : 1,
1641 : 5,
1642 : SHORT_RECONCILE_TIMEOUT,
1643 : &self.cancel,
1644 : )
1645 : .await
1646 : {
1647 : None => {
1648 : // We're shutting down (the Node's cancellation token can't have fired, because
1649 : // we're the only scope that has a reference to it, and we didn't fire it).
1650 : return Err(ApiError::ShuttingDown);
1651 : }
1652 : Some(Err(e)) => {
1653 : // This node didn't succeed listing its locations: it may not proceed to active state
1654 : // as it is apparently unavailable.
1655 : return Err(ApiError::PreconditionFailed(
1656 : format!("Failed to query node location configs, cannot activate ({e})").into(),
1657 : ));
1658 : }
1659 : Some(Ok(configs)) => configs,
1660 : };
1661 : tracing::info!("Loaded {} LocationConfigs", configs.tenant_shards.len());
1662 :
1663 : let mut cleanup = Vec::new();
1664 : {
1665 : let mut locked = self.inner.write().unwrap();
1666 :
1667 : for (tenant_shard_id, observed_loc) in configs.tenant_shards {
1668 : let Some(tenant_shard) = locked.tenants.get_mut(&tenant_shard_id) else {
1669 : cleanup.push(tenant_shard_id);
1670 : continue;
1671 : };
1672 : tenant_shard
1673 : .observed
1674 : .locations
1675 : .insert(node.get_id(), ObservedStateLocation { conf: observed_loc });
1676 : }
1677 : }
1678 :
1679 : for tenant_shard_id in cleanup {
1680 : tracing::info!("Detaching {tenant_shard_id}");
1681 : match node
1682 : .with_client_retries(
1683 0 : |client| async move {
1684 0 : let config = LocationConfig {
1685 0 : mode: LocationConfigMode::Detached,
1686 0 : generation: None,
1687 0 : secondary_conf: None,
1688 0 : shard_number: tenant_shard_id.shard_number.0,
1689 0 : shard_count: tenant_shard_id.shard_count.literal(),
1690 0 : shard_stripe_size: 0,
1691 0 : tenant_conf: models::TenantConfig::default(),
1692 0 : };
1693 0 : client
1694 0 : .location_config(tenant_shard_id, config, None, false)
1695 0 : .await
1696 0 : },
1697 : &self.config.jwt_token,
1698 : 1,
1699 : 5,
1700 : SHORT_RECONCILE_TIMEOUT,
1701 : &self.cancel,
1702 : )
1703 : .await
1704 : {
1705 : None => {
1706 : // We're shutting down (the Node's cancellation token can't have fired, because
1707 : // we're the only scope that has a reference to it, and we didn't fire it).
1708 : return Err(ApiError::ShuttingDown);
1709 : }
1710 : Some(Err(e)) => {
1711 : // Do not let the node proceed to Active state if it is not responsive to requests
1712 : // to detach. This could happen if e.g. a shutdown bug in the pageserver is preventing
1713 : // detach completing: we should not let this node back into the set of nodes considered
1714 : // okay for scheduling.
1715 : return Err(ApiError::Conflict(format!(
1716 : "Node {node} failed to detach {tenant_shard_id}: {e}"
1717 : )));
1718 : }
1719 : Some(Ok(_)) => {}
1720 : };
1721 : }
1722 :
1723 : Ok(())
1724 : }
1725 :
1726 0 : pub(crate) async fn re_attach(
1727 0 : &self,
1728 0 : reattach_req: ReAttachRequest,
1729 0 : ) -> Result<ReAttachResponse, ApiError> {
1730 0 : if let Some(register_req) = reattach_req.register {
1731 0 : self.node_register(register_req).await?;
1732 0 : }
1733 :
1734 : // Ordering: we must persist generation number updates before making them visible in the in-memory state
1735 0 : let incremented_generations = self.persistence.re_attach(reattach_req.node_id).await?;
1736 :
1737 0 : tracing::info!(
1738 : node_id=%reattach_req.node_id,
1739 0 : "Incremented {} tenant shards' generations",
1740 0 : incremented_generations.len()
1741 : );
1742 :
1743 : // Apply the updated generation to our in-memory state, and
1744 : // gather discover secondary locations.
1745 0 : let mut locked = self.inner.write().unwrap();
1746 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
1747 0 :
1748 0 : let mut response = ReAttachResponse {
1749 0 : tenants: Vec::new(),
1750 0 : };
1751 :
1752 : // TODO: cancel/restart any running reconciliation for this tenant, it might be trying
1753 : // to call location_conf API with an old generation. Wait for cancellation to complete
1754 : // before responding to this request. Requires well implemented CancellationToken logic
1755 : // all the way to where we call location_conf. Even then, there can still be a location_conf
1756 : // request in flight over the network: TODO handle that by making location_conf API refuse
1757 : // to go backward in generations.
1758 :
1759 : // Scan through all shards, applying updates for ones where we updated generation
1760 : // and identifying shards that intend to have a secondary location on this node.
1761 0 : for (tenant_shard_id, shard) in tenants {
1762 0 : if let Some(new_gen) = incremented_generations.get(tenant_shard_id) {
1763 0 : let new_gen = *new_gen;
1764 0 : response.tenants.push(ReAttachResponseTenant {
1765 0 : id: *tenant_shard_id,
1766 0 : gen: Some(new_gen.into().unwrap()),
1767 0 : // A tenant is only put into multi or stale modes in the middle of a [`Reconciler::live_migrate`]
1768 0 : // execution. If a pageserver is restarted during that process, then the reconcile pass will
1769 0 : // fail, and start from scratch, so it doesn't make sense for us to try and preserve
1770 0 : // the stale/multi states at this point.
1771 0 : mode: LocationConfigMode::AttachedSingle,
1772 0 : });
1773 0 :
1774 0 : shard.generation = std::cmp::max(shard.generation, Some(new_gen));
1775 0 : if let Some(observed) = shard.observed.locations.get_mut(&reattach_req.node_id) {
1776 : // Why can we update `observed` even though we're not sure our response will be received
1777 : // by the pageserver? Because the pageserver will not proceed with startup until
1778 : // it has processed response: if it loses it, we'll see another request and increment
1779 : // generation again, avoiding any uncertainty about dirtiness of tenant's state.
1780 0 : if let Some(conf) = observed.conf.as_mut() {
1781 0 : conf.generation = new_gen.into();
1782 0 : }
1783 0 : } else {
1784 0 : // This node has no observed state for the shard: perhaps it was offline
1785 0 : // when the pageserver restarted. Insert a None, so that the Reconciler
1786 0 : // will be prompted to learn the location's state before it makes changes.
1787 0 : shard
1788 0 : .observed
1789 0 : .locations
1790 0 : .insert(reattach_req.node_id, ObservedStateLocation { conf: None });
1791 0 : }
1792 0 : } else if shard.intent.get_secondary().contains(&reattach_req.node_id) {
1793 0 : // Ordering: pageserver will not accept /location_config requests until it has
1794 0 : // finished processing the response from re-attach. So we can update our in-memory state
1795 0 : // now, and be confident that we are not stamping on the result of some later location config.
1796 0 : // TODO: however, we are not strictly ordered wrt ReconcileResults queue,
1797 0 : // so we might update observed state here, and then get over-written by some racing
1798 0 : // ReconcileResult. The impact is low however, since we have set state on pageserver something
1799 0 : // that matches intent, so worst case if we race then we end up doing a spurious reconcile.
1800 0 :
1801 0 : response.tenants.push(ReAttachResponseTenant {
1802 0 : id: *tenant_shard_id,
1803 0 : gen: None,
1804 0 : mode: LocationConfigMode::Secondary,
1805 0 : });
1806 0 :
1807 0 : // We must not update observed, because we have no guarantee that our
1808 0 : // response will be received by the pageserver. This could leave it
1809 0 : // falsely dirty, but the resulting reconcile should be idempotent.
1810 0 : }
1811 : }
1812 :
1813 : // We consider a node Active once we have composed a re-attach response, but we
1814 : // do not call [`Self::node_activate_reconcile`]: the handling of the re-attach response
1815 : // implicitly synchronizes the LocationConfigs on the node.
1816 : //
1817 : // Setting a node active unblocks any Reconcilers that might write to the location config API,
1818 : // but those requests will not be accepted by the node until it has finished processing
1819 : // the re-attach response.
1820 : //
1821 : // Additionally, reset the nodes scheduling policy to match the conditional update done
1822 : // in [`Persistence::re_attach`].
1823 0 : if let Some(node) = nodes.get(&reattach_req.node_id) {
1824 0 : let reset_scheduling = matches!(
1825 0 : node.get_scheduling(),
1826 : NodeSchedulingPolicy::PauseForRestart
1827 : | NodeSchedulingPolicy::Draining
1828 : | NodeSchedulingPolicy::Filling
1829 : );
1830 :
1831 0 : let mut new_nodes = (**nodes).clone();
1832 0 : if let Some(node) = new_nodes.get_mut(&reattach_req.node_id) {
1833 0 : if reset_scheduling {
1834 0 : node.set_scheduling(NodeSchedulingPolicy::Active);
1835 0 : }
1836 :
1837 0 : tracing::info!("Marking {} warming-up on reattach", reattach_req.node_id);
1838 0 : node.set_availability(NodeAvailability::WarmingUp(std::time::Instant::now()));
1839 0 :
1840 0 : scheduler.node_upsert(node);
1841 0 : let new_nodes = Arc::new(new_nodes);
1842 0 : *nodes = new_nodes;
1843 : } else {
1844 0 : tracing::error!(
1845 0 : "Reattaching node {} was removed while processing the request",
1846 : reattach_req.node_id
1847 : );
1848 : }
1849 0 : }
1850 :
1851 0 : Ok(response)
1852 0 : }
1853 :
1854 0 : pub(crate) fn validate(&self, validate_req: ValidateRequest) -> ValidateResponse {
1855 0 : let locked = self.inner.read().unwrap();
1856 0 :
1857 0 : let mut response = ValidateResponse {
1858 0 : tenants: Vec::new(),
1859 0 : };
1860 :
1861 0 : for req_tenant in validate_req.tenants {
1862 0 : if let Some(tenant_shard) = locked.tenants.get(&req_tenant.id) {
1863 0 : let valid = tenant_shard.generation == Some(Generation::new(req_tenant.gen));
1864 0 : tracing::info!(
1865 0 : "handle_validate: {}(gen {}): valid={valid} (latest {:?})",
1866 : req_tenant.id,
1867 : req_tenant.gen,
1868 : tenant_shard.generation
1869 : );
1870 0 : response.tenants.push(ValidateResponseTenant {
1871 0 : id: req_tenant.id,
1872 0 : valid,
1873 0 : });
1874 0 : } else {
1875 0 : // After tenant deletion, we may approve any validation. This avoids
1876 0 : // spurious warnings on the pageserver if it has pending LSN updates
1877 0 : // at the point a deletion happens.
1878 0 : response.tenants.push(ValidateResponseTenant {
1879 0 : id: req_tenant.id,
1880 0 : valid: true,
1881 0 : });
1882 0 : }
1883 : }
1884 0 : response
1885 0 : }
1886 :
1887 0 : pub(crate) async fn tenant_create(
1888 0 : &self,
1889 0 : create_req: TenantCreateRequest,
1890 0 : ) -> Result<TenantCreateResponse, ApiError> {
1891 0 : let tenant_id = create_req.new_tenant_id.tenant_id;
1892 :
1893 : // Exclude any concurrent attempts to create/access the same tenant ID
1894 0 : let _tenant_lock = trace_exclusive_lock(
1895 0 : &self.tenant_op_locks,
1896 0 : create_req.new_tenant_id.tenant_id,
1897 0 : TenantOperations::Create,
1898 0 : )
1899 0 : .await;
1900 0 : let (response, waiters) = self.do_tenant_create(create_req).await?;
1901 :
1902 0 : if let Err(e) = self.await_waiters(waiters, RECONCILE_TIMEOUT).await {
1903 : // Avoid deadlock: reconcile may fail while notifying compute, if the cloud control plane refuses to
1904 : // accept compute notifications while it is in the process of creating. Reconciliation will
1905 : // be retried in the background.
1906 0 : tracing::warn!(%tenant_id, "Reconcile not done yet while creating tenant ({e})");
1907 0 : }
1908 0 : Ok(response)
1909 0 : }
1910 :
1911 0 : pub(crate) async fn do_tenant_create(
1912 0 : &self,
1913 0 : create_req: TenantCreateRequest,
1914 0 : ) -> Result<(TenantCreateResponse, Vec<ReconcilerWaiter>), ApiError> {
1915 0 : let placement_policy = create_req
1916 0 : .placement_policy
1917 0 : .clone()
1918 0 : // As a default, zero secondaries is convenient for tests that don't choose a policy.
1919 0 : .unwrap_or(PlacementPolicy::Attached(0));
1920 :
1921 : // This service expects to handle sharding itself: it is an error to try and directly create
1922 : // a particular shard here.
1923 0 : let tenant_id = if !create_req.new_tenant_id.is_unsharded() {
1924 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
1925 0 : "Attempted to create a specific shard, this API is for creating the whole tenant"
1926 0 : )));
1927 : } else {
1928 0 : create_req.new_tenant_id.tenant_id
1929 0 : };
1930 0 :
1931 0 : tracing::info!(
1932 0 : "Creating tenant {}, shard_count={:?}",
1933 : create_req.new_tenant_id,
1934 : create_req.shard_parameters.count,
1935 : );
1936 :
1937 0 : let create_ids = (0..create_req.shard_parameters.count.count())
1938 0 : .map(|i| TenantShardId {
1939 0 : tenant_id,
1940 0 : shard_number: ShardNumber(i),
1941 0 : shard_count: create_req.shard_parameters.count,
1942 0 : })
1943 0 : .collect::<Vec<_>>();
1944 :
1945 : // If the caller specifies a None generation, it means "start from default". This is different
1946 : // to [`Self::tenant_location_config`], where a None generation is used to represent
1947 : // an incompletely-onboarded tenant.
1948 0 : let initial_generation = if matches!(placement_policy, PlacementPolicy::Secondary) {
1949 0 : tracing::info!(
1950 0 : "tenant_create: secondary mode, generation is_some={}",
1951 0 : create_req.generation.is_some()
1952 : );
1953 0 : create_req.generation.map(Generation::new)
1954 : } else {
1955 0 : tracing::info!(
1956 0 : "tenant_create: not secondary mode, generation is_some={}",
1957 0 : create_req.generation.is_some()
1958 : );
1959 0 : Some(
1960 0 : create_req
1961 0 : .generation
1962 0 : .map(Generation::new)
1963 0 : .unwrap_or(INITIAL_GENERATION),
1964 0 : )
1965 : };
1966 :
1967 : // Ordering: we persist tenant shards before creating them on the pageserver. This enables a caller
1968 : // to clean up after themselves by issuing a tenant deletion if something goes wrong and we restart
1969 : // during the creation, rather than risking leaving orphan objects in S3.
1970 0 : let persist_tenant_shards = create_ids
1971 0 : .iter()
1972 0 : .map(|tenant_shard_id| TenantShardPersistence {
1973 0 : tenant_id: tenant_shard_id.tenant_id.to_string(),
1974 0 : shard_number: tenant_shard_id.shard_number.0 as i32,
1975 0 : shard_count: tenant_shard_id.shard_count.literal() as i32,
1976 0 : shard_stripe_size: create_req.shard_parameters.stripe_size.0 as i32,
1977 0 : generation: initial_generation.map(|g| g.into().unwrap() as i32),
1978 0 : // The pageserver is not known until scheduling happens: we will set this column when
1979 0 : // incrementing the generation the first time we attach to a pageserver.
1980 0 : generation_pageserver: None,
1981 0 : placement_policy: serde_json::to_string(&placement_policy).unwrap(),
1982 0 : config: serde_json::to_string(&create_req.config).unwrap(),
1983 0 : splitting: SplitState::default(),
1984 0 : scheduling_policy: serde_json::to_string(&ShardSchedulingPolicy::default())
1985 0 : .unwrap(),
1986 0 : })
1987 0 : .collect();
1988 0 :
1989 0 : match self
1990 0 : .persistence
1991 0 : .insert_tenant_shards(persist_tenant_shards)
1992 0 : .await
1993 : {
1994 0 : Ok(_) => {}
1995 : Err(DatabaseError::Query(diesel::result::Error::DatabaseError(
1996 : DatabaseErrorKind::UniqueViolation,
1997 : _,
1998 : ))) => {
1999 : // Unique key violation: this is probably a retry. Because the shard count is part of the unique key,
2000 : // if we see a unique key violation it means that the creation request's shard count matches the previous
2001 : // creation's shard count.
2002 0 : tracing::info!("Tenant shards already present in database, proceeding with idempotent creation...");
2003 : }
2004 : // Any other database error is unexpected and a bug.
2005 0 : Err(e) => return Err(ApiError::InternalServerError(anyhow::anyhow!(e))),
2006 : };
2007 :
2008 0 : let mut schedule_context = ScheduleContext::default();
2009 :
2010 0 : let (waiters, response_shards) = {
2011 0 : let mut locked = self.inner.write().unwrap();
2012 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
2013 0 :
2014 0 : let mut response_shards = Vec::new();
2015 0 : let mut schcedule_error = None;
2016 :
2017 0 : for tenant_shard_id in create_ids {
2018 0 : tracing::info!("Creating shard {tenant_shard_id}...");
2019 :
2020 : use std::collections::btree_map::Entry;
2021 0 : match tenants.entry(tenant_shard_id) {
2022 0 : Entry::Occupied(mut entry) => {
2023 0 : tracing::info!(
2024 0 : "Tenant shard {tenant_shard_id} already exists while creating"
2025 : );
2026 :
2027 : // TODO: schedule() should take an anti-affinity expression that pushes
2028 : // attached and secondary locations (independently) away frorm those
2029 : // pageservers also holding a shard for this tenant.
2030 :
2031 0 : entry
2032 0 : .get_mut()
2033 0 : .schedule(scheduler, &mut schedule_context)
2034 0 : .map_err(|e| {
2035 0 : ApiError::Conflict(format!(
2036 0 : "Failed to schedule shard {tenant_shard_id}: {e}"
2037 0 : ))
2038 0 : })?;
2039 :
2040 0 : if let Some(node_id) = entry.get().intent.get_attached() {
2041 0 : let generation = entry
2042 0 : .get()
2043 0 : .generation
2044 0 : .expect("Generation is set when in attached mode");
2045 0 : response_shards.push(TenantCreateResponseShard {
2046 0 : shard_id: tenant_shard_id,
2047 0 : node_id: *node_id,
2048 0 : generation: generation.into().unwrap(),
2049 0 : });
2050 0 : }
2051 :
2052 0 : continue;
2053 : }
2054 0 : Entry::Vacant(entry) => {
2055 0 : let state = entry.insert(TenantShard::new(
2056 0 : tenant_shard_id,
2057 0 : ShardIdentity::from_params(
2058 0 : tenant_shard_id.shard_number,
2059 0 : &create_req.shard_parameters,
2060 0 : ),
2061 0 : placement_policy.clone(),
2062 0 : ));
2063 0 :
2064 0 : state.generation = initial_generation;
2065 0 : state.config = create_req.config.clone();
2066 0 : if let Err(e) = state.schedule(scheduler, &mut schedule_context) {
2067 0 : schcedule_error = Some(e);
2068 0 : }
2069 :
2070 : // Only include shards in result if we are attaching: the purpose
2071 : // of the response is to tell the caller where the shards are attached.
2072 0 : if let Some(node_id) = state.intent.get_attached() {
2073 0 : let generation = state
2074 0 : .generation
2075 0 : .expect("Generation is set when in attached mode");
2076 0 : response_shards.push(TenantCreateResponseShard {
2077 0 : shard_id: tenant_shard_id,
2078 0 : node_id: *node_id,
2079 0 : generation: generation.into().unwrap(),
2080 0 : });
2081 0 : }
2082 : }
2083 : };
2084 : }
2085 :
2086 : // If we failed to schedule shards, then they are still created in the controller,
2087 : // but we return an error to the requester to avoid a silent failure when someone
2088 : // tries to e.g. create a tenant whose placement policy requires more nodes than
2089 : // are present in the system. We do this here rather than in the above loop, to
2090 : // avoid situations where we only create a subset of shards in the tenant.
2091 0 : if let Some(e) = schcedule_error {
2092 0 : return Err(ApiError::Conflict(format!(
2093 0 : "Failed to schedule shard(s): {e}"
2094 0 : )));
2095 0 : }
2096 0 :
2097 0 : let waiters = tenants
2098 0 : .range_mut(TenantShardId::tenant_range(tenant_id))
2099 0 : .filter_map(|(_shard_id, shard)| self.maybe_reconcile_shard(shard, nodes))
2100 0 : .collect::<Vec<_>>();
2101 0 : (waiters, response_shards)
2102 0 : };
2103 0 :
2104 0 : Ok((
2105 0 : TenantCreateResponse {
2106 0 : shards: response_shards,
2107 0 : },
2108 0 : waiters,
2109 0 : ))
2110 0 : }
2111 :
2112 : /// Helper for functions that reconcile a number of shards, and would like to do a timeout-bounded
2113 : /// wait for reconciliation to complete before responding.
2114 0 : async fn await_waiters(
2115 0 : &self,
2116 0 : waiters: Vec<ReconcilerWaiter>,
2117 0 : timeout: Duration,
2118 0 : ) -> Result<(), ReconcileWaitError> {
2119 0 : let deadline = Instant::now().checked_add(timeout).unwrap();
2120 0 : for waiter in waiters {
2121 0 : let timeout = deadline.duration_since(Instant::now());
2122 0 : waiter.wait_timeout(timeout).await?;
2123 : }
2124 :
2125 0 : Ok(())
2126 0 : }
2127 :
2128 : /// Same as [`Service::await_waiters`], but returns the waiters which are still
2129 : /// in progress
2130 0 : async fn await_waiters_remainder(
2131 0 : &self,
2132 0 : waiters: Vec<ReconcilerWaiter>,
2133 0 : timeout: Duration,
2134 0 : ) -> Vec<ReconcilerWaiter> {
2135 0 : let deadline = Instant::now().checked_add(timeout).unwrap();
2136 0 : for waiter in waiters.iter() {
2137 0 : let timeout = deadline.duration_since(Instant::now());
2138 0 : let _ = waiter.wait_timeout(timeout).await;
2139 : }
2140 :
2141 0 : waiters
2142 0 : .into_iter()
2143 0 : .filter(|waiter| matches!(waiter.get_status(), ReconcilerStatus::InProgress))
2144 0 : .collect::<Vec<_>>()
2145 0 : }
2146 :
2147 : /// Part of [`Self::tenant_location_config`]: dissect an incoming location config request,
2148 : /// and transform it into either a tenant creation of a series of shard updates.
2149 : ///
2150 : /// If the incoming request makes no changes, a [`TenantCreateOrUpdate::Update`] result will
2151 : /// still be returned.
2152 0 : fn tenant_location_config_prepare(
2153 0 : &self,
2154 0 : tenant_id: TenantId,
2155 0 : req: TenantLocationConfigRequest,
2156 0 : ) -> TenantCreateOrUpdate {
2157 0 : let mut updates = Vec::new();
2158 0 : let mut locked = self.inner.write().unwrap();
2159 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
2160 0 : let tenant_shard_id = TenantShardId::unsharded(tenant_id);
2161 :
2162 : // Use location config mode as an indicator of policy.
2163 0 : let placement_policy = match req.config.mode {
2164 0 : LocationConfigMode::Detached => PlacementPolicy::Detached,
2165 0 : LocationConfigMode::Secondary => PlacementPolicy::Secondary,
2166 : LocationConfigMode::AttachedMulti
2167 : | LocationConfigMode::AttachedSingle
2168 : | LocationConfigMode::AttachedStale => {
2169 0 : if nodes.len() > 1 {
2170 0 : PlacementPolicy::Attached(1)
2171 : } else {
2172 : // Convenience for dev/test: if we just have one pageserver, import
2173 : // tenants into non-HA mode so that scheduling will succeed.
2174 0 : PlacementPolicy::Attached(0)
2175 : }
2176 : }
2177 : };
2178 :
2179 0 : let mut create = true;
2180 0 : for (shard_id, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
2181 : // Saw an existing shard: this is not a creation
2182 0 : create = false;
2183 :
2184 : // Shards may have initially been created by a Secondary request, where we
2185 : // would have left generation as None.
2186 : //
2187 : // We only update generation the first time we see an attached-mode request,
2188 : // and if there is no existing generation set. The caller is responsible for
2189 : // ensuring that no non-storage-controller pageserver ever uses a higher
2190 : // generation than they passed in here.
2191 : use LocationConfigMode::*;
2192 0 : let set_generation = match req.config.mode {
2193 0 : AttachedMulti | AttachedSingle | AttachedStale if shard.generation.is_none() => {
2194 0 : req.config.generation.map(Generation::new)
2195 : }
2196 0 : _ => None,
2197 : };
2198 :
2199 0 : updates.push(ShardUpdate {
2200 0 : tenant_shard_id: *shard_id,
2201 0 : placement_policy: placement_policy.clone(),
2202 0 : tenant_config: req.config.tenant_conf.clone(),
2203 0 : generation: set_generation,
2204 0 : });
2205 : }
2206 :
2207 0 : if create {
2208 : use LocationConfigMode::*;
2209 0 : let generation = match req.config.mode {
2210 0 : AttachedMulti | AttachedSingle | AttachedStale => req.config.generation,
2211 : // If a caller provided a generation in a non-attached request, ignore it
2212 : // and leave our generation as None: this enables a subsequent update to set
2213 : // the generation when setting an attached mode for the first time.
2214 0 : _ => None,
2215 : };
2216 :
2217 0 : TenantCreateOrUpdate::Create(
2218 0 : // Synthesize a creation request
2219 0 : TenantCreateRequest {
2220 0 : new_tenant_id: tenant_shard_id,
2221 0 : generation,
2222 0 : shard_parameters: ShardParameters {
2223 0 : count: tenant_shard_id.shard_count,
2224 0 : // We only import un-sharded or single-sharded tenants, so stripe
2225 0 : // size can be made up arbitrarily here.
2226 0 : stripe_size: ShardParameters::DEFAULT_STRIPE_SIZE,
2227 0 : },
2228 0 : placement_policy: Some(placement_policy),
2229 0 : config: req.config.tenant_conf,
2230 0 : },
2231 0 : )
2232 : } else {
2233 0 : assert!(!updates.is_empty());
2234 0 : TenantCreateOrUpdate::Update(updates)
2235 : }
2236 0 : }
2237 :
2238 : /// This API is used by the cloud control plane to migrate unsharded tenants that it created
2239 : /// directly with pageservers into this service.
2240 : ///
2241 : /// Cloud control plane MUST NOT continue issuing GENERATION NUMBERS for this tenant once it
2242 : /// has attempted to call this API. Failure to oblige to this rule may lead to S3 corruption.
2243 : /// Think of the first attempt to call this API as a transfer of absolute authority over the
2244 : /// tenant's source of generation numbers.
2245 : ///
2246 : /// The mode in this request coarse-grained control of tenants:
2247 : /// - Call with mode Attached* to upsert the tenant.
2248 : /// - Call with mode Secondary to either onboard a tenant without attaching it, or
2249 : /// to set an existing tenant to PolicyMode::Secondary
2250 : /// - Call with mode Detached to switch to PolicyMode::Detached
2251 0 : pub(crate) async fn tenant_location_config(
2252 0 : &self,
2253 0 : tenant_shard_id: TenantShardId,
2254 0 : req: TenantLocationConfigRequest,
2255 0 : ) -> Result<TenantLocationConfigResponse, ApiError> {
2256 : // We require an exclusive lock, because we are updating both persistent and in-memory state
2257 0 : let _tenant_lock = trace_exclusive_lock(
2258 0 : &self.tenant_op_locks,
2259 0 : tenant_shard_id.tenant_id,
2260 0 : TenantOperations::LocationConfig,
2261 0 : )
2262 0 : .await;
2263 :
2264 0 : if !tenant_shard_id.is_unsharded() {
2265 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
2266 0 : "This API is for importing single-sharded or unsharded tenants"
2267 0 : )));
2268 0 : }
2269 0 :
2270 0 : // First check if this is a creation or an update
2271 0 : let create_or_update = self.tenant_location_config_prepare(tenant_shard_id.tenant_id, req);
2272 0 :
2273 0 : let mut result = TenantLocationConfigResponse {
2274 0 : shards: Vec::new(),
2275 0 : stripe_size: None,
2276 0 : };
2277 0 : let waiters = match create_or_update {
2278 0 : TenantCreateOrUpdate::Create(create_req) => {
2279 0 : let (create_resp, waiters) = self.do_tenant_create(create_req).await?;
2280 0 : result.shards = create_resp
2281 0 : .shards
2282 0 : .into_iter()
2283 0 : .map(|s| TenantShardLocation {
2284 0 : node_id: s.node_id,
2285 0 : shard_id: s.shard_id,
2286 0 : })
2287 0 : .collect();
2288 0 : waiters
2289 : }
2290 0 : TenantCreateOrUpdate::Update(updates) => {
2291 0 : // Persist updates
2292 0 : // Ordering: write to the database before applying changes in-memory, so that
2293 0 : // we will not appear time-travel backwards on a restart.
2294 0 : let mut schedule_context = ScheduleContext::default();
2295 : for ShardUpdate {
2296 0 : tenant_shard_id,
2297 0 : placement_policy,
2298 0 : tenant_config,
2299 0 : generation,
2300 0 : } in &updates
2301 : {
2302 0 : self.persistence
2303 0 : .update_tenant_shard(
2304 0 : TenantFilter::Shard(*tenant_shard_id),
2305 0 : Some(placement_policy.clone()),
2306 0 : Some(tenant_config.clone()),
2307 0 : *generation,
2308 0 : None,
2309 0 : )
2310 0 : .await?;
2311 : }
2312 :
2313 : // Apply updates in-memory
2314 0 : let mut waiters = Vec::new();
2315 0 : {
2316 0 : let mut locked = self.inner.write().unwrap();
2317 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
2318 :
2319 : for ShardUpdate {
2320 0 : tenant_shard_id,
2321 0 : placement_policy,
2322 0 : tenant_config,
2323 0 : generation: update_generation,
2324 0 : } in updates
2325 : {
2326 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
2327 0 : tracing::warn!("Shard {tenant_shard_id} removed while updating");
2328 0 : continue;
2329 : };
2330 :
2331 : // Update stripe size
2332 0 : if result.stripe_size.is_none() && shard.shard.count.count() > 1 {
2333 0 : result.stripe_size = Some(shard.shard.stripe_size);
2334 0 : }
2335 :
2336 0 : shard.policy = placement_policy;
2337 0 : shard.config = tenant_config;
2338 0 : if let Some(generation) = update_generation {
2339 0 : shard.generation = Some(generation);
2340 0 : }
2341 :
2342 0 : shard.schedule(scheduler, &mut schedule_context)?;
2343 :
2344 0 : let maybe_waiter = self.maybe_reconcile_shard(shard, nodes);
2345 0 : if let Some(waiter) = maybe_waiter {
2346 0 : waiters.push(waiter);
2347 0 : }
2348 :
2349 0 : if let Some(node_id) = shard.intent.get_attached() {
2350 0 : result.shards.push(TenantShardLocation {
2351 0 : shard_id: tenant_shard_id,
2352 0 : node_id: *node_id,
2353 0 : })
2354 0 : }
2355 : }
2356 : }
2357 0 : waiters
2358 : }
2359 : };
2360 :
2361 0 : if let Err(e) = self.await_waiters(waiters, SHORT_RECONCILE_TIMEOUT).await {
2362 : // Do not treat a reconcile error as fatal: we have already applied any requested
2363 : // Intent changes, and the reconcile can fail for external reasons like unavailable
2364 : // compute notification API. In these cases, it is important that we do not
2365 : // cause the cloud control plane to retry forever on this API.
2366 0 : tracing::warn!(
2367 0 : "Failed to reconcile after /location_config: {e}, returning success anyway"
2368 : );
2369 0 : }
2370 :
2371 : // Logging the full result is useful because it lets us cross-check what the cloud control
2372 : // plane's tenant_shards table should contain.
2373 0 : tracing::info!("Complete, returning {result:?}");
2374 :
2375 0 : Ok(result)
2376 0 : }
2377 :
2378 0 : pub(crate) async fn tenant_config_set(&self, req: TenantConfigRequest) -> Result<(), ApiError> {
2379 : // We require an exclusive lock, because we are updating persistent and in-memory state
2380 0 : let _tenant_lock = trace_exclusive_lock(
2381 0 : &self.tenant_op_locks,
2382 0 : req.tenant_id,
2383 0 : TenantOperations::ConfigSet,
2384 0 : )
2385 0 : .await;
2386 :
2387 0 : let tenant_id = req.tenant_id;
2388 0 : let config = req.config;
2389 0 :
2390 0 : self.persistence
2391 0 : .update_tenant_shard(
2392 0 : TenantFilter::Tenant(req.tenant_id),
2393 0 : None,
2394 0 : Some(config.clone()),
2395 0 : None,
2396 0 : None,
2397 0 : )
2398 0 : .await?;
2399 :
2400 0 : let waiters = {
2401 0 : let mut waiters = Vec::new();
2402 0 : let mut locked = self.inner.write().unwrap();
2403 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
2404 0 : for (_shard_id, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
2405 0 : shard.config = config.clone();
2406 0 : if let Some(waiter) = self.maybe_reconcile_shard(shard, nodes) {
2407 0 : waiters.push(waiter);
2408 0 : }
2409 : }
2410 0 : waiters
2411 : };
2412 :
2413 0 : if let Err(e) = self.await_waiters(waiters, SHORT_RECONCILE_TIMEOUT).await {
2414 : // Treat this as success because we have stored the configuration. If e.g.
2415 : // a node was unavailable at this time, it should not stop us accepting a
2416 : // configuration change.
2417 0 : tracing::warn!(%tenant_id, "Accepted configuration update but reconciliation failed: {e}");
2418 0 : }
2419 :
2420 0 : Ok(())
2421 0 : }
2422 :
2423 0 : pub(crate) fn tenant_config_get(
2424 0 : &self,
2425 0 : tenant_id: TenantId,
2426 0 : ) -> Result<HashMap<&str, serde_json::Value>, ApiError> {
2427 0 : let config = {
2428 0 : let locked = self.inner.read().unwrap();
2429 0 :
2430 0 : match locked
2431 0 : .tenants
2432 0 : .range(TenantShardId::tenant_range(tenant_id))
2433 0 : .next()
2434 : {
2435 0 : Some((_tenant_shard_id, shard)) => shard.config.clone(),
2436 : None => {
2437 0 : return Err(ApiError::NotFound(
2438 0 : anyhow::anyhow!("Tenant not found").into(),
2439 0 : ))
2440 : }
2441 : }
2442 : };
2443 :
2444 : // Unlike the pageserver, we do not have a set of global defaults: the config is
2445 : // entirely per-tenant. Therefore the distinction between `tenant_specific_overrides`
2446 : // and `effective_config` in the response is meaningless, but we retain that syntax
2447 : // in order to remain compatible with the pageserver API.
2448 :
2449 0 : let response = HashMap::from([
2450 : (
2451 : "tenant_specific_overrides",
2452 0 : serde_json::to_value(&config)
2453 0 : .context("serializing tenant specific overrides")
2454 0 : .map_err(ApiError::InternalServerError)?,
2455 : ),
2456 : (
2457 0 : "effective_config",
2458 0 : serde_json::to_value(&config)
2459 0 : .context("serializing effective config")
2460 0 : .map_err(ApiError::InternalServerError)?,
2461 : ),
2462 : ]);
2463 :
2464 0 : Ok(response)
2465 0 : }
2466 :
2467 0 : pub(crate) async fn tenant_time_travel_remote_storage(
2468 0 : &self,
2469 0 : time_travel_req: &TenantTimeTravelRequest,
2470 0 : tenant_id: TenantId,
2471 0 : timestamp: Cow<'_, str>,
2472 0 : done_if_after: Cow<'_, str>,
2473 0 : ) -> Result<(), ApiError> {
2474 0 : let _tenant_lock = trace_exclusive_lock(
2475 0 : &self.tenant_op_locks,
2476 0 : tenant_id,
2477 0 : TenantOperations::TimeTravelRemoteStorage,
2478 0 : )
2479 0 : .await;
2480 :
2481 0 : let node = {
2482 0 : let mut locked = self.inner.write().unwrap();
2483 : // Just a sanity check to prevent misuse: the API expects that the tenant is fully
2484 : // detached everywhere, and nothing writes to S3 storage. Here, we verify that,
2485 : // but only at the start of the process, so it's really just to prevent operator
2486 : // mistakes.
2487 0 : for (shard_id, shard) in locked.tenants.range(TenantShardId::tenant_range(tenant_id)) {
2488 0 : if shard.intent.get_attached().is_some() || !shard.intent.get_secondary().is_empty()
2489 : {
2490 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
2491 0 : "We want tenant to be attached in shard with tenant_shard_id={shard_id}"
2492 0 : )));
2493 0 : }
2494 0 : let maybe_attached = shard
2495 0 : .observed
2496 0 : .locations
2497 0 : .iter()
2498 0 : .filter_map(|(node_id, observed_location)| {
2499 0 : observed_location
2500 0 : .conf
2501 0 : .as_ref()
2502 0 : .map(|loc| (node_id, observed_location, loc.mode))
2503 0 : })
2504 0 : .find(|(_, _, mode)| *mode != LocationConfigMode::Detached);
2505 0 : if let Some((node_id, _observed_location, mode)) = maybe_attached {
2506 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!("We observed attached={mode:?} tenant in node_id={node_id} shard with tenant_shard_id={shard_id}")));
2507 0 : }
2508 : }
2509 0 : let scheduler = &mut locked.scheduler;
2510 : // Right now we only perform the operation on a single node without parallelization
2511 : // TODO fan out the operation to multiple nodes for better performance
2512 0 : let node_id = scheduler.schedule_shard(&[], &ScheduleContext::default())?;
2513 0 : let node = locked
2514 0 : .nodes
2515 0 : .get(&node_id)
2516 0 : .expect("Pageservers may not be deleted while lock is active");
2517 0 : node.clone()
2518 0 : };
2519 0 :
2520 0 : // The shard count is encoded in the remote storage's URL, so we need to handle all historically used shard counts
2521 0 : let mut counts = time_travel_req
2522 0 : .shard_counts
2523 0 : .iter()
2524 0 : .copied()
2525 0 : .collect::<HashSet<_>>()
2526 0 : .into_iter()
2527 0 : .collect::<Vec<_>>();
2528 0 : counts.sort_unstable();
2529 :
2530 0 : for count in counts {
2531 0 : let shard_ids = (0..count.count())
2532 0 : .map(|i| TenantShardId {
2533 0 : tenant_id,
2534 0 : shard_number: ShardNumber(i),
2535 0 : shard_count: count,
2536 0 : })
2537 0 : .collect::<Vec<_>>();
2538 0 : for tenant_shard_id in shard_ids {
2539 0 : let client = PageserverClient::new(
2540 0 : node.get_id(),
2541 0 : node.base_url(),
2542 0 : self.config.jwt_token.as_deref(),
2543 0 : );
2544 0 :
2545 0 : tracing::info!("Doing time travel recovery for shard {tenant_shard_id}",);
2546 :
2547 0 : client
2548 0 : .tenant_time_travel_remote_storage(
2549 0 : tenant_shard_id,
2550 0 : ×tamp,
2551 0 : &done_if_after,
2552 0 : )
2553 0 : .await
2554 0 : .map_err(|e| {
2555 0 : ApiError::InternalServerError(anyhow::anyhow!(
2556 0 : "Error doing time travel recovery for shard {tenant_shard_id} on node {}: {e}",
2557 0 : node
2558 0 : ))
2559 0 : })?;
2560 : }
2561 : }
2562 0 : Ok(())
2563 0 : }
2564 :
2565 0 : pub(crate) async fn tenant_secondary_download(
2566 0 : &self,
2567 0 : tenant_id: TenantId,
2568 0 : wait: Option<Duration>,
2569 0 : ) -> Result<(StatusCode, SecondaryProgress), ApiError> {
2570 0 : let _tenant_lock = trace_shared_lock(
2571 0 : &self.tenant_op_locks,
2572 0 : tenant_id,
2573 0 : TenantOperations::SecondaryDownload,
2574 0 : )
2575 0 : .await;
2576 :
2577 : // Acquire lock and yield the collection of shard-node tuples which we will send requests onward to
2578 0 : let targets = {
2579 0 : let locked = self.inner.read().unwrap();
2580 0 : let mut targets = Vec::new();
2581 :
2582 0 : for (tenant_shard_id, shard) in
2583 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
2584 : {
2585 0 : for node_id in shard.intent.get_secondary() {
2586 0 : let node = locked
2587 0 : .nodes
2588 0 : .get(node_id)
2589 0 : .expect("Pageservers may not be deleted while referenced");
2590 0 :
2591 0 : targets.push((*tenant_shard_id, node.clone()));
2592 0 : }
2593 : }
2594 0 : targets
2595 0 : };
2596 0 :
2597 0 : // Issue concurrent requests to all shards' locations
2598 0 : let mut futs = FuturesUnordered::new();
2599 0 : for (tenant_shard_id, node) in targets {
2600 0 : let client = PageserverClient::new(
2601 0 : node.get_id(),
2602 0 : node.base_url(),
2603 0 : self.config.jwt_token.as_deref(),
2604 0 : );
2605 0 : futs.push(async move {
2606 0 : let result = client
2607 0 : .tenant_secondary_download(tenant_shard_id, wait)
2608 0 : .await;
2609 0 : (result, node, tenant_shard_id)
2610 0 : })
2611 : }
2612 :
2613 : // Handle any errors returned by pageservers. This includes cases like this request racing with
2614 : // a scheduling operation, such that the tenant shard we're calling doesn't exist on that pageserver any more, as
2615 : // well as more general cases like 503s, 500s, or timeouts.
2616 0 : let mut aggregate_progress = SecondaryProgress::default();
2617 0 : let mut aggregate_status: Option<StatusCode> = None;
2618 0 : let mut error: Option<mgmt_api::Error> = None;
2619 0 : while let Some((result, node, tenant_shard_id)) = futs.next().await {
2620 0 : match result {
2621 0 : Err(e) => {
2622 0 : // Secondary downloads are always advisory: if something fails, we nevertheless report success, so that whoever
2623 0 : // is calling us will proceed with whatever migration they're doing, albeit with a slightly less warm cache
2624 0 : // than they had hoped for.
2625 0 : tracing::warn!("Secondary download error from pageserver {node}: {e}",);
2626 0 : error = Some(e)
2627 : }
2628 0 : Ok((status_code, progress)) => {
2629 0 : tracing::info!(%tenant_shard_id, "Shard status={status_code} progress: {progress:?}");
2630 0 : aggregate_progress.layers_downloaded += progress.layers_downloaded;
2631 0 : aggregate_progress.layers_total += progress.layers_total;
2632 0 : aggregate_progress.bytes_downloaded += progress.bytes_downloaded;
2633 0 : aggregate_progress.bytes_total += progress.bytes_total;
2634 0 : aggregate_progress.heatmap_mtime =
2635 0 : std::cmp::max(aggregate_progress.heatmap_mtime, progress.heatmap_mtime);
2636 0 : aggregate_status = match aggregate_status {
2637 0 : None => Some(status_code),
2638 0 : Some(StatusCode::OK) => Some(status_code),
2639 0 : Some(cur) => {
2640 0 : // Other status codes (e.g. 202) -- do not overwrite.
2641 0 : Some(cur)
2642 : }
2643 : };
2644 : }
2645 : }
2646 : }
2647 :
2648 : // If any of the shards return 202, indicate our result as 202.
2649 0 : match aggregate_status {
2650 : None => {
2651 0 : match error {
2652 0 : Some(e) => {
2653 0 : // No successes, and an error: surface it
2654 0 : Err(ApiError::Conflict(format!("Error from pageserver: {e}")))
2655 : }
2656 : None => {
2657 : // No shards found
2658 0 : Err(ApiError::NotFound(
2659 0 : anyhow::anyhow!("Tenant {} not found", tenant_id).into(),
2660 0 : ))
2661 : }
2662 : }
2663 : }
2664 0 : Some(aggregate_status) => Ok((aggregate_status, aggregate_progress)),
2665 : }
2666 0 : }
2667 :
2668 0 : pub(crate) async fn tenant_delete(&self, tenant_id: TenantId) -> Result<StatusCode, ApiError> {
2669 0 : let _tenant_lock =
2670 0 : trace_exclusive_lock(&self.tenant_op_locks, tenant_id, TenantOperations::Delete).await;
2671 :
2672 : // Detach all shards
2673 0 : let (detach_waiters, shard_ids, node) = {
2674 0 : let mut shard_ids = Vec::new();
2675 0 : let mut detach_waiters = Vec::new();
2676 0 : let mut locked = self.inner.write().unwrap();
2677 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
2678 0 : for (tenant_shard_id, shard) in
2679 0 : tenants.range_mut(TenantShardId::tenant_range(tenant_id))
2680 : {
2681 0 : shard_ids.push(*tenant_shard_id);
2682 0 :
2683 0 : // Update the tenant's intent to remove all attachments
2684 0 : shard.policy = PlacementPolicy::Detached;
2685 0 : shard
2686 0 : .schedule(scheduler, &mut ScheduleContext::default())
2687 0 : .expect("De-scheduling is infallible");
2688 0 : debug_assert!(shard.intent.get_attached().is_none());
2689 0 : debug_assert!(shard.intent.get_secondary().is_empty());
2690 :
2691 0 : if let Some(waiter) = self.maybe_reconcile_shard(shard, nodes) {
2692 0 : detach_waiters.push(waiter);
2693 0 : }
2694 : }
2695 :
2696 : // Pick an arbitrary node to use for remote deletions (does not have to be where the tenant
2697 : // was attached, just has to be able to see the S3 content)
2698 0 : let node_id = scheduler.schedule_shard(&[], &ScheduleContext::default())?;
2699 0 : let node = nodes
2700 0 : .get(&node_id)
2701 0 : .expect("Pageservers may not be deleted while lock is active");
2702 0 : (detach_waiters, shard_ids, node.clone())
2703 0 : };
2704 0 :
2705 0 : // This reconcile wait can fail in a few ways:
2706 0 : // A there is a very long queue for the reconciler semaphore
2707 0 : // B some pageserver is failing to handle a detach promptly
2708 0 : // C some pageserver goes offline right at the moment we send it a request.
2709 0 : //
2710 0 : // A and C are transient: the semaphore will eventually become available, and once a node is marked offline
2711 0 : // the next attempt to reconcile will silently skip detaches for an offline node and succeed. If B happens,
2712 0 : // it's a bug, and needs resolving at the pageserver level (we shouldn't just leave attachments behind while
2713 0 : // deleting the underlying data).
2714 0 : self.await_waiters(detach_waiters, RECONCILE_TIMEOUT)
2715 0 : .await?;
2716 :
2717 0 : let locations = shard_ids
2718 0 : .into_iter()
2719 0 : .map(|s| (s, node.clone()))
2720 0 : .collect::<Vec<_>>();
2721 0 : let results = self.tenant_for_shards_api(
2722 0 : locations,
2723 0 : |tenant_shard_id, client| async move { client.tenant_delete(tenant_shard_id).await },
2724 0 : 1,
2725 0 : 3,
2726 0 : RECONCILE_TIMEOUT,
2727 0 : &self.cancel,
2728 0 : )
2729 0 : .await;
2730 0 : for result in results {
2731 0 : match result {
2732 : Ok(StatusCode::ACCEPTED) => {
2733 : // This should never happen: we waited for detaches to finish above
2734 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
2735 0 : "Unexpectedly still attached on {}",
2736 0 : node
2737 0 : )));
2738 : }
2739 0 : Ok(_) => {}
2740 : Err(mgmt_api::Error::Cancelled) => {
2741 0 : return Err(ApiError::ShuttingDown);
2742 : }
2743 0 : Err(e) => {
2744 0 : // This is unexpected: remote deletion should be infallible, unless the object store
2745 0 : // at large is unavailable.
2746 0 : tracing::error!("Error deleting via node {}: {e}", node);
2747 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(e)));
2748 : }
2749 : }
2750 : }
2751 :
2752 : // Fall through: deletion of the tenant on pageservers is complete, we may proceed to drop
2753 : // our in-memory state and database state.
2754 :
2755 : // Ordering: we delete persistent state first: if we then
2756 : // crash, we will drop the in-memory state.
2757 :
2758 : // Drop persistent state.
2759 0 : self.persistence.delete_tenant(tenant_id).await?;
2760 :
2761 : // Drop in-memory state
2762 : {
2763 0 : let mut locked = self.inner.write().unwrap();
2764 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
2765 :
2766 : // Dereference Scheduler from shards before dropping them
2767 0 : for (_tenant_shard_id, shard) in
2768 0 : tenants.range_mut(TenantShardId::tenant_range(tenant_id))
2769 0 : {
2770 0 : shard.intent.clear(scheduler);
2771 0 : }
2772 :
2773 0 : tenants.retain(|tenant_shard_id, _shard| tenant_shard_id.tenant_id != tenant_id);
2774 0 : tracing::info!(
2775 0 : "Deleted tenant {tenant_id}, now have {} tenants",
2776 0 : locked.tenants.len()
2777 : );
2778 : };
2779 :
2780 : // Success is represented as 404, to imitate the existing pageserver deletion API
2781 0 : Ok(StatusCode::NOT_FOUND)
2782 0 : }
2783 :
2784 : /// Naming: this configures the storage controller's policies for a tenant, whereas [`Self::tenant_config_set`] is "set the TenantConfig"
2785 : /// for a tenant. The TenantConfig is passed through to pageservers, whereas this function modifies
2786 : /// the tenant's policies (configuration) within the storage controller
2787 0 : pub(crate) async fn tenant_update_policy(
2788 0 : &self,
2789 0 : tenant_id: TenantId,
2790 0 : req: TenantPolicyRequest,
2791 0 : ) -> Result<(), ApiError> {
2792 : // We require an exclusive lock, because we are updating persistent and in-memory state
2793 0 : let _tenant_lock = trace_exclusive_lock(
2794 0 : &self.tenant_op_locks,
2795 0 : tenant_id,
2796 0 : TenantOperations::UpdatePolicy,
2797 0 : )
2798 0 : .await;
2799 :
2800 0 : failpoint_support::sleep_millis_async!("tenant-update-policy-exclusive-lock");
2801 :
2802 : let TenantPolicyRequest {
2803 0 : placement,
2804 0 : scheduling,
2805 0 : } = req;
2806 0 :
2807 0 : self.persistence
2808 0 : .update_tenant_shard(
2809 0 : TenantFilter::Tenant(tenant_id),
2810 0 : placement.clone(),
2811 0 : None,
2812 0 : None,
2813 0 : scheduling,
2814 0 : )
2815 0 : .await?;
2816 :
2817 0 : let mut schedule_context = ScheduleContext::default();
2818 0 : let mut locked = self.inner.write().unwrap();
2819 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
2820 0 : for (shard_id, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
2821 0 : if let Some(placement) = &placement {
2822 0 : shard.policy = placement.clone();
2823 0 :
2824 0 : tracing::info!(tenant_id=%shard_id.tenant_id, shard_id=%shard_id.shard_slug(),
2825 0 : "Updated placement policy to {placement:?}");
2826 0 : }
2827 :
2828 0 : if let Some(scheduling) = &scheduling {
2829 0 : shard.set_scheduling_policy(*scheduling);
2830 0 :
2831 0 : tracing::info!(tenant_id=%shard_id.tenant_id, shard_id=%shard_id.shard_slug(),
2832 0 : "Updated scheduling policy to {scheduling:?}");
2833 0 : }
2834 :
2835 : // In case scheduling is being switched back on, try it now.
2836 0 : shard.schedule(scheduler, &mut schedule_context).ok();
2837 0 : self.maybe_reconcile_shard(shard, nodes);
2838 : }
2839 :
2840 0 : Ok(())
2841 0 : }
2842 :
2843 0 : pub(crate) async fn tenant_timeline_create(
2844 0 : &self,
2845 0 : tenant_id: TenantId,
2846 0 : mut create_req: TimelineCreateRequest,
2847 0 : ) -> Result<TimelineInfo, ApiError> {
2848 0 : tracing::info!(
2849 0 : "Creating timeline {}/{}",
2850 : tenant_id,
2851 : create_req.new_timeline_id,
2852 : );
2853 :
2854 0 : let _tenant_lock = trace_shared_lock(
2855 0 : &self.tenant_op_locks,
2856 0 : tenant_id,
2857 0 : TenantOperations::TimelineCreate,
2858 0 : )
2859 0 : .await;
2860 0 : failpoint_support::sleep_millis_async!("tenant-create-timeline-shared-lock");
2861 :
2862 0 : self.tenant_remote_mutation(tenant_id, move |mut targets| async move {
2863 0 : if targets.is_empty() {
2864 0 : return Err(ApiError::NotFound(
2865 0 : anyhow::anyhow!("Tenant not found").into(),
2866 0 : ));
2867 0 : };
2868 0 : let shard_zero = targets.remove(0);
2869 0 :
2870 0 : async fn create_one(
2871 0 : tenant_shard_id: TenantShardId,
2872 0 : node: Node,
2873 0 : jwt: Option<String>,
2874 0 : create_req: TimelineCreateRequest,
2875 0 : ) -> Result<TimelineInfo, ApiError> {
2876 0 : tracing::info!(
2877 0 : "Creating timeline on shard {}/{}, attached to node {node}",
2878 0 : tenant_shard_id,
2879 0 : create_req.new_timeline_id,
2880 0 : );
2881 0 : let client = PageserverClient::new(node.get_id(), node.base_url(), jwt.as_deref());
2882 0 :
2883 0 : client
2884 0 : .timeline_create(tenant_shard_id, &create_req)
2885 0 : .await
2886 0 : .map_err(|e| passthrough_api_error(&node, e))
2887 0 : }
2888 0 :
2889 0 : // Because the caller might not provide an explicit LSN, we must do the creation first on a single shard, and then
2890 0 : // use whatever LSN that shard picked when creating on subsequent shards. We arbitrarily use shard zero as the shard
2891 0 : // that will get the first creation request, and propagate the LSN to all the >0 shards.
2892 0 : let timeline_info = create_one(
2893 0 : shard_zero.0,
2894 0 : shard_zero.1,
2895 0 : self.config.jwt_token.clone(),
2896 0 : create_req.clone(),
2897 0 : )
2898 0 : .await?;
2899 0 :
2900 0 : // Propagate the LSN that shard zero picked, if caller didn't provide one
2901 0 : if create_req.ancestor_timeline_id.is_some() && create_req.ancestor_start_lsn.is_none()
2902 0 : {
2903 0 : create_req.ancestor_start_lsn = timeline_info.ancestor_lsn;
2904 0 : }
2905 0 :
2906 0 : // Create timeline on remaining shards with number >0
2907 0 : if !targets.is_empty() {
2908 0 : // If we had multiple shards, issue requests for the remainder now.
2909 0 : let jwt = &self.config.jwt_token;
2910 0 : self.tenant_for_shards(
2911 0 : targets.iter().map(|t| (t.0, t.1.clone())).collect(),
2912 0 : |tenant_shard_id: TenantShardId, node: Node| {
2913 0 : let create_req = create_req.clone();
2914 0 : Box::pin(create_one(tenant_shard_id, node, jwt.clone(), create_req))
2915 0 : },
2916 0 : )
2917 0 : .await?;
2918 0 : }
2919 0 :
2920 0 : Ok(timeline_info)
2921 0 : })
2922 0 : .await?
2923 0 : }
2924 :
2925 0 : pub(crate) async fn tenant_timeline_archival_config(
2926 0 : &self,
2927 0 : tenant_id: TenantId,
2928 0 : timeline_id: TimelineId,
2929 0 : req: TimelineArchivalConfigRequest,
2930 0 : ) -> Result<(), ApiError> {
2931 0 : tracing::info!(
2932 0 : "Setting archival config of timeline {tenant_id}/{timeline_id} to '{:?}'",
2933 : req.state
2934 : );
2935 :
2936 0 : let _tenant_lock = trace_shared_lock(
2937 0 : &self.tenant_op_locks,
2938 0 : tenant_id,
2939 0 : TenantOperations::TimelineArchivalConfig,
2940 0 : )
2941 0 : .await;
2942 :
2943 0 : self.tenant_remote_mutation(tenant_id, move |targets| async move {
2944 0 : if targets.is_empty() {
2945 0 : return Err(ApiError::NotFound(
2946 0 : anyhow::anyhow!("Tenant not found").into(),
2947 0 : ));
2948 0 : }
2949 0 : async fn config_one(
2950 0 : tenant_shard_id: TenantShardId,
2951 0 : timeline_id: TimelineId,
2952 0 : node: Node,
2953 0 : jwt: Option<String>,
2954 0 : req: TimelineArchivalConfigRequest,
2955 0 : ) -> Result<(), ApiError> {
2956 0 : tracing::info!(
2957 0 : "Setting archival config of timeline on shard {tenant_shard_id}/{timeline_id}, attached to node {node}",
2958 0 : );
2959 0 :
2960 0 : let client = PageserverClient::new(node.get_id(), node.base_url(), jwt.as_deref());
2961 0 :
2962 0 : client
2963 0 : .timeline_archival_config(tenant_shard_id, timeline_id, &req)
2964 0 : .await
2965 0 : .map_err(|e| match e {
2966 0 : mgmt_api::Error::ApiError(StatusCode::PRECONDITION_FAILED, msg) => {
2967 0 : ApiError::PreconditionFailed(msg.into_boxed_str())
2968 0 : }
2969 0 : _ => passthrough_api_error(&node, e),
2970 0 : })
2971 0 : }
2972 0 :
2973 0 : // no shard needs to go first/last; the operation should be idempotent
2974 0 : // TODO: it would be great to ensure that all shards return the same error
2975 0 : let results = self
2976 0 : .tenant_for_shards(targets, |tenant_shard_id, node| {
2977 0 : futures::FutureExt::boxed(config_one(
2978 0 : tenant_shard_id,
2979 0 : timeline_id,
2980 0 : node,
2981 0 : self.config.jwt_token.clone(),
2982 0 : req.clone(),
2983 0 : ))
2984 0 : })
2985 0 : .await?;
2986 0 : assert!(!results.is_empty(), "must have at least one result");
2987 0 :
2988 0 : Ok(())
2989 0 : }).await?
2990 0 : }
2991 :
2992 0 : pub(crate) async fn tenant_timeline_detach_ancestor(
2993 0 : &self,
2994 0 : tenant_id: TenantId,
2995 0 : timeline_id: TimelineId,
2996 0 : ) -> Result<models::detach_ancestor::AncestorDetached, ApiError> {
2997 0 : tracing::info!("Detaching timeline {tenant_id}/{timeline_id}",);
2998 :
2999 0 : let _tenant_lock = trace_shared_lock(
3000 0 : &self.tenant_op_locks,
3001 0 : tenant_id,
3002 0 : TenantOperations::TimelineDetachAncestor,
3003 0 : )
3004 0 : .await;
3005 :
3006 0 : self.tenant_remote_mutation(tenant_id, move |targets| async move {
3007 0 : if targets.is_empty() {
3008 0 : return Err(ApiError::NotFound(
3009 0 : anyhow::anyhow!("Tenant not found").into(),
3010 0 : ));
3011 0 : }
3012 0 :
3013 0 : async fn detach_one(
3014 0 : tenant_shard_id: TenantShardId,
3015 0 : timeline_id: TimelineId,
3016 0 : node: Node,
3017 0 : jwt: Option<String>,
3018 0 : ) -> Result<(ShardNumber, models::detach_ancestor::AncestorDetached), ApiError> {
3019 0 : tracing::info!(
3020 0 : "Detaching timeline on shard {tenant_shard_id}/{timeline_id}, attached to node {node}",
3021 0 : );
3022 0 :
3023 0 : let client = PageserverClient::new(node.get_id(), node.base_url(), jwt.as_deref());
3024 0 :
3025 0 : client
3026 0 : .timeline_detach_ancestor(tenant_shard_id, timeline_id)
3027 0 : .await
3028 0 : .map_err(|e| {
3029 0 : use mgmt_api::Error;
3030 0 :
3031 0 : match e {
3032 0 : // no ancestor (ever)
3033 0 : Error::ApiError(StatusCode::CONFLICT, msg) => ApiError::Conflict(format!(
3034 0 : "{node}: {}",
3035 0 : msg.strip_prefix("Conflict: ").unwrap_or(&msg)
3036 0 : )),
3037 0 : // too many ancestors
3038 0 : Error::ApiError(StatusCode::BAD_REQUEST, msg) => {
3039 0 : ApiError::BadRequest(anyhow::anyhow!("{node}: {msg}"))
3040 0 : }
3041 0 : Error::ApiError(StatusCode::INTERNAL_SERVER_ERROR, msg) => {
3042 0 : // avoid turning these into conflicts to remain compatible with
3043 0 : // pageservers, 500 errors are sadly retryable with timeline ancestor
3044 0 : // detach
3045 0 : ApiError::InternalServerError(anyhow::anyhow!("{node}: {msg}"))
3046 0 : }
3047 0 : // rest can be mapped as usual
3048 0 : other => passthrough_api_error(&node, other),
3049 0 : }
3050 0 : })
3051 0 : .map(|res| (tenant_shard_id.shard_number, res))
3052 0 : }
3053 0 :
3054 0 : // no shard needs to go first/last; the operation should be idempotent
3055 0 : let mut results = self
3056 0 : .tenant_for_shards(targets, |tenant_shard_id, node| {
3057 0 : futures::FutureExt::boxed(detach_one(
3058 0 : tenant_shard_id,
3059 0 : timeline_id,
3060 0 : node,
3061 0 : self.config.jwt_token.clone(),
3062 0 : ))
3063 0 : })
3064 0 : .await?;
3065 0 :
3066 0 : let any = results.pop().expect("we must have at least one response");
3067 0 :
3068 0 : let mismatching = results
3069 0 : .iter()
3070 0 : .filter(|(_, res)| res != &any.1)
3071 0 : .collect::<Vec<_>>();
3072 0 : if !mismatching.is_empty() {
3073 0 : // this can be hit by races which should not happen because operation lock on cplane
3074 0 : let matching = results.len() - mismatching.len();
3075 0 : tracing::error!(
3076 0 : matching,
3077 0 : compared_against=?any,
3078 0 : ?mismatching,
3079 0 : "shards returned different results"
3080 0 : );
3081 0 :
3082 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!("pageservers returned mixed results for ancestor detach; manual intervention is required.")));
3083 0 : }
3084 0 :
3085 0 : Ok(any.1)
3086 0 : }).await?
3087 0 : }
3088 :
3089 : /// Helper for concurrently calling a pageserver API on a number of shards, such as timeline creation.
3090 : ///
3091 : /// On success, the returned vector contains exactly the same number of elements as the input `locations`.
3092 0 : async fn tenant_for_shards<F, R>(
3093 0 : &self,
3094 0 : locations: Vec<(TenantShardId, Node)>,
3095 0 : mut req_fn: F,
3096 0 : ) -> Result<Vec<R>, ApiError>
3097 0 : where
3098 0 : F: FnMut(
3099 0 : TenantShardId,
3100 0 : Node,
3101 0 : )
3102 0 : -> std::pin::Pin<Box<dyn futures::Future<Output = Result<R, ApiError>> + Send>>,
3103 0 : {
3104 0 : let mut futs = FuturesUnordered::new();
3105 0 : let mut results = Vec::with_capacity(locations.len());
3106 :
3107 0 : for (tenant_shard_id, node) in locations {
3108 0 : futs.push(req_fn(tenant_shard_id, node));
3109 0 : }
3110 :
3111 0 : while let Some(r) = futs.next().await {
3112 0 : results.push(r?);
3113 : }
3114 :
3115 0 : Ok(results)
3116 0 : }
3117 :
3118 : /// Concurrently invoke a pageserver API call on many shards at once
3119 0 : pub(crate) async fn tenant_for_shards_api<T, O, F>(
3120 0 : &self,
3121 0 : locations: Vec<(TenantShardId, Node)>,
3122 0 : op: O,
3123 0 : warn_threshold: u32,
3124 0 : max_retries: u32,
3125 0 : timeout: Duration,
3126 0 : cancel: &CancellationToken,
3127 0 : ) -> Vec<mgmt_api::Result<T>>
3128 0 : where
3129 0 : O: Fn(TenantShardId, PageserverClient) -> F + Copy,
3130 0 : F: std::future::Future<Output = mgmt_api::Result<T>>,
3131 0 : {
3132 0 : let mut futs = FuturesUnordered::new();
3133 0 : let mut results = Vec::with_capacity(locations.len());
3134 :
3135 0 : for (tenant_shard_id, node) in locations {
3136 0 : futs.push(async move {
3137 0 : node.with_client_retries(
3138 0 : |client| op(tenant_shard_id, client),
3139 0 : &self.config.jwt_token,
3140 0 : warn_threshold,
3141 0 : max_retries,
3142 0 : timeout,
3143 0 : cancel,
3144 0 : )
3145 0 : .await
3146 0 : });
3147 0 : }
3148 :
3149 0 : while let Some(r) = futs.next().await {
3150 0 : let r = r.unwrap_or(Err(mgmt_api::Error::Cancelled));
3151 0 : results.push(r);
3152 0 : }
3153 :
3154 0 : results
3155 0 : }
3156 :
3157 : /// Helper for safely working with the shards in a tenant remotely on pageservers, for example
3158 : /// when creating and deleting timelines:
3159 : /// - Makes sure shards are attached somewhere if they weren't already
3160 : /// - Looks up the shards and the nodes where they were most recently attached
3161 : /// - Guarantees that after the inner function returns, the shards' generations haven't moved on: this
3162 : /// ensures that the remote operation acted on the most recent generation, and is therefore durable.
3163 0 : async fn tenant_remote_mutation<R, O, F>(
3164 0 : &self,
3165 0 : tenant_id: TenantId,
3166 0 : op: O,
3167 0 : ) -> Result<R, ApiError>
3168 0 : where
3169 0 : O: FnOnce(Vec<(TenantShardId, Node)>) -> F,
3170 0 : F: std::future::Future<Output = R>,
3171 0 : {
3172 0 : let target_gens = {
3173 0 : let mut targets = Vec::new();
3174 :
3175 : // Load the currently attached pageservers for the latest generation of each shard. This can
3176 : // run concurrently with reconciliations, and it is not guaranteed that the node we find here
3177 : // will still be the latest when we're done: we will check generations again at the end of
3178 : // this function to handle that.
3179 0 : let generations = self.persistence.peek_generations(tenant_id).await?;
3180 :
3181 0 : if generations
3182 0 : .iter()
3183 0 : .any(|i| i.generation.is_none() || i.generation_pageserver.is_none())
3184 : {
3185 : // One or more shards has not been attached to a pageserver. Check if this is because it's configured
3186 : // to be detached (409: caller should give up), or because it's meant to be attached but isn't yet (503: caller should retry)
3187 0 : let locked = self.inner.read().unwrap();
3188 0 : for (shard_id, shard) in
3189 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
3190 : {
3191 0 : match shard.policy {
3192 0 : PlacementPolicy::Attached(_) => {
3193 0 : // This shard is meant to be attached: the caller is not wrong to try and
3194 0 : // use this function, but we can't service the request right now.
3195 0 : }
3196 : PlacementPolicy::Secondary | PlacementPolicy::Detached => {
3197 0 : return Err(ApiError::Conflict(format!(
3198 0 : "Shard {shard_id} tenant has policy {:?}",
3199 0 : shard.policy
3200 0 : )));
3201 : }
3202 : }
3203 : }
3204 :
3205 0 : return Err(ApiError::ResourceUnavailable(
3206 0 : "One or more shards in tenant is not yet attached".into(),
3207 0 : ));
3208 0 : }
3209 0 :
3210 0 : let locked = self.inner.read().unwrap();
3211 : for ShardGenerationState {
3212 0 : tenant_shard_id,
3213 0 : generation,
3214 0 : generation_pageserver,
3215 0 : } in generations
3216 : {
3217 0 : let node_id = generation_pageserver.expect("We checked for None above");
3218 0 : let node = locked
3219 0 : .nodes
3220 0 : .get(&node_id)
3221 0 : .ok_or(ApiError::Conflict(format!(
3222 0 : "Raced with removal of node {node_id}"
3223 0 : )))?;
3224 0 : targets.push((tenant_shard_id, node.clone(), generation));
3225 : }
3226 :
3227 0 : targets
3228 0 : };
3229 0 :
3230 0 : let targets = target_gens.iter().map(|t| (t.0, t.1.clone())).collect();
3231 0 : let result = op(targets).await;
3232 :
3233 : // Post-check: are all the generations of all the shards the same as they were initially? This proves that
3234 : // our remote operation executed on the latest generation and is therefore persistent.
3235 : {
3236 0 : let latest_generations = self.persistence.peek_generations(tenant_id).await?;
3237 0 : if latest_generations
3238 0 : .into_iter()
3239 0 : .map(
3240 0 : |ShardGenerationState {
3241 : tenant_shard_id,
3242 : generation,
3243 : generation_pageserver: _,
3244 0 : }| (tenant_shard_id, generation),
3245 0 : )
3246 0 : .collect::<Vec<_>>()
3247 0 : != target_gens
3248 0 : .into_iter()
3249 0 : .map(|i| (i.0, i.2))
3250 0 : .collect::<Vec<_>>()
3251 : {
3252 : // We raced with something that incremented the generation, and therefore cannot be
3253 : // confident that our actions are persistent (they might have hit an old generation).
3254 : //
3255 : // This is safe but requires a retry: ask the client to do that by giving them a 503 response.
3256 0 : return Err(ApiError::ResourceUnavailable(
3257 0 : "Tenant attachment changed, please retry".into(),
3258 0 : ));
3259 0 : }
3260 0 : }
3261 0 :
3262 0 : Ok(result)
3263 0 : }
3264 :
3265 0 : pub(crate) async fn tenant_timeline_delete(
3266 0 : &self,
3267 0 : tenant_id: TenantId,
3268 0 : timeline_id: TimelineId,
3269 0 : ) -> Result<StatusCode, ApiError> {
3270 0 : tracing::info!("Deleting timeline {}/{}", tenant_id, timeline_id,);
3271 0 : let _tenant_lock = trace_shared_lock(
3272 0 : &self.tenant_op_locks,
3273 0 : tenant_id,
3274 0 : TenantOperations::TimelineDelete,
3275 0 : )
3276 0 : .await;
3277 :
3278 0 : self.tenant_remote_mutation(tenant_id, move |mut targets| async move {
3279 0 : if targets.is_empty() {
3280 0 : return Err(ApiError::NotFound(
3281 0 : anyhow::anyhow!("Tenant not found").into(),
3282 0 : ));
3283 0 : }
3284 0 : let shard_zero = targets.remove(0);
3285 0 :
3286 0 : async fn delete_one(
3287 0 : tenant_shard_id: TenantShardId,
3288 0 : timeline_id: TimelineId,
3289 0 : node: Node,
3290 0 : jwt: Option<String>,
3291 0 : ) -> Result<StatusCode, ApiError> {
3292 0 : tracing::info!(
3293 0 : "Deleting timeline on shard {tenant_shard_id}/{timeline_id}, attached to node {node}",
3294 0 : );
3295 0 :
3296 0 : let client = PageserverClient::new(node.get_id(), node.base_url(), jwt.as_deref());
3297 0 : client
3298 0 : .timeline_delete(tenant_shard_id, timeline_id)
3299 0 : .await
3300 0 : .map_err(|e| {
3301 0 : ApiError::InternalServerError(anyhow::anyhow!(
3302 0 : "Error deleting timeline {timeline_id} on {tenant_shard_id} on node {node}: {e}",
3303 0 : ))
3304 0 : })
3305 0 : }
3306 0 :
3307 0 : let statuses = self
3308 0 : .tenant_for_shards(targets, |tenant_shard_id: TenantShardId, node: Node| {
3309 0 : Box::pin(delete_one(
3310 0 : tenant_shard_id,
3311 0 : timeline_id,
3312 0 : node,
3313 0 : self.config.jwt_token.clone(),
3314 0 : ))
3315 0 : })
3316 0 : .await?;
3317 0 :
3318 0 : // If any shards >0 haven't finished deletion yet, don't start deletion on shard zero
3319 0 : if statuses.iter().any(|s| s != &StatusCode::NOT_FOUND) {
3320 0 : return Ok(StatusCode::ACCEPTED);
3321 0 : }
3322 0 :
3323 0 : // Delete shard zero last: this is not strictly necessary, but since a caller's GET on a timeline will be routed
3324 0 : // to shard zero, it gives a more obvious behavior that a GET returns 404 once the deletion is done.
3325 0 : let shard_zero_status = delete_one(
3326 0 : shard_zero.0,
3327 0 : timeline_id,
3328 0 : shard_zero.1,
3329 0 : self.config.jwt_token.clone(),
3330 0 : )
3331 0 : .await?;
3332 0 : Ok(shard_zero_status)
3333 0 : }).await?
3334 0 : }
3335 :
3336 : /// When you need to send an HTTP request to the pageserver that holds shard0 of a tenant, this
3337 : /// function looks up and returns node. If the tenant isn't found, returns Err(ApiError::NotFound)
3338 0 : pub(crate) fn tenant_shard0_node(
3339 0 : &self,
3340 0 : tenant_id: TenantId,
3341 0 : ) -> Result<(Node, TenantShardId), ApiError> {
3342 0 : let locked = self.inner.read().unwrap();
3343 0 : let Some((tenant_shard_id, shard)) = locked
3344 0 : .tenants
3345 0 : .range(TenantShardId::tenant_range(tenant_id))
3346 0 : .next()
3347 : else {
3348 0 : return Err(ApiError::NotFound(
3349 0 : anyhow::anyhow!("Tenant {tenant_id} not found").into(),
3350 0 : ));
3351 : };
3352 :
3353 : // TODO: should use the ID last published to compute_hook, rather than the intent: the intent might
3354 : // point to somewhere we haven't attached yet.
3355 0 : let Some(node_id) = shard.intent.get_attached() else {
3356 0 : tracing::warn!(
3357 0 : tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(),
3358 0 : "Shard not scheduled (policy {:?}), cannot generate pass-through URL",
3359 : shard.policy
3360 : );
3361 0 : return Err(ApiError::Conflict(
3362 0 : "Cannot call timeline API on non-attached tenant".to_string(),
3363 0 : ));
3364 : };
3365 :
3366 0 : let Some(node) = locked.nodes.get(node_id) else {
3367 : // This should never happen
3368 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
3369 0 : "Shard refers to nonexistent node"
3370 0 : )));
3371 : };
3372 :
3373 0 : Ok((node.clone(), *tenant_shard_id))
3374 0 : }
3375 :
3376 0 : pub(crate) fn tenant_locate(
3377 0 : &self,
3378 0 : tenant_id: TenantId,
3379 0 : ) -> Result<TenantLocateResponse, ApiError> {
3380 0 : let locked = self.inner.read().unwrap();
3381 0 : tracing::info!("Locating shards for tenant {tenant_id}");
3382 :
3383 0 : let mut result = Vec::new();
3384 0 : let mut shard_params: Option<ShardParameters> = None;
3385 :
3386 0 : for (tenant_shard_id, shard) in locked.tenants.range(TenantShardId::tenant_range(tenant_id))
3387 : {
3388 0 : let node_id =
3389 0 : shard
3390 0 : .intent
3391 0 : .get_attached()
3392 0 : .ok_or(ApiError::BadRequest(anyhow::anyhow!(
3393 0 : "Cannot locate a tenant that is not attached"
3394 0 : )))?;
3395 :
3396 0 : let node = locked
3397 0 : .nodes
3398 0 : .get(&node_id)
3399 0 : .expect("Pageservers may not be deleted while referenced");
3400 0 :
3401 0 : result.push(node.shard_location(*tenant_shard_id));
3402 0 :
3403 0 : match &shard_params {
3404 0 : None => {
3405 0 : shard_params = Some(ShardParameters {
3406 0 : stripe_size: shard.shard.stripe_size,
3407 0 : count: shard.shard.count,
3408 0 : });
3409 0 : }
3410 0 : Some(params) => {
3411 0 : if params.stripe_size != shard.shard.stripe_size {
3412 : // This should never happen. We enforce at runtime because it's simpler than
3413 : // adding an extra per-tenant data structure to store the things that should be the same
3414 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
3415 0 : "Inconsistent shard stripe size parameters!"
3416 0 : )));
3417 0 : }
3418 : }
3419 : }
3420 : }
3421 :
3422 0 : if result.is_empty() {
3423 0 : return Err(ApiError::NotFound(
3424 0 : anyhow::anyhow!("No shards for this tenant ID found").into(),
3425 0 : ));
3426 0 : }
3427 0 : let shard_params = shard_params.expect("result is non-empty, therefore this is set");
3428 0 : tracing::info!(
3429 0 : "Located tenant {} with params {:?} on shards {}",
3430 0 : tenant_id,
3431 0 : shard_params,
3432 0 : result
3433 0 : .iter()
3434 0 : .map(|s| format!("{:?}", s))
3435 0 : .collect::<Vec<_>>()
3436 0 : .join(",")
3437 : );
3438 :
3439 0 : Ok(TenantLocateResponse {
3440 0 : shards: result,
3441 0 : shard_params,
3442 0 : })
3443 0 : }
3444 :
3445 : /// Returns None if the input iterator of shards does not include a shard with number=0
3446 0 : fn tenant_describe_impl<'a>(
3447 0 : &self,
3448 0 : shards: impl Iterator<Item = &'a TenantShard>,
3449 0 : ) -> Option<TenantDescribeResponse> {
3450 0 : let mut shard_zero = None;
3451 0 : let mut describe_shards = Vec::new();
3452 :
3453 0 : for shard in shards {
3454 0 : if shard.tenant_shard_id.is_shard_zero() {
3455 0 : shard_zero = Some(shard);
3456 0 : }
3457 :
3458 0 : describe_shards.push(TenantDescribeResponseShard {
3459 0 : tenant_shard_id: shard.tenant_shard_id,
3460 0 : node_attached: *shard.intent.get_attached(),
3461 0 : node_secondary: shard.intent.get_secondary().to_vec(),
3462 0 : last_error: shard
3463 0 : .last_error
3464 0 : .lock()
3465 0 : .unwrap()
3466 0 : .as_ref()
3467 0 : .map(|e| format!("{e}"))
3468 0 : .unwrap_or("".to_string())
3469 0 : .clone(),
3470 0 : is_reconciling: shard.reconciler.is_some(),
3471 0 : is_pending_compute_notification: shard.pending_compute_notification,
3472 0 : is_splitting: matches!(shard.splitting, SplitState::Splitting),
3473 0 : scheduling_policy: *shard.get_scheduling_policy(),
3474 : })
3475 : }
3476 :
3477 0 : let shard_zero = shard_zero?;
3478 :
3479 0 : Some(TenantDescribeResponse {
3480 0 : tenant_id: shard_zero.tenant_shard_id.tenant_id,
3481 0 : shards: describe_shards,
3482 0 : stripe_size: shard_zero.shard.stripe_size,
3483 0 : policy: shard_zero.policy.clone(),
3484 0 : config: shard_zero.config.clone(),
3485 0 : })
3486 0 : }
3487 :
3488 0 : pub(crate) fn tenant_describe(
3489 0 : &self,
3490 0 : tenant_id: TenantId,
3491 0 : ) -> Result<TenantDescribeResponse, ApiError> {
3492 0 : let locked = self.inner.read().unwrap();
3493 0 :
3494 0 : self.tenant_describe_impl(
3495 0 : locked
3496 0 : .tenants
3497 0 : .range(TenantShardId::tenant_range(tenant_id))
3498 0 : .map(|(_k, v)| v),
3499 0 : )
3500 0 : .ok_or_else(|| ApiError::NotFound(anyhow::anyhow!("Tenant {tenant_id} not found").into()))
3501 0 : }
3502 :
3503 0 : pub(crate) fn tenant_list(&self) -> Vec<TenantDescribeResponse> {
3504 0 : let locked = self.inner.read().unwrap();
3505 0 :
3506 0 : let mut result = Vec::new();
3507 0 : for (_tenant_id, tenant_shards) in
3508 0 : &locked.tenants.iter().group_by(|(id, _shard)| id.tenant_id)
3509 0 : {
3510 0 : result.push(
3511 0 : self.tenant_describe_impl(tenant_shards.map(|(_k, v)| v))
3512 0 : .expect("Groups are always non-empty"),
3513 0 : );
3514 0 : }
3515 :
3516 0 : result
3517 0 : }
3518 :
3519 0 : #[instrument(skip_all, fields(tenant_id=%op.tenant_id))]
3520 : async fn abort_tenant_shard_split(
3521 : &self,
3522 : op: &TenantShardSplitAbort,
3523 : ) -> Result<(), TenantShardSplitAbortError> {
3524 : // Cleaning up a split:
3525 : // - Parent shards are not destroyed during a split, just detached.
3526 : // - Failed pageserver split API calls can leave the remote node with just the parent attached,
3527 : // just the children attached, or both.
3528 : //
3529 : // Therefore our work to do is to:
3530 : // 1. Clean up storage controller's internal state to just refer to parents, no children
3531 : // 2. Call out to pageservers to ensure that children are detached
3532 : // 3. Call out to pageservers to ensure that parents are attached.
3533 : //
3534 : // Crash safety:
3535 : // - If the storage controller stops running during this cleanup *after* clearing the splitting state
3536 : // from our database, then [`Self::startup_reconcile`] will regard child attachments as garbage
3537 : // and detach them.
3538 : // - TODO: If the storage controller stops running during this cleanup *before* clearing the splitting state
3539 : // from our database, then we will re-enter this cleanup routine on startup.
3540 :
3541 : let TenantShardSplitAbort {
3542 : tenant_id,
3543 : new_shard_count,
3544 : new_stripe_size,
3545 : ..
3546 : } = op;
3547 :
3548 : // First abort persistent state, if any exists.
3549 : match self
3550 : .persistence
3551 : .abort_shard_split(*tenant_id, *new_shard_count)
3552 : .await?
3553 : {
3554 : AbortShardSplitStatus::Aborted => {
3555 : // Proceed to roll back any child shards created on pageservers
3556 : }
3557 : AbortShardSplitStatus::Complete => {
3558 : // The split completed (we might hit that path if e.g. our database transaction
3559 : // to write the completion landed in the database, but we dropped connection
3560 : // before seeing the result).
3561 : //
3562 : // We must update in-memory state to reflect the successful split.
3563 : self.tenant_shard_split_commit_inmem(
3564 : *tenant_id,
3565 : *new_shard_count,
3566 : *new_stripe_size,
3567 : );
3568 : return Ok(());
3569 : }
3570 : }
3571 :
3572 : // Clean up in-memory state, and accumulate the list of child locations that need detaching
3573 : let detach_locations: Vec<(Node, TenantShardId)> = {
3574 : let mut detach_locations = Vec::new();
3575 : let mut locked = self.inner.write().unwrap();
3576 : let (nodes, tenants, scheduler) = locked.parts_mut();
3577 :
3578 : for (tenant_shard_id, shard) in
3579 : tenants.range_mut(TenantShardId::tenant_range(op.tenant_id))
3580 : {
3581 : if shard.shard.count == op.new_shard_count {
3582 : // Surprising: the phase of [`Self::do_tenant_shard_split`] which inserts child shards in-memory
3583 : // is infallible, so if we got an error we shouldn't have got that far.
3584 : tracing::warn!(
3585 : "During split abort, child shard {tenant_shard_id} found in-memory"
3586 : );
3587 : continue;
3588 : }
3589 :
3590 : // Add the children of this shard to this list of things to detach
3591 : if let Some(node_id) = shard.intent.get_attached() {
3592 : for child_id in tenant_shard_id.split(*new_shard_count) {
3593 : detach_locations.push((
3594 : nodes
3595 : .get(node_id)
3596 : .expect("Intent references nonexistent node")
3597 : .clone(),
3598 : child_id,
3599 : ));
3600 : }
3601 : } else {
3602 : tracing::warn!(
3603 : "During split abort, shard {tenant_shard_id} has no attached location"
3604 : );
3605 : }
3606 :
3607 : tracing::info!("Restoring parent shard {tenant_shard_id}");
3608 : shard.splitting = SplitState::Idle;
3609 : if let Err(e) = shard.schedule(scheduler, &mut ScheduleContext::default()) {
3610 : // If this shard can't be scheduled now (perhaps due to offline nodes or
3611 : // capacity issues), that must not prevent us rolling back a split. In this
3612 : // case it should be eventually scheduled in the background.
3613 : tracing::warn!("Failed to schedule {tenant_shard_id} during shard abort: {e}")
3614 : }
3615 :
3616 : self.maybe_reconcile_shard(shard, nodes);
3617 : }
3618 :
3619 : // We don't expect any new_shard_count shards to exist here, but drop them just in case
3620 0 : tenants.retain(|_id, s| s.shard.count != *new_shard_count);
3621 :
3622 : detach_locations
3623 : };
3624 :
3625 : for (node, child_id) in detach_locations {
3626 : if !node.is_available() {
3627 : // An unavailable node cannot be cleaned up now: to avoid blocking forever, we will permit this, and
3628 : // rely on the reconciliation that happens when a node transitions to Active to clean up. Since we have
3629 : // removed child shards from our in-memory state and database, the reconciliation will implicitly remove
3630 : // them from the node.
3631 : tracing::warn!("Node {node} unavailable, can't clean up during split abort. It will be cleaned up when it is reactivated.");
3632 : continue;
3633 : }
3634 :
3635 : // Detach the remote child. If the pageserver split API call is still in progress, this call will get
3636 : // a 503 and retry, up to our limit.
3637 : tracing::info!("Detaching {child_id} on {node}...");
3638 : match node
3639 : .with_client_retries(
3640 0 : |client| async move {
3641 0 : let config = LocationConfig {
3642 0 : mode: LocationConfigMode::Detached,
3643 0 : generation: None,
3644 0 : secondary_conf: None,
3645 0 : shard_number: child_id.shard_number.0,
3646 0 : shard_count: child_id.shard_count.literal(),
3647 0 : // Stripe size and tenant config don't matter when detaching
3648 0 : shard_stripe_size: 0,
3649 0 : tenant_conf: TenantConfig::default(),
3650 0 : };
3651 0 :
3652 0 : client.location_config(child_id, config, None, false).await
3653 0 : },
3654 : &self.config.jwt_token,
3655 : 1,
3656 : 10,
3657 : Duration::from_secs(5),
3658 : &self.cancel,
3659 : )
3660 : .await
3661 : {
3662 : Some(Ok(_)) => {}
3663 : Some(Err(e)) => {
3664 : // We failed to communicate with the remote node. This is problematic: we may be
3665 : // leaving it with a rogue child shard.
3666 : tracing::warn!(
3667 : "Failed to detach child {child_id} from node {node} during abort"
3668 : );
3669 : return Err(e.into());
3670 : }
3671 : None => {
3672 : // Cancellation: we were shutdown or the node went offline. Shutdown is fine, we'll
3673 : // clean up on restart. The node going offline requires a retry.
3674 : return Err(TenantShardSplitAbortError::Unavailable);
3675 : }
3676 : };
3677 : }
3678 :
3679 : tracing::info!("Successfully aborted split");
3680 : Ok(())
3681 : }
3682 :
3683 : /// Infallible final stage of [`Self::tenant_shard_split`]: update the contents
3684 : /// of the tenant map to reflect the child shards that exist after the split.
3685 0 : fn tenant_shard_split_commit_inmem(
3686 0 : &self,
3687 0 : tenant_id: TenantId,
3688 0 : new_shard_count: ShardCount,
3689 0 : new_stripe_size: Option<ShardStripeSize>,
3690 0 : ) -> (
3691 0 : TenantShardSplitResponse,
3692 0 : Vec<(TenantShardId, NodeId, ShardStripeSize)>,
3693 0 : Vec<ReconcilerWaiter>,
3694 0 : ) {
3695 0 : let mut response = TenantShardSplitResponse {
3696 0 : new_shards: Vec::new(),
3697 0 : };
3698 0 : let mut child_locations = Vec::new();
3699 0 : let mut waiters = Vec::new();
3700 0 :
3701 0 : {
3702 0 : let mut locked = self.inner.write().unwrap();
3703 0 :
3704 0 : let parent_ids = locked
3705 0 : .tenants
3706 0 : .range(TenantShardId::tenant_range(tenant_id))
3707 0 : .map(|(shard_id, _)| *shard_id)
3708 0 : .collect::<Vec<_>>();
3709 0 :
3710 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
3711 0 : for parent_id in parent_ids {
3712 0 : let child_ids = parent_id.split(new_shard_count);
3713 :
3714 0 : let (pageserver, generation, policy, parent_ident, config) = {
3715 0 : let mut old_state = tenants
3716 0 : .remove(&parent_id)
3717 0 : .expect("It was present, we just split it");
3718 0 :
3719 0 : // A non-splitting state is impossible, because [`Self::tenant_shard_split`] holds
3720 0 : // a TenantId lock and passes it through to [`TenantShardSplitAbort`] in case of cleanup:
3721 0 : // nothing else can clear this.
3722 0 : assert!(matches!(old_state.splitting, SplitState::Splitting));
3723 :
3724 0 : let old_attached = old_state.intent.get_attached().unwrap();
3725 0 : old_state.intent.clear(scheduler);
3726 0 : let generation = old_state.generation.expect("Shard must have been attached");
3727 0 : (
3728 0 : old_attached,
3729 0 : generation,
3730 0 : old_state.policy,
3731 0 : old_state.shard,
3732 0 : old_state.config,
3733 0 : )
3734 0 : };
3735 0 :
3736 0 : let mut schedule_context = ScheduleContext::default();
3737 0 : for child in child_ids {
3738 0 : let mut child_shard = parent_ident;
3739 0 : child_shard.number = child.shard_number;
3740 0 : child_shard.count = child.shard_count;
3741 0 : if let Some(stripe_size) = new_stripe_size {
3742 0 : child_shard.stripe_size = stripe_size;
3743 0 : }
3744 :
3745 0 : let mut child_observed: HashMap<NodeId, ObservedStateLocation> = HashMap::new();
3746 0 : child_observed.insert(
3747 0 : pageserver,
3748 0 : ObservedStateLocation {
3749 0 : conf: Some(attached_location_conf(
3750 0 : generation,
3751 0 : &child_shard,
3752 0 : &config,
3753 0 : &policy,
3754 0 : )),
3755 0 : },
3756 0 : );
3757 0 :
3758 0 : let mut child_state = TenantShard::new(child, child_shard, policy.clone());
3759 0 : child_state.intent = IntentState::single(scheduler, Some(pageserver));
3760 0 : child_state.observed = ObservedState {
3761 0 : locations: child_observed,
3762 0 : };
3763 0 : child_state.generation = Some(generation);
3764 0 : child_state.config = config.clone();
3765 0 :
3766 0 : // The child's TenantShard::splitting is intentionally left at the default value of Idle,
3767 0 : // as at this point in the split process we have succeeded and this part is infallible:
3768 0 : // we will never need to do any special recovery from this state.
3769 0 :
3770 0 : child_locations.push((child, pageserver, child_shard.stripe_size));
3771 :
3772 0 : if let Err(e) = child_state.schedule(scheduler, &mut schedule_context) {
3773 : // This is not fatal, because we've implicitly already got an attached
3774 : // location for the child shard. Failure here just means we couldn't
3775 : // find a secondary (e.g. because cluster is overloaded).
3776 0 : tracing::warn!("Failed to schedule child shard {child}: {e}");
3777 0 : }
3778 : // In the background, attach secondary locations for the new shards
3779 0 : if let Some(waiter) = self.maybe_reconcile_shard(&mut child_state, nodes) {
3780 0 : waiters.push(waiter);
3781 0 : }
3782 :
3783 0 : tenants.insert(child, child_state);
3784 0 : response.new_shards.push(child);
3785 : }
3786 : }
3787 0 : (response, child_locations, waiters)
3788 0 : }
3789 0 : }
3790 :
3791 0 : async fn tenant_shard_split_start_secondaries(
3792 0 : &self,
3793 0 : tenant_id: TenantId,
3794 0 : waiters: Vec<ReconcilerWaiter>,
3795 0 : ) {
3796 : // Wait for initial reconcile of child shards, this creates the secondary locations
3797 0 : if let Err(e) = self.await_waiters(waiters, RECONCILE_TIMEOUT).await {
3798 : // This is not a failure to split: it's some issue reconciling the new child shards, perhaps
3799 : // their secondaries couldn't be attached.
3800 0 : tracing::warn!("Failed to reconcile after split: {e}");
3801 0 : return;
3802 0 : }
3803 :
3804 : // Take the state lock to discover the attached & secondary intents for all shards
3805 0 : let (attached, secondary) = {
3806 0 : let locked = self.inner.read().unwrap();
3807 0 : let mut attached = Vec::new();
3808 0 : let mut secondary = Vec::new();
3809 :
3810 0 : for (tenant_shard_id, shard) in
3811 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
3812 : {
3813 0 : let Some(node_id) = shard.intent.get_attached() else {
3814 : // Unexpected. Race with a PlacementPolicy change?
3815 0 : tracing::warn!(
3816 0 : "No attached node on {tenant_shard_id} immediately after shard split!"
3817 : );
3818 0 : continue;
3819 : };
3820 :
3821 0 : let Some(secondary_node_id) = shard.intent.get_secondary().first() else {
3822 : // No secondary location. Nothing for us to do.
3823 0 : continue;
3824 : };
3825 :
3826 0 : let attached_node = locked
3827 0 : .nodes
3828 0 : .get(node_id)
3829 0 : .expect("Pageservers may not be deleted while referenced");
3830 0 :
3831 0 : let secondary_node = locked
3832 0 : .nodes
3833 0 : .get(secondary_node_id)
3834 0 : .expect("Pageservers may not be deleted while referenced");
3835 0 :
3836 0 : attached.push((*tenant_shard_id, attached_node.clone()));
3837 0 : secondary.push((*tenant_shard_id, secondary_node.clone()));
3838 : }
3839 0 : (attached, secondary)
3840 0 : };
3841 0 :
3842 0 : if secondary.is_empty() {
3843 : // No secondary locations; nothing for us to do
3844 0 : return;
3845 0 : }
3846 :
3847 0 : for result in self
3848 0 : .tenant_for_shards_api(
3849 0 : attached,
3850 0 : |tenant_shard_id, client| async move {
3851 0 : client.tenant_heatmap_upload(tenant_shard_id).await
3852 0 : },
3853 0 : 1,
3854 0 : 1,
3855 0 : SHORT_RECONCILE_TIMEOUT,
3856 0 : &self.cancel,
3857 0 : )
3858 0 : .await
3859 : {
3860 0 : if let Err(e) = result {
3861 0 : tracing::warn!("Error calling heatmap upload after shard split: {e}");
3862 0 : return;
3863 0 : }
3864 : }
3865 :
3866 0 : for result in self
3867 0 : .tenant_for_shards_api(
3868 0 : secondary,
3869 0 : |tenant_shard_id, client| async move {
3870 0 : client
3871 0 : .tenant_secondary_download(tenant_shard_id, Some(Duration::ZERO))
3872 0 : .await
3873 0 : },
3874 0 : 1,
3875 0 : 1,
3876 0 : SHORT_RECONCILE_TIMEOUT,
3877 0 : &self.cancel,
3878 0 : )
3879 0 : .await
3880 : {
3881 0 : if let Err(e) = result {
3882 0 : tracing::warn!("Error calling secondary download after shard split: {e}");
3883 0 : return;
3884 0 : }
3885 : }
3886 0 : }
3887 :
3888 0 : pub(crate) async fn tenant_shard_split(
3889 0 : &self,
3890 0 : tenant_id: TenantId,
3891 0 : split_req: TenantShardSplitRequest,
3892 0 : ) -> Result<TenantShardSplitResponse, ApiError> {
3893 : // TODO: return 503 if we get stuck waiting for this lock
3894 : // (issue https://github.com/neondatabase/neon/issues/7108)
3895 0 : let _tenant_lock = trace_exclusive_lock(
3896 0 : &self.tenant_op_locks,
3897 0 : tenant_id,
3898 0 : TenantOperations::ShardSplit,
3899 0 : )
3900 0 : .await;
3901 :
3902 0 : let new_shard_count = ShardCount::new(split_req.new_shard_count);
3903 0 : let new_stripe_size = split_req.new_stripe_size;
3904 :
3905 : // Validate the request and construct parameters. This phase is fallible, but does not require
3906 : // rollback on errors, as it does no I/O and mutates no state.
3907 0 : let shard_split_params = match self.prepare_tenant_shard_split(tenant_id, split_req)? {
3908 0 : ShardSplitAction::NoOp(resp) => return Ok(resp),
3909 0 : ShardSplitAction::Split(params) => params,
3910 : };
3911 :
3912 : // Execute this split: this phase mutates state and does remote I/O on pageservers. If it fails,
3913 : // we must roll back.
3914 0 : let r = self
3915 0 : .do_tenant_shard_split(tenant_id, shard_split_params)
3916 0 : .await;
3917 :
3918 0 : let (response, waiters) = match r {
3919 0 : Ok(r) => r,
3920 0 : Err(e) => {
3921 0 : // Split might be part-done, we must do work to abort it.
3922 0 : tracing::warn!("Enqueuing background abort of split on {tenant_id}");
3923 0 : self.abort_tx
3924 0 : .send(TenantShardSplitAbort {
3925 0 : tenant_id,
3926 0 : new_shard_count,
3927 0 : new_stripe_size,
3928 0 : _tenant_lock,
3929 0 : })
3930 0 : // Ignore error sending: that just means we're shutting down: aborts are ephemeral so it's fine to drop it.
3931 0 : .ok();
3932 0 : return Err(e);
3933 : }
3934 : };
3935 :
3936 : // The split is now complete. As an optimization, we will trigger all the child shards to upload
3937 : // a heatmap immediately, and all their secondary locations to start downloading: this avoids waiting
3938 : // for the background heatmap/download interval before secondaries get warm enough to migrate shards
3939 : // in [`Self::optimize_all`]
3940 0 : self.tenant_shard_split_start_secondaries(tenant_id, waiters)
3941 0 : .await;
3942 0 : Ok(response)
3943 0 : }
3944 :
3945 0 : fn prepare_tenant_shard_split(
3946 0 : &self,
3947 0 : tenant_id: TenantId,
3948 0 : split_req: TenantShardSplitRequest,
3949 0 : ) -> Result<ShardSplitAction, ApiError> {
3950 0 : fail::fail_point!("shard-split-validation", |_| Err(ApiError::BadRequest(
3951 0 : anyhow::anyhow!("failpoint")
3952 0 : )));
3953 :
3954 0 : let mut policy = None;
3955 0 : let mut config = None;
3956 0 : let mut shard_ident = None;
3957 : // Validate input, and calculate which shards we will create
3958 0 : let (old_shard_count, targets) =
3959 : {
3960 0 : let locked = self.inner.read().unwrap();
3961 0 :
3962 0 : let pageservers = locked.nodes.clone();
3963 0 :
3964 0 : let mut targets = Vec::new();
3965 0 :
3966 0 : // In case this is a retry, count how many already-split shards we found
3967 0 : let mut children_found = Vec::new();
3968 0 : let mut old_shard_count = None;
3969 :
3970 0 : for (tenant_shard_id, shard) in
3971 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
3972 : {
3973 0 : match shard.shard.count.count().cmp(&split_req.new_shard_count) {
3974 : Ordering::Equal => {
3975 : // Already split this
3976 0 : children_found.push(*tenant_shard_id);
3977 0 : continue;
3978 : }
3979 : Ordering::Greater => {
3980 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
3981 0 : "Requested count {} but already have shards at count {}",
3982 0 : split_req.new_shard_count,
3983 0 : shard.shard.count.count()
3984 0 : )));
3985 : }
3986 0 : Ordering::Less => {
3987 0 : // Fall through: this shard has lower count than requested,
3988 0 : // is a candidate for splitting.
3989 0 : }
3990 0 : }
3991 0 :
3992 0 : match old_shard_count {
3993 0 : None => old_shard_count = Some(shard.shard.count),
3994 0 : Some(old_shard_count) => {
3995 0 : if old_shard_count != shard.shard.count {
3996 : // We may hit this case if a caller asked for two splits to
3997 : // different sizes, before the first one is complete.
3998 : // e.g. 1->2, 2->4, where the 4 call comes while we have a mixture
3999 : // of shard_count=1 and shard_count=2 shards in the map.
4000 0 : return Err(ApiError::Conflict(
4001 0 : "Cannot split, currently mid-split".to_string(),
4002 0 : ));
4003 0 : }
4004 : }
4005 : }
4006 0 : if policy.is_none() {
4007 0 : policy = Some(shard.policy.clone());
4008 0 : }
4009 0 : if shard_ident.is_none() {
4010 0 : shard_ident = Some(shard.shard);
4011 0 : }
4012 0 : if config.is_none() {
4013 0 : config = Some(shard.config.clone());
4014 0 : }
4015 :
4016 0 : if tenant_shard_id.shard_count.count() == split_req.new_shard_count {
4017 0 : tracing::info!(
4018 0 : "Tenant shard {} already has shard count {}",
4019 : tenant_shard_id,
4020 : split_req.new_shard_count
4021 : );
4022 0 : continue;
4023 0 : }
4024 :
4025 0 : let node_id = shard.intent.get_attached().ok_or(ApiError::BadRequest(
4026 0 : anyhow::anyhow!("Cannot split a tenant that is not attached"),
4027 0 : ))?;
4028 :
4029 0 : let node = pageservers
4030 0 : .get(&node_id)
4031 0 : .expect("Pageservers may not be deleted while referenced");
4032 0 :
4033 0 : targets.push(ShardSplitTarget {
4034 0 : parent_id: *tenant_shard_id,
4035 0 : node: node.clone(),
4036 0 : child_ids: tenant_shard_id
4037 0 : .split(ShardCount::new(split_req.new_shard_count)),
4038 0 : });
4039 : }
4040 :
4041 0 : if targets.is_empty() {
4042 0 : if children_found.len() == split_req.new_shard_count as usize {
4043 0 : return Ok(ShardSplitAction::NoOp(TenantShardSplitResponse {
4044 0 : new_shards: children_found,
4045 0 : }));
4046 : } else {
4047 : // No shards found to split, and no existing children found: the
4048 : // tenant doesn't exist at all.
4049 0 : return Err(ApiError::NotFound(
4050 0 : anyhow::anyhow!("Tenant {} not found", tenant_id).into(),
4051 0 : ));
4052 : }
4053 0 : }
4054 0 :
4055 0 : (old_shard_count, targets)
4056 0 : };
4057 0 :
4058 0 : // unwrap safety: we would have returned above if we didn't find at least one shard to split
4059 0 : let old_shard_count = old_shard_count.unwrap();
4060 0 : let shard_ident = if let Some(new_stripe_size) = split_req.new_stripe_size {
4061 : // This ShardIdentity will be used as the template for all children, so this implicitly
4062 : // applies the new stripe size to the children.
4063 0 : let mut shard_ident = shard_ident.unwrap();
4064 0 : if shard_ident.count.count() > 1 && shard_ident.stripe_size != new_stripe_size {
4065 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)));
4066 0 : }
4067 0 :
4068 0 : shard_ident.stripe_size = new_stripe_size;
4069 0 : tracing::info!("applied stripe size {}", shard_ident.stripe_size.0);
4070 0 : shard_ident
4071 : } else {
4072 0 : shard_ident.unwrap()
4073 : };
4074 0 : let policy = policy.unwrap();
4075 0 : let config = config.unwrap();
4076 0 :
4077 0 : Ok(ShardSplitAction::Split(ShardSplitParams {
4078 0 : old_shard_count,
4079 0 : new_shard_count: ShardCount::new(split_req.new_shard_count),
4080 0 : new_stripe_size: split_req.new_stripe_size,
4081 0 : targets,
4082 0 : policy,
4083 0 : config,
4084 0 : shard_ident,
4085 0 : }))
4086 0 : }
4087 :
4088 0 : async fn do_tenant_shard_split(
4089 0 : &self,
4090 0 : tenant_id: TenantId,
4091 0 : params: ShardSplitParams,
4092 0 : ) -> Result<(TenantShardSplitResponse, Vec<ReconcilerWaiter>), ApiError> {
4093 0 : // FIXME: we have dropped self.inner lock, and not yet written anything to the database: another
4094 0 : // request could occur here, deleting or mutating the tenant. begin_shard_split checks that the
4095 0 : // parent shards exist as expected, but it would be neater to do the above pre-checks within the
4096 0 : // same database transaction rather than pre-check in-memory and then maybe-fail the database write.
4097 0 : // (https://github.com/neondatabase/neon/issues/6676)
4098 0 :
4099 0 : let ShardSplitParams {
4100 0 : old_shard_count,
4101 0 : new_shard_count,
4102 0 : new_stripe_size,
4103 0 : mut targets,
4104 0 : policy,
4105 0 : config,
4106 0 : shard_ident,
4107 0 : } = params;
4108 :
4109 : // Drop any secondary locations: pageservers do not support splitting these, and in any case the
4110 : // end-state for a split tenant will usually be to have secondary locations on different nodes.
4111 : // The reconciliation calls in this block also implicitly cancel+barrier wrt any ongoing reconciliation
4112 : // at the time of split.
4113 0 : let waiters = {
4114 0 : let mut locked = self.inner.write().unwrap();
4115 0 : let mut waiters = Vec::new();
4116 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
4117 0 : for target in &mut targets {
4118 0 : let Some(shard) = tenants.get_mut(&target.parent_id) else {
4119 : // Paranoia check: this shouldn't happen: we have the oplock for this tenant ID.
4120 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
4121 0 : "Shard {} not found",
4122 0 : target.parent_id
4123 0 : )));
4124 : };
4125 :
4126 0 : if shard.intent.get_attached() != &Some(target.node.get_id()) {
4127 : // Paranoia check: this shouldn't happen: we have the oplock for this tenant ID.
4128 0 : return Err(ApiError::Conflict(format!(
4129 0 : "Shard {} unexpectedly rescheduled during split",
4130 0 : target.parent_id
4131 0 : )));
4132 0 : }
4133 0 :
4134 0 : // Irrespective of PlacementPolicy, clear secondary locations from intent
4135 0 : shard.intent.clear_secondary(scheduler);
4136 :
4137 : // Run Reconciler to execute detach fo secondary locations.
4138 0 : if let Some(waiter) = self.maybe_reconcile_shard(shard, nodes) {
4139 0 : waiters.push(waiter);
4140 0 : }
4141 : }
4142 0 : waiters
4143 0 : };
4144 0 : self.await_waiters(waiters, RECONCILE_TIMEOUT).await?;
4145 :
4146 : // Before creating any new child shards in memory or on the pageservers, persist them: this
4147 : // enables us to ensure that we will always be able to clean up if something goes wrong. This also
4148 : // acts as the protection against two concurrent attempts to split: one of them will get a database
4149 : // error trying to insert the child shards.
4150 0 : let mut child_tsps = Vec::new();
4151 0 : for target in &targets {
4152 0 : let mut this_child_tsps = Vec::new();
4153 0 : for child in &target.child_ids {
4154 0 : let mut child_shard = shard_ident;
4155 0 : child_shard.number = child.shard_number;
4156 0 : child_shard.count = child.shard_count;
4157 0 :
4158 0 : tracing::info!(
4159 0 : "Create child shard persistence with stripe size {}",
4160 : shard_ident.stripe_size.0
4161 : );
4162 :
4163 0 : this_child_tsps.push(TenantShardPersistence {
4164 0 : tenant_id: child.tenant_id.to_string(),
4165 0 : shard_number: child.shard_number.0 as i32,
4166 0 : shard_count: child.shard_count.literal() as i32,
4167 0 : shard_stripe_size: shard_ident.stripe_size.0 as i32,
4168 0 : // Note: this generation is a placeholder, [`Persistence::begin_shard_split`] will
4169 0 : // populate the correct generation as part of its transaction, to protect us
4170 0 : // against racing with changes in the state of the parent.
4171 0 : generation: None,
4172 0 : generation_pageserver: Some(target.node.get_id().0 as i64),
4173 0 : placement_policy: serde_json::to_string(&policy).unwrap(),
4174 0 : config: serde_json::to_string(&config).unwrap(),
4175 0 : splitting: SplitState::Splitting,
4176 0 :
4177 0 : // Scheduling policies do not carry through to children
4178 0 : scheduling_policy: serde_json::to_string(&ShardSchedulingPolicy::default())
4179 0 : .unwrap(),
4180 0 : });
4181 : }
4182 :
4183 0 : child_tsps.push((target.parent_id, this_child_tsps));
4184 : }
4185 :
4186 0 : if let Err(e) = self
4187 0 : .persistence
4188 0 : .begin_shard_split(old_shard_count, tenant_id, child_tsps)
4189 0 : .await
4190 : {
4191 0 : match e {
4192 : DatabaseError::Query(diesel::result::Error::DatabaseError(
4193 : DatabaseErrorKind::UniqueViolation,
4194 : _,
4195 : )) => {
4196 : // Inserting a child shard violated a unique constraint: we raced with another call to
4197 : // this function
4198 0 : tracing::warn!("Conflicting attempt to split {tenant_id}: {e}");
4199 0 : return Err(ApiError::Conflict("Tenant is already splitting".into()));
4200 : }
4201 0 : _ => return Err(ApiError::InternalServerError(e.into())),
4202 : }
4203 0 : }
4204 0 : fail::fail_point!("shard-split-post-begin", |_| Err(
4205 0 : ApiError::InternalServerError(anyhow::anyhow!("failpoint"))
4206 0 : ));
4207 :
4208 : // Now that I have persisted the splitting state, apply it in-memory. This is infallible, so
4209 : // callers may assume that if splitting is set in memory, then it was persisted, and if splitting
4210 : // is not set in memory, then it was not persisted.
4211 : {
4212 0 : let mut locked = self.inner.write().unwrap();
4213 0 : for target in &targets {
4214 0 : if let Some(parent_shard) = locked.tenants.get_mut(&target.parent_id) {
4215 0 : parent_shard.splitting = SplitState::Splitting;
4216 0 : // Put the observed state to None, to reflect that it is indeterminate once we start the
4217 0 : // split operation.
4218 0 : parent_shard
4219 0 : .observed
4220 0 : .locations
4221 0 : .insert(target.node.get_id(), ObservedStateLocation { conf: None });
4222 0 : }
4223 : }
4224 : }
4225 :
4226 : // TODO: issue split calls concurrently (this only matters once we're splitting
4227 : // N>1 shards into M shards -- initially we're usually splitting 1 shard into N).
4228 :
4229 0 : for target in &targets {
4230 : let ShardSplitTarget {
4231 0 : parent_id,
4232 0 : node,
4233 0 : child_ids,
4234 0 : } = target;
4235 0 : let client = PageserverClient::new(
4236 0 : node.get_id(),
4237 0 : node.base_url(),
4238 0 : self.config.jwt_token.as_deref(),
4239 0 : );
4240 0 : let response = client
4241 0 : .tenant_shard_split(
4242 0 : *parent_id,
4243 0 : TenantShardSplitRequest {
4244 0 : new_shard_count: new_shard_count.literal(),
4245 0 : new_stripe_size,
4246 0 : },
4247 0 : )
4248 0 : .await
4249 0 : .map_err(|e| ApiError::Conflict(format!("Failed to split {}: {}", parent_id, e)))?;
4250 :
4251 0 : fail::fail_point!("shard-split-post-remote", |_| Err(ApiError::Conflict(
4252 0 : "failpoint".to_string()
4253 0 : )));
4254 :
4255 0 : failpoint_support::sleep_millis_async!("shard-split-post-remote-sleep", &self.cancel);
4256 :
4257 0 : tracing::info!(
4258 0 : "Split {} into {}",
4259 0 : parent_id,
4260 0 : response
4261 0 : .new_shards
4262 0 : .iter()
4263 0 : .map(|s| format!("{:?}", s))
4264 0 : .collect::<Vec<_>>()
4265 0 : .join(",")
4266 : );
4267 :
4268 0 : if &response.new_shards != child_ids {
4269 : // This should never happen: the pageserver should agree with us on how shard splits work.
4270 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
4271 0 : "Splitting shard {} resulted in unexpected IDs: {:?} (expected {:?})",
4272 0 : parent_id,
4273 0 : response.new_shards,
4274 0 : child_ids
4275 0 : )));
4276 0 : }
4277 : }
4278 :
4279 : // TODO: if the pageserver restarted concurrently with our split API call,
4280 : // the actual generation of the child shard might differ from the generation
4281 : // we expect it to have. In order for our in-database generation to end up
4282 : // correct, we should carry the child generation back in the response and apply it here
4283 : // in complete_shard_split (and apply the correct generation in memory)
4284 : // (or, we can carry generation in the request and reject the request if
4285 : // it doesn't match, but that requires more retry logic on this side)
4286 :
4287 0 : self.persistence
4288 0 : .complete_shard_split(tenant_id, old_shard_count)
4289 0 : .await?;
4290 :
4291 0 : fail::fail_point!("shard-split-post-complete", |_| Err(
4292 0 : ApiError::InternalServerError(anyhow::anyhow!("failpoint"))
4293 0 : ));
4294 :
4295 : // Replace all the shards we just split with their children: this phase is infallible.
4296 0 : let (response, child_locations, waiters) =
4297 0 : self.tenant_shard_split_commit_inmem(tenant_id, new_shard_count, new_stripe_size);
4298 0 :
4299 0 : // Send compute notifications for all the new shards
4300 0 : let mut failed_notifications = Vec::new();
4301 0 : for (child_id, child_ps, stripe_size) in child_locations {
4302 0 : if let Err(e) = self
4303 0 : .compute_hook
4304 0 : .notify(child_id, child_ps, stripe_size, &self.cancel)
4305 0 : .await
4306 : {
4307 0 : tracing::warn!("Failed to update compute of {}->{} during split, proceeding anyway to complete split ({e})",
4308 : child_id, child_ps);
4309 0 : failed_notifications.push(child_id);
4310 0 : }
4311 : }
4312 :
4313 : // If we failed any compute notifications, make a note to retry later.
4314 0 : if !failed_notifications.is_empty() {
4315 0 : let mut locked = self.inner.write().unwrap();
4316 0 : for failed in failed_notifications {
4317 0 : if let Some(shard) = locked.tenants.get_mut(&failed) {
4318 0 : shard.pending_compute_notification = true;
4319 0 : }
4320 : }
4321 0 : }
4322 :
4323 0 : Ok((response, waiters))
4324 0 : }
4325 :
4326 0 : pub(crate) async fn tenant_shard_migrate(
4327 0 : &self,
4328 0 : tenant_shard_id: TenantShardId,
4329 0 : migrate_req: TenantShardMigrateRequest,
4330 0 : ) -> Result<TenantShardMigrateResponse, ApiError> {
4331 0 : let waiter = {
4332 0 : let mut locked = self.inner.write().unwrap();
4333 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
4334 :
4335 0 : let Some(node) = nodes.get(&migrate_req.node_id) else {
4336 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
4337 0 : "Node {} not found",
4338 0 : migrate_req.node_id
4339 0 : )));
4340 : };
4341 :
4342 0 : if !node.is_available() {
4343 : // Warn but proceed: the caller may intend to manually adjust the placement of
4344 : // a shard even if the node is down, e.g. if intervening during an incident.
4345 0 : tracing::warn!("Migrating to unavailable node {node}");
4346 0 : }
4347 :
4348 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
4349 0 : return Err(ApiError::NotFound(
4350 0 : anyhow::anyhow!("Tenant shard not found").into(),
4351 0 : ));
4352 : };
4353 :
4354 0 : if shard.intent.get_attached() == &Some(migrate_req.node_id) {
4355 : // No-op case: we will still proceed to wait for reconciliation in case it is
4356 : // incomplete from an earlier update to the intent.
4357 0 : tracing::info!("Migrating: intent is unchanged {:?}", shard.intent);
4358 : } else {
4359 0 : let old_attached = *shard.intent.get_attached();
4360 0 :
4361 0 : match shard.policy {
4362 0 : PlacementPolicy::Attached(n) => {
4363 0 : // If our new attached node was a secondary, it no longer should be.
4364 0 : shard.intent.remove_secondary(scheduler, migrate_req.node_id);
4365 :
4366 : // If we were already attached to something, demote that to a secondary
4367 0 : if let Some(old_attached) = old_attached {
4368 0 : if n > 0 {
4369 : // Remove other secondaries to make room for the location we'll demote
4370 0 : while shard.intent.get_secondary().len() >= n {
4371 0 : shard.intent.pop_secondary(scheduler);
4372 0 : }
4373 :
4374 0 : shard.intent.push_secondary(scheduler, old_attached);
4375 0 : }
4376 0 : }
4377 :
4378 0 : shard.intent.set_attached(scheduler, Some(migrate_req.node_id));
4379 : }
4380 0 : PlacementPolicy::Secondary => {
4381 0 : shard.intent.clear(scheduler);
4382 0 : shard.intent.push_secondary(scheduler, migrate_req.node_id);
4383 0 : }
4384 : PlacementPolicy::Detached => {
4385 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
4386 0 : "Cannot migrate a tenant that is PlacementPolicy::Detached: configure it to an attached policy first"
4387 0 : )))
4388 : }
4389 : }
4390 :
4391 0 : tracing::info!("Migrating: new intent {:?}", shard.intent);
4392 0 : shard.sequence = shard.sequence.next();
4393 : }
4394 :
4395 0 : self.maybe_reconcile_shard(shard, nodes)
4396 : };
4397 :
4398 0 : if let Some(waiter) = waiter {
4399 0 : waiter.wait_timeout(RECONCILE_TIMEOUT).await?;
4400 : } else {
4401 0 : tracing::info!("Migration is a no-op");
4402 : }
4403 :
4404 0 : Ok(TenantShardMigrateResponse {})
4405 0 : }
4406 :
4407 : /// This is for debug/support only: we simply drop all state for a tenant, without
4408 : /// detaching or deleting it on pageservers.
4409 0 : pub(crate) async fn tenant_drop(&self, tenant_id: TenantId) -> Result<(), ApiError> {
4410 0 : self.persistence.delete_tenant(tenant_id).await?;
4411 :
4412 0 : let mut locked = self.inner.write().unwrap();
4413 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
4414 0 : let mut shards = Vec::new();
4415 0 : for (tenant_shard_id, _) in tenants.range(TenantShardId::tenant_range(tenant_id)) {
4416 0 : shards.push(*tenant_shard_id);
4417 0 : }
4418 :
4419 0 : for shard_id in shards {
4420 0 : if let Some(mut shard) = tenants.remove(&shard_id) {
4421 0 : shard.intent.clear(scheduler);
4422 0 : }
4423 : }
4424 :
4425 0 : Ok(())
4426 0 : }
4427 :
4428 : /// This is for debug/support only: assuming tenant data is already present in S3, we "create" a
4429 : /// tenant with a very high generation number so that it will see the existing data.
4430 0 : pub(crate) async fn tenant_import(
4431 0 : &self,
4432 0 : tenant_id: TenantId,
4433 0 : ) -> Result<TenantCreateResponse, ApiError> {
4434 0 : // Pick an arbitrary available pageserver to use for scanning the tenant in remote storage
4435 0 : let maybe_node = {
4436 0 : self.inner
4437 0 : .read()
4438 0 : .unwrap()
4439 0 : .nodes
4440 0 : .values()
4441 0 : .find(|n| n.is_available())
4442 0 : .cloned()
4443 : };
4444 0 : let Some(node) = maybe_node else {
4445 0 : return Err(ApiError::BadRequest(anyhow::anyhow!("No nodes available")));
4446 : };
4447 :
4448 0 : let client = PageserverClient::new(
4449 0 : node.get_id(),
4450 0 : node.base_url(),
4451 0 : self.config.jwt_token.as_deref(),
4452 0 : );
4453 :
4454 0 : let scan_result = client
4455 0 : .tenant_scan_remote_storage(tenant_id)
4456 0 : .await
4457 0 : .map_err(|e| passthrough_api_error(&node, e))?;
4458 :
4459 : // A post-split tenant may contain a mixture of shard counts in remote storage: pick the highest count.
4460 0 : let Some(shard_count) = scan_result
4461 0 : .shards
4462 0 : .iter()
4463 0 : .map(|s| s.tenant_shard_id.shard_count)
4464 0 : .max()
4465 : else {
4466 0 : return Err(ApiError::NotFound(
4467 0 : anyhow::anyhow!("No shards found").into(),
4468 0 : ));
4469 : };
4470 :
4471 : // Ideally we would set each newly imported shard's generation independently, but for correctness it is sufficient
4472 : // to
4473 0 : let generation = scan_result
4474 0 : .shards
4475 0 : .iter()
4476 0 : .map(|s| s.generation)
4477 0 : .max()
4478 0 : .expect("We already validated >0 shards");
4479 0 :
4480 0 : // FIXME: we have no way to recover the shard stripe size from contents of remote storage: this will
4481 0 : // only work if they were using the default stripe size.
4482 0 : let stripe_size = ShardParameters::DEFAULT_STRIPE_SIZE;
4483 :
4484 0 : let (response, waiters) = self
4485 0 : .do_tenant_create(TenantCreateRequest {
4486 0 : new_tenant_id: TenantShardId::unsharded(tenant_id),
4487 0 : generation,
4488 0 :
4489 0 : shard_parameters: ShardParameters {
4490 0 : count: shard_count,
4491 0 : stripe_size,
4492 0 : },
4493 0 : placement_policy: Some(PlacementPolicy::Attached(0)), // No secondaries, for convenient debug/hacking
4494 0 :
4495 0 : // There is no way to know what the tenant's config was: revert to defaults
4496 0 : //
4497 0 : // TODO: remove `switch_aux_file_policy` once we finish auxv2 migration
4498 0 : //
4499 0 : // we write to both v1+v2 storage, so that the test case can use either storage format for testing
4500 0 : config: TenantConfig {
4501 0 : switch_aux_file_policy: Some(models::AuxFilePolicy::CrossValidation),
4502 0 : ..TenantConfig::default()
4503 0 : },
4504 0 : })
4505 0 : .await?;
4506 :
4507 0 : if let Err(e) = self.await_waiters(waiters, SHORT_RECONCILE_TIMEOUT).await {
4508 : // Since this is a debug/support operation, all kinds of weird issues are possible (e.g. this
4509 : // tenant doesn't exist in the control plane), so don't fail the request if it can't fully
4510 : // reconcile, as reconciliation includes notifying compute.
4511 0 : tracing::warn!(%tenant_id, "Reconcile not done yet while importing tenant ({e})");
4512 0 : }
4513 :
4514 0 : Ok(response)
4515 0 : }
4516 :
4517 : /// For debug/support: a full JSON dump of TenantShards. Returns a response so that
4518 : /// we don't have to make TenantShard clonable in the return path.
4519 0 : pub(crate) fn tenants_dump(&self) -> Result<hyper::Response<hyper::Body>, ApiError> {
4520 0 : let serialized = {
4521 0 : let locked = self.inner.read().unwrap();
4522 0 : let result = locked.tenants.values().collect::<Vec<_>>();
4523 0 : serde_json::to_string(&result).map_err(|e| ApiError::InternalServerError(e.into()))?
4524 : };
4525 :
4526 0 : hyper::Response::builder()
4527 0 : .status(hyper::StatusCode::OK)
4528 0 : .header(hyper::header::CONTENT_TYPE, "application/json")
4529 0 : .body(hyper::Body::from(serialized))
4530 0 : .map_err(|e| ApiError::InternalServerError(e.into()))
4531 0 : }
4532 :
4533 : /// Check the consistency of in-memory state vs. persistent state, and check that the
4534 : /// scheduler's statistics are up to date.
4535 : ///
4536 : /// These consistency checks expect an **idle** system. If changes are going on while
4537 : /// we run, then we can falsely indicate a consistency issue. This is sufficient for end-of-test
4538 : /// checks, but not suitable for running continuously in the background in the field.
4539 0 : pub(crate) async fn consistency_check(&self) -> Result<(), ApiError> {
4540 0 : let (mut expect_nodes, mut expect_shards) = {
4541 0 : let locked = self.inner.read().unwrap();
4542 0 :
4543 0 : locked
4544 0 : .scheduler
4545 0 : .consistency_check(locked.nodes.values(), locked.tenants.values())
4546 0 : .context("Scheduler checks")
4547 0 : .map_err(ApiError::InternalServerError)?;
4548 :
4549 0 : let expect_nodes = locked
4550 0 : .nodes
4551 0 : .values()
4552 0 : .map(|n| n.to_persistent())
4553 0 : .collect::<Vec<_>>();
4554 0 :
4555 0 : let expect_shards = locked
4556 0 : .tenants
4557 0 : .values()
4558 0 : .map(|t| t.to_persistent())
4559 0 : .collect::<Vec<_>>();
4560 :
4561 : // This method can only validate the state of an idle system: if a reconcile is in
4562 : // progress, fail out early to avoid giving false errors on state that won't match
4563 : // between database and memory under a ReconcileResult is processed.
4564 0 : for t in locked.tenants.values() {
4565 0 : if t.reconciler.is_some() {
4566 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
4567 0 : "Shard {} reconciliation in progress",
4568 0 : t.tenant_shard_id
4569 0 : )));
4570 0 : }
4571 : }
4572 :
4573 0 : (expect_nodes, expect_shards)
4574 : };
4575 :
4576 0 : let mut nodes = self.persistence.list_nodes().await?;
4577 0 : expect_nodes.sort_by_key(|n| n.node_id);
4578 0 : nodes.sort_by_key(|n| n.node_id);
4579 0 :
4580 0 : if nodes != expect_nodes {
4581 0 : tracing::error!("Consistency check failed on nodes.");
4582 0 : tracing::error!(
4583 0 : "Nodes in memory: {}",
4584 0 : serde_json::to_string(&expect_nodes)
4585 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
4586 : );
4587 0 : tracing::error!(
4588 0 : "Nodes in database: {}",
4589 0 : serde_json::to_string(&nodes)
4590 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
4591 : );
4592 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
4593 0 : "Node consistency failure"
4594 0 : )));
4595 0 : }
4596 :
4597 0 : let mut shards = self.persistence.list_tenant_shards().await?;
4598 0 : shards.sort_by_key(|tsp| (tsp.tenant_id.clone(), tsp.shard_number, tsp.shard_count));
4599 0 : expect_shards.sort_by_key(|tsp| (tsp.tenant_id.clone(), tsp.shard_number, tsp.shard_count));
4600 0 :
4601 0 : if shards != expect_shards {
4602 0 : tracing::error!("Consistency check failed on shards.");
4603 0 : tracing::error!(
4604 0 : "Shards in memory: {}",
4605 0 : serde_json::to_string(&expect_shards)
4606 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
4607 : );
4608 0 : tracing::error!(
4609 0 : "Shards in database: {}",
4610 0 : serde_json::to_string(&shards)
4611 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
4612 : );
4613 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
4614 0 : "Shard consistency failure"
4615 0 : )));
4616 0 : }
4617 0 :
4618 0 : Ok(())
4619 0 : }
4620 :
4621 : /// For debug/support: a JSON dump of the [`Scheduler`]. Returns a response so that
4622 : /// we don't have to make TenantShard clonable in the return path.
4623 0 : pub(crate) fn scheduler_dump(&self) -> Result<hyper::Response<hyper::Body>, ApiError> {
4624 0 : let serialized = {
4625 0 : let locked = self.inner.read().unwrap();
4626 0 : serde_json::to_string(&locked.scheduler)
4627 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
4628 : };
4629 :
4630 0 : hyper::Response::builder()
4631 0 : .status(hyper::StatusCode::OK)
4632 0 : .header(hyper::header::CONTENT_TYPE, "application/json")
4633 0 : .body(hyper::Body::from(serialized))
4634 0 : .map_err(|e| ApiError::InternalServerError(e.into()))
4635 0 : }
4636 :
4637 : /// This is for debug/support only: we simply drop all state for a tenant, without
4638 : /// detaching or deleting it on pageservers. We do not try and re-schedule any
4639 : /// tenants that were on this node.
4640 0 : pub(crate) async fn node_drop(&self, node_id: NodeId) -> Result<(), ApiError> {
4641 0 : self.persistence.delete_node(node_id).await?;
4642 :
4643 0 : let mut locked = self.inner.write().unwrap();
4644 :
4645 0 : for shard in locked.tenants.values_mut() {
4646 0 : shard.deref_node(node_id);
4647 0 : shard.observed.locations.remove(&node_id);
4648 0 : }
4649 :
4650 0 : let mut nodes = (*locked.nodes).clone();
4651 0 : nodes.remove(&node_id);
4652 0 : locked.nodes = Arc::new(nodes);
4653 0 :
4654 0 : locked.scheduler.node_remove(node_id);
4655 0 :
4656 0 : Ok(())
4657 0 : }
4658 :
4659 : /// If a node has any work on it, it will be rescheduled: this is "clean" in the sense
4660 : /// that we don't leave any bad state behind in the storage controller, but unclean
4661 : /// in the sense that we are not carefully draining the node.
4662 0 : pub(crate) async fn node_delete(&self, node_id: NodeId) -> Result<(), ApiError> {
4663 0 : let _node_lock =
4664 0 : trace_exclusive_lock(&self.node_op_locks, node_id, NodeOperations::Delete).await;
4665 :
4666 : // 1. Atomically update in-memory state:
4667 : // - set the scheduling state to Pause to make subsequent scheduling ops skip it
4668 : // - update shards' intents to exclude the node, and reschedule any shards whose intents we modified.
4669 : // - drop the node from the main nodes map, so that when running reconciles complete they do not
4670 : // re-insert references to this node into the ObservedState of shards
4671 : // - drop the node from the scheduler
4672 : {
4673 0 : let mut locked = self.inner.write().unwrap();
4674 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
4675 0 :
4676 0 : {
4677 0 : let mut nodes_mut = (*nodes).deref().clone();
4678 0 : match nodes_mut.get_mut(&node_id) {
4679 0 : Some(node) => {
4680 0 : // We do not bother setting this in the database, because we're about to delete the row anyway, and
4681 0 : // if we crash it would not be desirable to leave the node paused after a restart.
4682 0 : node.set_scheduling(NodeSchedulingPolicy::Pause);
4683 0 : }
4684 : None => {
4685 0 : tracing::info!(
4686 0 : "Node not found: presuming this is a retry and returning success"
4687 : );
4688 0 : return Ok(());
4689 : }
4690 : }
4691 :
4692 0 : *nodes = Arc::new(nodes_mut);
4693 : }
4694 :
4695 0 : for (tenant_shard_id, shard) in tenants {
4696 0 : if shard.deref_node(node_id) {
4697 : // FIXME: we need to build a ScheduleContext that reflects this shard's peers, otherwise
4698 : // it won't properly do anti-affinity.
4699 0 : let mut schedule_context = ScheduleContext::default();
4700 :
4701 0 : if let Err(e) = shard.schedule(scheduler, &mut schedule_context) {
4702 : // TODO: implement force flag to remove a node even if we can't reschedule
4703 : // a tenant
4704 0 : tracing::error!("Refusing to delete node, shard {tenant_shard_id} can't be rescheduled: {e}");
4705 0 : return Err(e.into());
4706 : } else {
4707 0 : tracing::info!(
4708 0 : "Rescheduled shard {tenant_shard_id} away from node during deletion"
4709 : )
4710 : }
4711 :
4712 0 : self.maybe_reconcile_shard(shard, nodes);
4713 0 : }
4714 :
4715 : // Here we remove an existing observed location for the node we're removing, and it will
4716 : // not be re-added by a reconciler's completion because we filter out removed nodes in
4717 : // process_result.
4718 : //
4719 : // Note that we update the shard's observed state _after_ calling maybe_reconcile_shard: that
4720 : // means any reconciles we spawned will know about the node we're deleting, enabling them
4721 : // to do live migrations if it's still online.
4722 0 : shard.observed.locations.remove(&node_id);
4723 : }
4724 :
4725 0 : scheduler.node_remove(node_id);
4726 0 :
4727 0 : {
4728 0 : let mut nodes_mut = (**nodes).clone();
4729 0 : nodes_mut.remove(&node_id);
4730 0 : *nodes = Arc::new(nodes_mut);
4731 0 : }
4732 0 : }
4733 0 :
4734 0 : // Note: some `generation_pageserver` columns on tenant shards in the database may still refer to
4735 0 : // the removed node, as this column means "The pageserver to which this generation was issued", and
4736 0 : // their generations won't get updated until the reconcilers moving them away from this node complete.
4737 0 : // That is safe because in Service::spawn we only use generation_pageserver if it refers to a node
4738 0 : // that exists.
4739 0 :
4740 0 : // 2. Actually delete the node from the database and from in-memory state
4741 0 : tracing::info!("Deleting node from database");
4742 0 : self.persistence.delete_node(node_id).await?;
4743 :
4744 0 : Ok(())
4745 0 : }
4746 :
4747 0 : pub(crate) async fn node_list(&self) -> Result<Vec<Node>, ApiError> {
4748 0 : let nodes = {
4749 0 : self.inner
4750 0 : .read()
4751 0 : .unwrap()
4752 0 : .nodes
4753 0 : .values()
4754 0 : .cloned()
4755 0 : .collect::<Vec<_>>()
4756 0 : };
4757 0 :
4758 0 : Ok(nodes)
4759 0 : }
4760 :
4761 0 : pub(crate) async fn get_node(&self, node_id: NodeId) -> Result<Node, ApiError> {
4762 0 : self.inner
4763 0 : .read()
4764 0 : .unwrap()
4765 0 : .nodes
4766 0 : .get(&node_id)
4767 0 : .cloned()
4768 0 : .ok_or(ApiError::NotFound(
4769 0 : format!("Node {node_id} not registered").into(),
4770 0 : ))
4771 0 : }
4772 :
4773 0 : pub(crate) async fn get_leader(&self) -> DatabaseResult<Option<ControllerPersistence>> {
4774 0 : self.persistence.get_leader().await
4775 0 : }
4776 :
4777 0 : pub(crate) async fn node_register(
4778 0 : &self,
4779 0 : register_req: NodeRegisterRequest,
4780 0 : ) -> Result<(), ApiError> {
4781 0 : let _node_lock = trace_exclusive_lock(
4782 0 : &self.node_op_locks,
4783 0 : register_req.node_id,
4784 0 : NodeOperations::Register,
4785 0 : )
4786 0 : .await;
4787 :
4788 0 : if register_req.availability_zone_id.is_none() {
4789 0 : tracing::warn!(
4790 0 : "Node {} registering without specific availability zone id",
4791 : register_req.node_id
4792 : );
4793 0 : }
4794 :
4795 : enum RegistrationStatus {
4796 : Matched(Node),
4797 : Mismatched,
4798 : New,
4799 : }
4800 :
4801 0 : let registration_status = {
4802 0 : let locked = self.inner.read().unwrap();
4803 0 : if let Some(node) = locked.nodes.get(®ister_req.node_id) {
4804 0 : if node.registration_match(®ister_req) {
4805 0 : RegistrationStatus::Matched(node.clone())
4806 : } else {
4807 0 : RegistrationStatus::Mismatched
4808 : }
4809 : } else {
4810 0 : RegistrationStatus::New
4811 : }
4812 : };
4813 :
4814 0 : match registration_status {
4815 0 : RegistrationStatus::Matched(node) => {
4816 0 : tracing::info!(
4817 0 : "Node {} re-registered with matching address",
4818 : register_req.node_id
4819 : );
4820 :
4821 0 : if node.get_availability_zone_id().is_none() {
4822 0 : if let Some(az_id) = register_req.availability_zone_id.clone() {
4823 0 : tracing::info!("Extracting availability zone id from registration request for node {}: {}",
4824 : register_req.node_id, az_id);
4825 :
4826 : // Persist to the database and update in memory state. See comment below
4827 : // on ordering.
4828 0 : self.persistence
4829 0 : .set_node_availability_zone_id(register_req.node_id, az_id)
4830 0 : .await?;
4831 0 : let node_with_az = Node::new(
4832 0 : register_req.node_id,
4833 0 : register_req.listen_http_addr,
4834 0 : register_req.listen_http_port,
4835 0 : register_req.listen_pg_addr,
4836 0 : register_req.listen_pg_port,
4837 0 : register_req.availability_zone_id,
4838 0 : );
4839 0 :
4840 0 : let mut locked = self.inner.write().unwrap();
4841 0 : let mut new_nodes = (*locked.nodes).clone();
4842 0 :
4843 0 : locked.scheduler.node_upsert(&node_with_az);
4844 0 : new_nodes.insert(register_req.node_id, node_with_az);
4845 0 :
4846 0 : locked.nodes = Arc::new(new_nodes);
4847 0 : }
4848 0 : }
4849 :
4850 0 : return Ok(());
4851 : }
4852 : RegistrationStatus::Mismatched => {
4853 : // TODO: decide if we want to allow modifying node addresses without removing and re-adding
4854 : // the node. Safest/simplest thing is to refuse it, and usually we deploy with
4855 : // a fixed address through the lifetime of a node.
4856 0 : tracing::warn!(
4857 0 : "Node {} tried to register with different address",
4858 : register_req.node_id
4859 : );
4860 0 : return Err(ApiError::Conflict(
4861 0 : "Node is already registered with different address".to_string(),
4862 0 : ));
4863 : }
4864 0 : RegistrationStatus::New => {
4865 0 : // fallthrough
4866 0 : }
4867 0 : }
4868 0 :
4869 0 : // We do not require that a node is actually online when registered (it will start life
4870 0 : // with it's availability set to Offline), but we _do_ require that its DNS record exists. We're
4871 0 : // therefore not immune to asymmetric L3 connectivity issues, but we are protected against nodes
4872 0 : // that register themselves with a broken DNS config. We check only the HTTP hostname, because
4873 0 : // the postgres hostname might only be resolvable to clients (e.g. if we're on a different VPC than clients).
4874 0 : if tokio::net::lookup_host(format!(
4875 0 : "{}:{}",
4876 0 : register_req.listen_http_addr, register_req.listen_http_port
4877 0 : ))
4878 0 : .await
4879 0 : .is_err()
4880 : {
4881 : // If we have a transient DNS issue, it's up to the caller to retry their registration. Because
4882 : // we can't robustly distinguish between an intermittent issue and a totally bogus DNS situation,
4883 : // we return a soft 503 error, to encourage callers to retry past transient issues.
4884 0 : return Err(ApiError::ResourceUnavailable(
4885 0 : format!(
4886 0 : "Node {} tried to register with unknown DNS name '{}'",
4887 0 : register_req.node_id, register_req.listen_http_addr
4888 0 : )
4889 0 : .into(),
4890 0 : ));
4891 0 : }
4892 0 :
4893 0 : // Ordering: we must persist the new node _before_ adding it to in-memory state.
4894 0 : // This ensures that before we use it for anything or expose it via any external
4895 0 : // API, it is guaranteed to be available after a restart.
4896 0 : let new_node = Node::new(
4897 0 : register_req.node_id,
4898 0 : register_req.listen_http_addr,
4899 0 : register_req.listen_http_port,
4900 0 : register_req.listen_pg_addr,
4901 0 : register_req.listen_pg_port,
4902 0 : register_req.availability_zone_id,
4903 0 : );
4904 0 :
4905 0 : // TODO: idempotency if the node already exists in the database
4906 0 : self.persistence.insert_node(&new_node).await?;
4907 :
4908 0 : let mut locked = self.inner.write().unwrap();
4909 0 : let mut new_nodes = (*locked.nodes).clone();
4910 0 :
4911 0 : locked.scheduler.node_upsert(&new_node);
4912 0 : new_nodes.insert(register_req.node_id, new_node);
4913 0 :
4914 0 : locked.nodes = Arc::new(new_nodes);
4915 0 :
4916 0 : tracing::info!(
4917 0 : "Registered pageserver {}, now have {} pageservers",
4918 0 : register_req.node_id,
4919 0 : locked.nodes.len()
4920 : );
4921 0 : Ok(())
4922 0 : }
4923 :
4924 0 : pub(crate) async fn node_configure(
4925 0 : &self,
4926 0 : node_id: NodeId,
4927 0 : availability: Option<NodeAvailability>,
4928 0 : scheduling: Option<NodeSchedulingPolicy>,
4929 0 : ) -> Result<(), ApiError> {
4930 0 : let _node_lock =
4931 0 : trace_exclusive_lock(&self.node_op_locks, node_id, NodeOperations::Configure).await;
4932 :
4933 0 : if let Some(scheduling) = scheduling {
4934 : // Scheduling is a persistent part of Node: we must write updates to the database before
4935 : // applying them in memory
4936 0 : self.persistence.update_node(node_id, scheduling).await?;
4937 0 : }
4938 :
4939 : // If we're activating a node, then before setting it active we must reconcile any shard locations
4940 : // on that node, in case it is out of sync, e.g. due to being unavailable during controller startup,
4941 : // by calling [`Self::node_activate_reconcile`]
4942 : //
4943 : // The transition we calculate here remains valid later in the function because we hold the op lock on the node:
4944 : // nothing else can mutate its availability while we run.
4945 0 : let availability_transition = if let Some(input_availability) = availability.as_ref() {
4946 0 : let (activate_node, availability_transition) = {
4947 0 : let locked = self.inner.read().unwrap();
4948 0 : let Some(node) = locked.nodes.get(&node_id) else {
4949 0 : return Err(ApiError::NotFound(
4950 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
4951 0 : ));
4952 : };
4953 :
4954 0 : (
4955 0 : node.clone(),
4956 0 : node.get_availability_transition(input_availability),
4957 0 : )
4958 : };
4959 :
4960 0 : if matches!(availability_transition, AvailabilityTransition::ToActive) {
4961 0 : self.node_activate_reconcile(activate_node, &_node_lock)
4962 0 : .await?;
4963 0 : }
4964 0 : availability_transition
4965 : } else {
4966 0 : AvailabilityTransition::Unchanged
4967 : };
4968 :
4969 : // Apply changes from the request to our in-memory state for the Node
4970 0 : let mut locked = self.inner.write().unwrap();
4971 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
4972 0 :
4973 0 : let mut new_nodes = (**nodes).clone();
4974 :
4975 0 : let Some(node) = new_nodes.get_mut(&node_id) else {
4976 0 : return Err(ApiError::NotFound(
4977 0 : anyhow::anyhow!("Node not registered").into(),
4978 0 : ));
4979 : };
4980 :
4981 0 : if let Some(availability) = availability.as_ref() {
4982 0 : node.set_availability(availability.clone());
4983 0 : }
4984 :
4985 0 : if let Some(scheduling) = scheduling {
4986 0 : node.set_scheduling(scheduling);
4987 0 : }
4988 :
4989 : // Update the scheduler, in case the elegibility of the node for new shards has changed
4990 0 : scheduler.node_upsert(node);
4991 0 :
4992 0 : let new_nodes = Arc::new(new_nodes);
4993 0 :
4994 0 : // Modify scheduling state for any Tenants that are affected by a change in the node's availability state.
4995 0 : match availability_transition {
4996 : AvailabilityTransition::ToOffline => {
4997 0 : tracing::info!("Node {} transition to offline", node_id);
4998 0 : let mut tenants_affected: usize = 0;
4999 :
5000 0 : for (tenant_shard_id, tenant_shard) in tenants {
5001 0 : if let Some(observed_loc) = tenant_shard.observed.locations.get_mut(&node_id) {
5002 0 : // When a node goes offline, we set its observed configuration to None, indicating unknown: we will
5003 0 : // not assume our knowledge of the node's configuration is accurate until it comes back online
5004 0 : observed_loc.conf = None;
5005 0 : }
5006 :
5007 0 : if new_nodes.len() == 1 {
5008 : // Special case for single-node cluster: there is no point trying to reschedule
5009 : // any tenant shards: avoid doing so, in order to avoid spewing warnings about
5010 : // failures to schedule them.
5011 0 : continue;
5012 0 : }
5013 0 :
5014 0 : if !new_nodes
5015 0 : .values()
5016 0 : .any(|n| matches!(n.may_schedule(), MaySchedule::Yes(_)))
5017 : {
5018 : // Special case for when all nodes are unavailable and/or unschedulable: there is no point
5019 : // trying to reschedule since there's nowhere else to go. Without this
5020 : // branch we incorrectly detach tenants in response to node unavailability.
5021 0 : continue;
5022 0 : }
5023 0 :
5024 0 : if tenant_shard.intent.demote_attached(scheduler, node_id) {
5025 0 : tenant_shard.sequence = tenant_shard.sequence.next();
5026 0 :
5027 0 : // TODO: populate a ScheduleContext including all shards in the same tenant_id (only matters
5028 0 : // for tenants without secondary locations: if they have a secondary location, then this
5029 0 : // schedule() call is just promoting an existing secondary)
5030 0 : let mut schedule_context = ScheduleContext::default();
5031 0 :
5032 0 : match tenant_shard.schedule(scheduler, &mut schedule_context) {
5033 0 : Err(e) => {
5034 0 : // It is possible that some tenants will become unschedulable when too many pageservers
5035 0 : // go offline: in this case there isn't much we can do other than make the issue observable.
5036 0 : // TODO: give TenantShard a scheduling error attribute to be queried later.
5037 0 : tracing::warn!(%tenant_shard_id, "Scheduling error when marking pageserver {} offline: {e}", node_id);
5038 : }
5039 : Ok(()) => {
5040 0 : if self
5041 0 : .maybe_reconcile_shard(tenant_shard, &new_nodes)
5042 0 : .is_some()
5043 0 : {
5044 0 : tenants_affected += 1;
5045 0 : };
5046 : }
5047 : }
5048 0 : }
5049 : }
5050 0 : tracing::info!(
5051 0 : "Launched {} reconciler tasks for tenants affected by node {} going offline",
5052 : tenants_affected,
5053 : node_id
5054 : )
5055 : }
5056 : AvailabilityTransition::ToActive => {
5057 0 : tracing::info!("Node {} transition to active", node_id);
5058 : // When a node comes back online, we must reconcile any tenant that has a None observed
5059 : // location on the node.
5060 0 : for tenant_shard in locked.tenants.values_mut() {
5061 : // If a reconciliation is already in progress, rely on the previous scheduling
5062 : // decision and skip triggering a new reconciliation.
5063 0 : if tenant_shard.reconciler.is_some() {
5064 0 : continue;
5065 0 : }
5066 :
5067 0 : if let Some(observed_loc) = tenant_shard.observed.locations.get_mut(&node_id) {
5068 0 : if observed_loc.conf.is_none() {
5069 0 : self.maybe_reconcile_shard(tenant_shard, &new_nodes);
5070 0 : }
5071 0 : }
5072 : }
5073 :
5074 : // TODO: in the background, we should balance work back onto this pageserver
5075 : }
5076 : // No action required for the intermediate unavailable state.
5077 : // When we transition into active or offline from the unavailable state,
5078 : // the correct handling above will kick in.
5079 : AvailabilityTransition::ToWarmingUpFromActive => {
5080 0 : tracing::info!("Node {} transition to unavailable from active", node_id);
5081 : }
5082 : AvailabilityTransition::ToWarmingUpFromOffline => {
5083 0 : tracing::info!("Node {} transition to unavailable from offline", node_id);
5084 : }
5085 : AvailabilityTransition::Unchanged => {
5086 0 : tracing::debug!("Node {} no availability change during config", node_id);
5087 : }
5088 : }
5089 :
5090 0 : locked.nodes = new_nodes;
5091 0 :
5092 0 : Ok(())
5093 0 : }
5094 :
5095 : /// Wrapper around [`Self::node_configure`] which only allows changes while there is no ongoing
5096 : /// operation for HTTP api.
5097 0 : pub(crate) async fn external_node_configure(
5098 0 : &self,
5099 0 : node_id: NodeId,
5100 0 : availability: Option<NodeAvailability>,
5101 0 : scheduling: Option<NodeSchedulingPolicy>,
5102 0 : ) -> Result<(), ApiError> {
5103 0 : {
5104 0 : let locked = self.inner.read().unwrap();
5105 0 : if let Some(op) = locked.ongoing_operation.as_ref().map(|op| op.operation) {
5106 0 : return Err(ApiError::PreconditionFailed(
5107 0 : format!("Ongoing background operation forbids configuring: {op}").into(),
5108 0 : ));
5109 0 : }
5110 0 : }
5111 0 :
5112 0 : self.node_configure(node_id, availability, scheduling).await
5113 0 : }
5114 :
5115 0 : pub(crate) async fn start_node_drain(
5116 0 : self: &Arc<Self>,
5117 0 : node_id: NodeId,
5118 0 : ) -> Result<(), ApiError> {
5119 0 : let (ongoing_op, node_available, node_policy, schedulable_nodes_count) = {
5120 0 : let locked = self.inner.read().unwrap();
5121 0 : let nodes = &locked.nodes;
5122 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
5123 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
5124 0 : ))?;
5125 0 : let schedulable_nodes_count = nodes
5126 0 : .iter()
5127 0 : .filter(|(_, n)| matches!(n.may_schedule(), MaySchedule::Yes(_)))
5128 0 : .count();
5129 0 :
5130 0 : (
5131 0 : locked
5132 0 : .ongoing_operation
5133 0 : .as_ref()
5134 0 : .map(|ongoing| ongoing.operation),
5135 0 : node.is_available(),
5136 0 : node.get_scheduling(),
5137 0 : schedulable_nodes_count,
5138 0 : )
5139 0 : };
5140 :
5141 0 : if let Some(ongoing) = ongoing_op {
5142 0 : return Err(ApiError::PreconditionFailed(
5143 0 : format!("Background operation already ongoing for node: {}", ongoing).into(),
5144 0 : ));
5145 0 : }
5146 0 :
5147 0 : if !node_available {
5148 0 : return Err(ApiError::ResourceUnavailable(
5149 0 : format!("Node {node_id} is currently unavailable").into(),
5150 0 : ));
5151 0 : }
5152 0 :
5153 0 : if schedulable_nodes_count == 0 {
5154 0 : return Err(ApiError::PreconditionFailed(
5155 0 : "No other schedulable nodes to drain to".into(),
5156 0 : ));
5157 0 : }
5158 0 :
5159 0 : match node_policy {
5160 : NodeSchedulingPolicy::Active | NodeSchedulingPolicy::Pause => {
5161 0 : self.node_configure(node_id, None, Some(NodeSchedulingPolicy::Draining))
5162 0 : .await?;
5163 :
5164 0 : let cancel = self.cancel.child_token();
5165 0 : let gate_guard = self.gate.enter().map_err(|_| ApiError::ShuttingDown)?;
5166 :
5167 0 : self.inner.write().unwrap().ongoing_operation = Some(OperationHandler {
5168 0 : operation: Operation::Drain(Drain { node_id }),
5169 0 : cancel: cancel.clone(),
5170 0 : });
5171 :
5172 0 : let span = tracing::info_span!(parent: None, "drain_node", %node_id);
5173 :
5174 0 : tokio::task::spawn({
5175 0 : let service = self.clone();
5176 0 : let cancel = cancel.clone();
5177 0 : async move {
5178 0 : let _gate_guard = gate_guard;
5179 :
5180 : scopeguard::defer! {
5181 : let prev = service.inner.write().unwrap().ongoing_operation.take();
5182 :
5183 0 : if let Some(Operation::Drain(removed_drain)) = prev.map(|h| h.operation) {
5184 : assert_eq!(removed_drain.node_id, node_id, "We always take the same operation");
5185 : } else {
5186 : panic!("We always remove the same operation")
5187 : }
5188 : }
5189 :
5190 0 : tracing::info!("Drain background operation starting");
5191 0 : let res = service.drain_node(node_id, cancel).await;
5192 0 : match res {
5193 : Ok(()) => {
5194 0 : tracing::info!("Drain background operation completed successfully");
5195 : }
5196 : Err(OperationError::Cancelled) => {
5197 0 : tracing::info!("Drain background operation was cancelled");
5198 : }
5199 0 : Err(err) => {
5200 0 : tracing::error!("Drain background operation encountered: {err}")
5201 : }
5202 : }
5203 0 : }
5204 0 : }.instrument(span));
5205 0 : }
5206 : NodeSchedulingPolicy::Draining => {
5207 0 : return Err(ApiError::Conflict(format!(
5208 0 : "Node {node_id} has drain in progress"
5209 0 : )));
5210 : }
5211 0 : policy => {
5212 0 : return Err(ApiError::PreconditionFailed(
5213 0 : format!("Node {node_id} cannot be drained due to {policy:?} policy").into(),
5214 0 : ));
5215 : }
5216 : }
5217 :
5218 0 : Ok(())
5219 0 : }
5220 :
5221 0 : pub(crate) async fn cancel_node_drain(&self, node_id: NodeId) -> Result<(), ApiError> {
5222 0 : let node_available = {
5223 0 : let locked = self.inner.read().unwrap();
5224 0 : let nodes = &locked.nodes;
5225 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
5226 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
5227 0 : ))?;
5228 :
5229 0 : node.is_available()
5230 0 : };
5231 0 :
5232 0 : if !node_available {
5233 0 : return Err(ApiError::ResourceUnavailable(
5234 0 : format!("Node {node_id} is currently unavailable").into(),
5235 0 : ));
5236 0 : }
5237 :
5238 0 : if let Some(op_handler) = self.inner.read().unwrap().ongoing_operation.as_ref() {
5239 0 : if let Operation::Drain(drain) = op_handler.operation {
5240 0 : if drain.node_id == node_id {
5241 0 : tracing::info!("Cancelling background drain operation for node {node_id}");
5242 0 : op_handler.cancel.cancel();
5243 0 : return Ok(());
5244 0 : }
5245 0 : }
5246 0 : }
5247 :
5248 0 : Err(ApiError::PreconditionFailed(
5249 0 : format!("Node {node_id} has no drain in progress").into(),
5250 0 : ))
5251 0 : }
5252 :
5253 0 : pub(crate) async fn start_node_fill(self: &Arc<Self>, node_id: NodeId) -> Result<(), ApiError> {
5254 0 : let (ongoing_op, node_available, node_policy, total_nodes_count) = {
5255 0 : let locked = self.inner.read().unwrap();
5256 0 : let nodes = &locked.nodes;
5257 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
5258 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
5259 0 : ))?;
5260 :
5261 0 : (
5262 0 : locked
5263 0 : .ongoing_operation
5264 0 : .as_ref()
5265 0 : .map(|ongoing| ongoing.operation),
5266 0 : node.is_available(),
5267 0 : node.get_scheduling(),
5268 0 : nodes.len(),
5269 0 : )
5270 0 : };
5271 :
5272 0 : if let Some(ongoing) = ongoing_op {
5273 0 : return Err(ApiError::PreconditionFailed(
5274 0 : format!("Background operation already ongoing for node: {}", ongoing).into(),
5275 0 : ));
5276 0 : }
5277 0 :
5278 0 : if !node_available {
5279 0 : return Err(ApiError::ResourceUnavailable(
5280 0 : format!("Node {node_id} is currently unavailable").into(),
5281 0 : ));
5282 0 : }
5283 0 :
5284 0 : if total_nodes_count <= 1 {
5285 0 : return Err(ApiError::PreconditionFailed(
5286 0 : "No other nodes to fill from".into(),
5287 0 : ));
5288 0 : }
5289 0 :
5290 0 : match node_policy {
5291 : NodeSchedulingPolicy::Active => {
5292 0 : self.node_configure(node_id, None, Some(NodeSchedulingPolicy::Filling))
5293 0 : .await?;
5294 :
5295 0 : let cancel = self.cancel.child_token();
5296 0 : let gate_guard = self.gate.enter().map_err(|_| ApiError::ShuttingDown)?;
5297 :
5298 0 : self.inner.write().unwrap().ongoing_operation = Some(OperationHandler {
5299 0 : operation: Operation::Fill(Fill { node_id }),
5300 0 : cancel: cancel.clone(),
5301 0 : });
5302 :
5303 0 : let span = tracing::info_span!(parent: None, "fill_node", %node_id);
5304 :
5305 0 : tokio::task::spawn({
5306 0 : let service = self.clone();
5307 0 : let cancel = cancel.clone();
5308 0 : async move {
5309 0 : let _gate_guard = gate_guard;
5310 :
5311 : scopeguard::defer! {
5312 : let prev = service.inner.write().unwrap().ongoing_operation.take();
5313 :
5314 0 : if let Some(Operation::Fill(removed_fill)) = prev.map(|h| h.operation) {
5315 : assert_eq!(removed_fill.node_id, node_id, "We always take the same operation");
5316 : } else {
5317 : panic!("We always remove the same operation")
5318 : }
5319 : }
5320 :
5321 0 : tracing::info!("Fill background operation starting");
5322 0 : let res = service.fill_node(node_id, cancel).await;
5323 0 : match res {
5324 : Ok(()) => {
5325 0 : tracing::info!("Fill background operation completed successfully");
5326 : }
5327 : Err(OperationError::Cancelled) => {
5328 0 : tracing::info!("Fill background operation was cancelled");
5329 : }
5330 0 : Err(err) => {
5331 0 : tracing::error!("Fill background operation encountered: {err}")
5332 : }
5333 : }
5334 0 : }
5335 0 : }.instrument(span));
5336 0 : }
5337 : NodeSchedulingPolicy::Filling => {
5338 0 : return Err(ApiError::Conflict(format!(
5339 0 : "Node {node_id} has fill in progress"
5340 0 : )));
5341 : }
5342 0 : policy => {
5343 0 : return Err(ApiError::PreconditionFailed(
5344 0 : format!("Node {node_id} cannot be filled due to {policy:?} policy").into(),
5345 0 : ));
5346 : }
5347 : }
5348 :
5349 0 : Ok(())
5350 0 : }
5351 :
5352 0 : pub(crate) async fn cancel_node_fill(&self, node_id: NodeId) -> Result<(), ApiError> {
5353 0 : let node_available = {
5354 0 : let locked = self.inner.read().unwrap();
5355 0 : let nodes = &locked.nodes;
5356 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
5357 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
5358 0 : ))?;
5359 :
5360 0 : node.is_available()
5361 0 : };
5362 0 :
5363 0 : if !node_available {
5364 0 : return Err(ApiError::ResourceUnavailable(
5365 0 : format!("Node {node_id} is currently unavailable").into(),
5366 0 : ));
5367 0 : }
5368 :
5369 0 : if let Some(op_handler) = self.inner.read().unwrap().ongoing_operation.as_ref() {
5370 0 : if let Operation::Fill(fill) = op_handler.operation {
5371 0 : if fill.node_id == node_id {
5372 0 : tracing::info!("Cancelling background drain operation for node {node_id}");
5373 0 : op_handler.cancel.cancel();
5374 0 : return Ok(());
5375 0 : }
5376 0 : }
5377 0 : }
5378 :
5379 0 : Err(ApiError::PreconditionFailed(
5380 0 : format!("Node {node_id} has no fill in progress").into(),
5381 0 : ))
5382 0 : }
5383 :
5384 : /// Like [`Self::maybe_configured_reconcile_shard`], but uses the default reconciler
5385 : /// configuration
5386 0 : fn maybe_reconcile_shard(
5387 0 : &self,
5388 0 : shard: &mut TenantShard,
5389 0 : nodes: &Arc<HashMap<NodeId, Node>>,
5390 0 : ) -> Option<ReconcilerWaiter> {
5391 0 : self.maybe_configured_reconcile_shard(shard, nodes, ReconcilerConfig::default())
5392 0 : }
5393 :
5394 : /// Wrap [`TenantShard`] reconciliation methods with acquisition of [`Gate`] and [`ReconcileUnits`],
5395 0 : fn maybe_configured_reconcile_shard(
5396 0 : &self,
5397 0 : shard: &mut TenantShard,
5398 0 : nodes: &Arc<HashMap<NodeId, Node>>,
5399 0 : reconciler_config: ReconcilerConfig,
5400 0 : ) -> Option<ReconcilerWaiter> {
5401 0 : let reconcile_needed = shard.get_reconcile_needed(nodes);
5402 0 :
5403 0 : match reconcile_needed {
5404 0 : ReconcileNeeded::No => return None,
5405 0 : ReconcileNeeded::WaitExisting(waiter) => return Some(waiter),
5406 0 : ReconcileNeeded::Yes => {
5407 0 : // Fall through to try and acquire units for spawning reconciler
5408 0 : }
5409 : };
5410 :
5411 0 : let units = match self.reconciler_concurrency.clone().try_acquire_owned() {
5412 0 : Ok(u) => ReconcileUnits::new(u),
5413 : Err(_) => {
5414 0 : tracing::info!(tenant_id=%shard.tenant_shard_id.tenant_id, shard_id=%shard.tenant_shard_id.shard_slug(),
5415 0 : "Concurrency limited: enqueued for reconcile later");
5416 0 : if !shard.delayed_reconcile {
5417 0 : match self.delayed_reconcile_tx.try_send(shard.tenant_shard_id) {
5418 0 : Err(TrySendError::Closed(_)) => {
5419 0 : // Weird mid-shutdown case?
5420 0 : }
5421 : Err(TrySendError::Full(_)) => {
5422 : // It is safe to skip sending our ID in the channel: we will eventually get retried by the background reconcile task.
5423 0 : tracing::warn!(
5424 0 : "Many shards are waiting to reconcile: delayed_reconcile queue is full"
5425 : );
5426 : }
5427 0 : Ok(()) => {
5428 0 : shard.delayed_reconcile = true;
5429 0 : }
5430 : }
5431 0 : }
5432 :
5433 : // We won't spawn a reconciler, but we will construct a waiter that waits for the shard's sequence
5434 : // number to advance. When this function is eventually called again and succeeds in getting units,
5435 : // it will spawn a reconciler that makes this waiter complete.
5436 0 : return Some(shard.future_reconcile_waiter());
5437 : }
5438 : };
5439 :
5440 0 : let Ok(gate_guard) = self.reconcilers_gate.enter() else {
5441 : // Gate closed: we're shutting down, drop out.
5442 0 : return None;
5443 : };
5444 :
5445 0 : shard.spawn_reconciler(
5446 0 : &self.result_tx,
5447 0 : nodes,
5448 0 : &self.compute_hook,
5449 0 : reconciler_config,
5450 0 : &self.config,
5451 0 : &self.persistence,
5452 0 : units,
5453 0 : gate_guard,
5454 0 : &self.reconcilers_cancel,
5455 0 : )
5456 0 : }
5457 :
5458 : /// Check all tenants for pending reconciliation work, and reconcile those in need.
5459 : /// Additionally, reschedule tenants that require it.
5460 : ///
5461 : /// Returns how many reconciliation tasks were started, or `1` if no reconciles were
5462 : /// spawned but some _would_ have been spawned if `reconciler_concurrency` units where
5463 : /// available. A return value of 0 indicates that everything is fully reconciled already.
5464 0 : fn reconcile_all(&self) -> usize {
5465 0 : let mut locked = self.inner.write().unwrap();
5466 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
5467 0 : let pageservers = nodes.clone();
5468 0 :
5469 0 : let mut schedule_context = ScheduleContext::default();
5470 0 :
5471 0 : let mut reconciles_spawned = 0;
5472 0 : for (tenant_shard_id, shard) in tenants.iter_mut() {
5473 0 : if tenant_shard_id.is_shard_zero() {
5474 0 : schedule_context = ScheduleContext::default();
5475 0 : }
5476 :
5477 : // Skip checking if this shard is already enqueued for reconciliation
5478 0 : if shard.delayed_reconcile && self.reconciler_concurrency.available_permits() == 0 {
5479 : // If there is something delayed, then return a nonzero count so that
5480 : // callers like reconcile_all_now do not incorrectly get the impression
5481 : // that the system is in a quiescent state.
5482 0 : reconciles_spawned = std::cmp::max(1, reconciles_spawned);
5483 0 : continue;
5484 0 : }
5485 0 :
5486 0 : // Eventual consistency: if an earlier reconcile job failed, and the shard is still
5487 0 : // dirty, spawn another rone
5488 0 : if self.maybe_reconcile_shard(shard, &pageservers).is_some() {
5489 0 : reconciles_spawned += 1;
5490 0 : }
5491 :
5492 0 : schedule_context.avoid(&shard.intent.all_pageservers());
5493 : }
5494 :
5495 0 : reconciles_spawned
5496 0 : }
5497 :
5498 : /// `optimize` in this context means identifying shards which have valid scheduled locations, but
5499 : /// could be scheduled somewhere better:
5500 : /// - Cutting over to a secondary if the node with the secondary is more lightly loaded
5501 : /// * e.g. after a node fails then recovers, to move some work back to it
5502 : /// - Cutting over to a secondary if it improves the spread of shard attachments within a tenant
5503 : /// * e.g. after a shard split, the initial attached locations will all be on the node where
5504 : /// we did the split, but are probably better placed elsewhere.
5505 : /// - Creating new secondary locations if it improves the spreading of a sharded tenant
5506 : /// * e.g. after a shard split, some locations will be on the same node (where the split
5507 : /// happened), and will probably be better placed elsewhere.
5508 : ///
5509 : /// To put it more briefly: whereas the scheduler respects soft constraints in a ScheduleContext at
5510 : /// the time of scheduling, this function looks for cases where a better-scoring location is available
5511 : /// according to those same soft constraints.
5512 0 : async fn optimize_all(&self) -> usize {
5513 0 : // Limit on how many shards' optmizations each call to this function will execute. Combined
5514 0 : // with the frequency of background calls, this acts as an implicit rate limit that runs a small
5515 0 : // trickle of optimizations in the background, rather than executing a large number in parallel
5516 0 : // when a change occurs.
5517 0 : const MAX_OPTIMIZATIONS_EXEC_PER_PASS: usize = 2;
5518 0 :
5519 0 : // Synchronous prepare: scan shards for possible scheduling optimizations
5520 0 : let candidate_work = self.optimize_all_plan();
5521 0 : let candidate_work_len = candidate_work.len();
5522 :
5523 : // Asynchronous validate: I/O to pageservers to make sure shards are in a good state to apply validation
5524 0 : let validated_work = self.optimize_all_validate(candidate_work).await;
5525 :
5526 0 : let was_work_filtered = validated_work.len() != candidate_work_len;
5527 0 :
5528 0 : // Synchronous apply: update the shards' intent states according to validated optimisations
5529 0 : let mut reconciles_spawned = 0;
5530 0 : let mut optimizations_applied = 0;
5531 0 : let mut locked = self.inner.write().unwrap();
5532 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
5533 0 : for (tenant_shard_id, optimization) in validated_work {
5534 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
5535 : // Shard was dropped between planning and execution;
5536 0 : continue;
5537 : };
5538 0 : if shard.apply_optimization(scheduler, optimization) {
5539 0 : optimizations_applied += 1;
5540 0 : if self.maybe_reconcile_shard(shard, nodes).is_some() {
5541 0 : reconciles_spawned += 1;
5542 0 : }
5543 0 : }
5544 :
5545 0 : if optimizations_applied >= MAX_OPTIMIZATIONS_EXEC_PER_PASS {
5546 0 : break;
5547 0 : }
5548 : }
5549 :
5550 0 : if was_work_filtered {
5551 0 : // If we filtered any work out during validation, ensure we return a nonzero value to indicate
5552 0 : // to callers that the system is not in a truly quiet state, it's going to do some work as soon
5553 0 : // as these validations start passing.
5554 0 : reconciles_spawned = std::cmp::max(reconciles_spawned, 1);
5555 0 : }
5556 :
5557 0 : reconciles_spawned
5558 0 : }
5559 :
5560 0 : fn optimize_all_plan(&self) -> Vec<(TenantShardId, ScheduleOptimization)> {
5561 0 : let mut schedule_context = ScheduleContext::default();
5562 0 :
5563 0 : let mut tenant_shards: Vec<&TenantShard> = Vec::new();
5564 0 :
5565 0 : // How many candidate optimizations we will generate, before evaluating them for readniess: setting
5566 0 : // this higher than the execution limit gives us a chance to execute some work even if the first
5567 0 : // few optimizations we find are not ready.
5568 0 : const MAX_OPTIMIZATIONS_PLAN_PER_PASS: usize = 8;
5569 0 :
5570 0 : let mut work = Vec::new();
5571 0 :
5572 0 : let mut locked = self.inner.write().unwrap();
5573 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
5574 0 : for (tenant_shard_id, shard) in tenants.iter() {
5575 0 : if tenant_shard_id.is_shard_zero() {
5576 0 : // Reset accumulators on the first shard in a tenant
5577 0 : schedule_context = ScheduleContext::default();
5578 0 : schedule_context.mode = ScheduleMode::Speculative;
5579 0 : tenant_shards.clear();
5580 0 : }
5581 :
5582 0 : if work.len() >= MAX_OPTIMIZATIONS_PLAN_PER_PASS {
5583 0 : break;
5584 0 : }
5585 0 :
5586 0 : match shard.get_scheduling_policy() {
5587 0 : ShardSchedulingPolicy::Active => {
5588 0 : // Ok to do optimization
5589 0 : }
5590 : ShardSchedulingPolicy::Essential
5591 : | ShardSchedulingPolicy::Pause
5592 : | ShardSchedulingPolicy::Stop => {
5593 : // Policy prevents optimizing this shard.
5594 0 : continue;
5595 : }
5596 : }
5597 :
5598 : // Accumulate the schedule context for all the shards in a tenant: we must have
5599 : // the total view of all shards before we can try to optimize any of them.
5600 0 : schedule_context.avoid(&shard.intent.all_pageservers());
5601 0 : if let Some(attached) = shard.intent.get_attached() {
5602 0 : schedule_context.push_attached(*attached);
5603 0 : }
5604 0 : tenant_shards.push(shard);
5605 0 :
5606 0 : // Once we have seen the last shard in the tenant, proceed to search across all shards
5607 0 : // in the tenant for optimizations
5608 0 : if shard.shard.number.0 == shard.shard.count.count() - 1 {
5609 0 : if tenant_shards.iter().any(|s| s.reconciler.is_some()) {
5610 : // Do not start any optimizations while another change to the tenant is ongoing: this
5611 : // is not necessary for correctness, but simplifies operations and implicitly throttles
5612 : // optimization changes to happen in a "trickle" over time.
5613 0 : continue;
5614 0 : }
5615 0 :
5616 0 : if tenant_shards.iter().any(|s| {
5617 0 : !matches!(s.splitting, SplitState::Idle)
5618 0 : || matches!(s.policy, PlacementPolicy::Detached)
5619 0 : }) {
5620 : // Never attempt to optimize a tenant that is currently being split, or
5621 : // a tenant that is meant to be detached
5622 0 : continue;
5623 0 : }
5624 :
5625 : // TODO: optimization calculations are relatively expensive: create some fast-path for
5626 : // the common idle case (avoiding the search on tenants that we have recently checked)
5627 :
5628 0 : for shard in &tenant_shards {
5629 0 : if let Some(optimization) =
5630 : // If idle, maybe ptimize attachments: if a shard has a secondary location that is preferable to
5631 : // its primary location based on soft constraints, cut it over.
5632 0 : shard.optimize_attachment(nodes, &schedule_context)
5633 : {
5634 0 : work.push((shard.tenant_shard_id, optimization));
5635 0 : break;
5636 0 : } else if let Some(optimization) =
5637 : // If idle, maybe optimize secondary locations: if a shard has a secondary location that would be
5638 : // better placed on another node, based on ScheduleContext, then adjust it. This
5639 : // covers cases like after a shard split, where we might have too many shards
5640 : // in the same tenant with secondary locations on the node where they originally split.
5641 0 : shard.optimize_secondary(scheduler, &schedule_context)
5642 : {
5643 0 : work.push((shard.tenant_shard_id, optimization));
5644 0 : break;
5645 0 : }
5646 :
5647 : // TODO: extend this mechanism to prefer attaching on nodes with fewer attached
5648 : // tenants (i.e. extend schedule state to distinguish attached from secondary counts),
5649 : // for the total number of attachments on a node (not just within a tenant.)
5650 : }
5651 0 : }
5652 : }
5653 :
5654 0 : work
5655 0 : }
5656 :
5657 0 : async fn optimize_all_validate(
5658 0 : &self,
5659 0 : candidate_work: Vec<(TenantShardId, ScheduleOptimization)>,
5660 0 : ) -> Vec<(TenantShardId, ScheduleOptimization)> {
5661 0 : // Take a clone of the node map to use outside the lock in async validation phase
5662 0 : let validation_nodes = { self.inner.read().unwrap().nodes.clone() };
5663 0 :
5664 0 : let mut want_secondary_status = Vec::new();
5665 0 :
5666 0 : // Validate our plans: this is an async phase where we may do I/O to pageservers to
5667 0 : // check that the state of locations is acceptable to run the optimization, such as
5668 0 : // checking that a secondary location is sufficiently warmed-up to cleanly cut over
5669 0 : // in a live migration.
5670 0 : let mut validated_work = Vec::new();
5671 0 : for (tenant_shard_id, optimization) in candidate_work {
5672 0 : match optimization.action {
5673 : ScheduleOptimizationAction::MigrateAttachment(MigrateAttachment {
5674 : old_attached_node_id: _,
5675 0 : new_attached_node_id,
5676 0 : }) => {
5677 0 : match validation_nodes.get(&new_attached_node_id) {
5678 0 : None => {
5679 0 : // Node was dropped between planning and validation
5680 0 : }
5681 0 : Some(node) => {
5682 0 : if !node.is_available() {
5683 0 : tracing::info!("Skipping optimization migration of {tenant_shard_id} to {new_attached_node_id} because node unavailable");
5684 0 : } else {
5685 0 : // Accumulate optimizations that require fetching secondary status, so that we can execute these
5686 0 : // remote API requests concurrently.
5687 0 : want_secondary_status.push((
5688 0 : tenant_shard_id,
5689 0 : node.clone(),
5690 0 : optimization,
5691 0 : ));
5692 0 : }
5693 : }
5694 : }
5695 : }
5696 : ScheduleOptimizationAction::ReplaceSecondary(_) => {
5697 : // No extra checks needed to replace a secondary: this does not interrupt client access
5698 0 : validated_work.push((tenant_shard_id, optimization))
5699 : }
5700 : };
5701 : }
5702 :
5703 : // Call into pageserver API to find out if the destination secondary location is warm enough for a reasonably smooth migration: we
5704 : // do this so that we avoid spawning a Reconciler that would have to wait minutes/hours for a destination to warm up: that reconciler
5705 : // would hold a precious reconcile semaphore unit the whole time it was waiting for the destination to warm up.
5706 0 : let results = self
5707 0 : .tenant_for_shards_api(
5708 0 : want_secondary_status
5709 0 : .iter()
5710 0 : .map(|i| (i.0, i.1.clone()))
5711 0 : .collect(),
5712 0 : |tenant_shard_id, client| async move {
5713 0 : client.tenant_secondary_status(tenant_shard_id).await
5714 0 : },
5715 0 : 1,
5716 0 : 1,
5717 0 : SHORT_RECONCILE_TIMEOUT,
5718 0 : &self.cancel,
5719 0 : )
5720 0 : .await;
5721 :
5722 0 : for ((tenant_shard_id, node, optimization), secondary_status) in
5723 0 : want_secondary_status.into_iter().zip(results.into_iter())
5724 : {
5725 0 : match secondary_status {
5726 0 : Err(e) => {
5727 0 : tracing::info!("Skipping migration of {tenant_shard_id} to {node}, error querying secondary: {e}");
5728 : }
5729 0 : Ok(progress) => {
5730 0 : // We require secondary locations to have less than 10GiB of downloads pending before we will use
5731 0 : // them in an optimization
5732 0 : const DOWNLOAD_FRESHNESS_THRESHOLD: u64 = 10 * 1024 * 1024 * 1024;
5733 0 :
5734 0 : if progress.heatmap_mtime.is_none()
5735 0 : || progress.bytes_total < DOWNLOAD_FRESHNESS_THRESHOLD
5736 0 : && progress.bytes_downloaded != progress.bytes_total
5737 0 : || progress.bytes_total - progress.bytes_downloaded
5738 0 : > DOWNLOAD_FRESHNESS_THRESHOLD
5739 : {
5740 0 : tracing::info!("Skipping migration of {tenant_shard_id} to {node} because secondary isn't ready: {progress:?}");
5741 : } else {
5742 : // Location looks ready: proceed
5743 0 : tracing::info!(
5744 0 : "{tenant_shard_id} secondary on {node} is warm enough for migration: {progress:?}"
5745 : );
5746 0 : validated_work.push((tenant_shard_id, optimization))
5747 : }
5748 : }
5749 : }
5750 : }
5751 :
5752 0 : validated_work
5753 0 : }
5754 :
5755 : /// Look for shards which are oversized and in need of splitting
5756 0 : async fn autosplit_tenants(self: &Arc<Self>) {
5757 0 : let Some(split_threshold) = self.config.split_threshold else {
5758 : // Auto-splitting is disabled
5759 0 : return;
5760 : };
5761 :
5762 0 : let nodes = self.inner.read().unwrap().nodes.clone();
5763 0 :
5764 0 : const SPLIT_TO_MAX: ShardCount = ShardCount::new(8);
5765 0 :
5766 0 : let mut top_n = Vec::new();
5767 0 :
5768 0 : // Call into each node to look for big tenants
5769 0 : let top_n_request = TopTenantShardsRequest {
5770 0 : // We currently split based on logical size, for simplicity: logical size is a signal of
5771 0 : // the user's intent to run a large database, whereas physical/resident size can be symptoms
5772 0 : // of compaction issues. Eventually we should switch to using resident size to bound the
5773 0 : // disk space impact of one shard.
5774 0 : order_by: models::TenantSorting::MaxLogicalSize,
5775 0 : limit: 10,
5776 0 : where_shards_lt: Some(SPLIT_TO_MAX),
5777 0 : where_gt: Some(split_threshold),
5778 0 : };
5779 0 : for node in nodes.values() {
5780 0 : let request_ref = &top_n_request;
5781 0 : match node
5782 0 : .with_client_retries(
5783 0 : |client| async move {
5784 0 : let request = request_ref.clone();
5785 0 : client.top_tenant_shards(request.clone()).await
5786 0 : },
5787 0 : &self.config.jwt_token,
5788 0 : 3,
5789 0 : 3,
5790 0 : Duration::from_secs(5),
5791 0 : &self.cancel,
5792 0 : )
5793 0 : .await
5794 : {
5795 0 : Some(Ok(node_top_n)) => {
5796 0 : top_n.extend(node_top_n.shards.into_iter());
5797 0 : }
5798 : Some(Err(mgmt_api::Error::Cancelled)) => {
5799 0 : continue;
5800 : }
5801 0 : Some(Err(e)) => {
5802 0 : tracing::warn!("Failed to fetch top N tenants from {node}: {e}");
5803 0 : continue;
5804 : }
5805 : None => {
5806 : // Node is shutting down
5807 0 : continue;
5808 : }
5809 : };
5810 : }
5811 :
5812 : // Pick the biggest tenant to split first
5813 0 : top_n.sort_by_key(|i| i.resident_size);
5814 0 : let Some(split_candidate) = top_n.into_iter().next() else {
5815 0 : tracing::debug!("No split-elegible shards found");
5816 0 : return;
5817 : };
5818 :
5819 : // We spawn a task to run this, so it's exactly like some external API client requesting it. We don't
5820 : // want to block the background reconcile loop on this.
5821 0 : tracing::info!("Auto-splitting tenant for size threshold {split_threshold}: current size {split_candidate:?}");
5822 :
5823 0 : let this = self.clone();
5824 0 : tokio::spawn(
5825 0 : async move {
5826 0 : match this
5827 0 : .tenant_shard_split(
5828 0 : split_candidate.id.tenant_id,
5829 0 : TenantShardSplitRequest {
5830 0 : // Always split to the max number of shards: this avoids stepping through
5831 0 : // intervening shard counts and encountering the overrhead of a split+cleanup
5832 0 : // each time as a tenant grows, and is not too expensive because our max shard
5833 0 : // count is relatively low anyway.
5834 0 : // This policy will be adjusted in future once we support higher shard count.
5835 0 : new_shard_count: SPLIT_TO_MAX.literal(),
5836 0 : new_stripe_size: Some(ShardParameters::DEFAULT_STRIPE_SIZE),
5837 0 : },
5838 0 : )
5839 0 : .await
5840 : {
5841 : Ok(_) => {
5842 0 : tracing::info!("Successful auto-split");
5843 : }
5844 0 : Err(e) => {
5845 0 : tracing::error!("Auto-split failed: {e}");
5846 : }
5847 : }
5848 0 : }
5849 0 : .instrument(tracing::info_span!("auto_split", tenant_id=%split_candidate.id.tenant_id)),
5850 : );
5851 0 : }
5852 :
5853 : /// Useful for tests: run whatever work a background [`Self::reconcile_all`] would have done, but
5854 : /// also wait for any generated Reconcilers to complete. Calling this until it returns zero should
5855 : /// put the system into a quiescent state where future background reconciliations won't do anything.
5856 0 : pub(crate) async fn reconcile_all_now(&self) -> Result<usize, ReconcileWaitError> {
5857 0 : let reconciles_spawned = self.reconcile_all();
5858 0 : let reconciles_spawned = if reconciles_spawned == 0 {
5859 : // Only optimize when we are otherwise idle
5860 0 : self.optimize_all().await
5861 : } else {
5862 0 : reconciles_spawned
5863 : };
5864 :
5865 0 : let waiters = {
5866 0 : let mut waiters = Vec::new();
5867 0 : let locked = self.inner.read().unwrap();
5868 0 : for (_tenant_shard_id, shard) in locked.tenants.iter() {
5869 0 : if let Some(waiter) = shard.get_waiter() {
5870 0 : waiters.push(waiter);
5871 0 : }
5872 : }
5873 0 : waiters
5874 0 : };
5875 0 :
5876 0 : let waiter_count = waiters.len();
5877 0 : match self.await_waiters(waiters, RECONCILE_TIMEOUT).await {
5878 0 : Ok(()) => {}
5879 0 : Err(ReconcileWaitError::Failed(_, reconcile_error))
5880 0 : if matches!(*reconcile_error, ReconcileError::Cancel) =>
5881 0 : {
5882 0 : // Ignore reconciler cancel errors: this reconciler might have shut down
5883 0 : // because some other change superceded it. We will return a nonzero number,
5884 0 : // so the caller knows they might have to call again to quiesce the system.
5885 0 : }
5886 0 : Err(e) => {
5887 0 : return Err(e);
5888 : }
5889 : };
5890 :
5891 0 : tracing::info!(
5892 0 : "{} reconciles in reconcile_all, {} waiters",
5893 : reconciles_spawned,
5894 : waiter_count
5895 : );
5896 :
5897 0 : Ok(std::cmp::max(waiter_count, reconciles_spawned))
5898 0 : }
5899 :
5900 0 : async fn stop_reconciliations(&self, reason: StopReconciliationsReason) {
5901 0 : // Cancel all on-going reconciles and wait for them to exit the gate.
5902 0 : tracing::info!("{reason}: cancelling and waiting for in-flight reconciles");
5903 0 : self.reconcilers_cancel.cancel();
5904 0 : self.reconcilers_gate.close().await;
5905 :
5906 : // Signal the background loop in [`Service::process_results`] to exit once
5907 : // it has proccessed the results from all the reconciles we cancelled earlier.
5908 0 : tracing::info!("{reason}: processing results from previously in-flight reconciles");
5909 0 : self.result_tx.send(ReconcileResultRequest::Stop).ok();
5910 0 : self.result_tx.closed().await;
5911 0 : }
5912 :
5913 0 : pub async fn shutdown(&self) {
5914 0 : self.stop_reconciliations(StopReconciliationsReason::ShuttingDown)
5915 0 : .await;
5916 :
5917 : // Background tasks hold gate guards: this notifies them of the cancellation and
5918 : // waits for them all to complete.
5919 0 : tracing::info!("Shutting down: cancelling and waiting for background tasks to exit");
5920 0 : self.cancel.cancel();
5921 0 : self.gate.close().await;
5922 0 : }
5923 :
5924 : /// Spot check the download lag for a secondary location of a shard.
5925 : /// Should be used as a heuristic, since it's not always precise: the
5926 : /// secondary might have not downloaded the new heat map yet and, hence,
5927 : /// is not aware of the lag.
5928 : ///
5929 : /// Returns:
5930 : /// * Ok(None) if the lag could not be determined from the status,
5931 : /// * Ok(Some(_)) if the lag could be determind
5932 : /// * Err on failures to query the pageserver.
5933 0 : async fn secondary_lag(
5934 0 : &self,
5935 0 : secondary: &NodeId,
5936 0 : tenant_shard_id: TenantShardId,
5937 0 : ) -> Result<Option<u64>, mgmt_api::Error> {
5938 0 : let nodes = self.inner.read().unwrap().nodes.clone();
5939 0 : let node = nodes.get(secondary).ok_or(mgmt_api::Error::ApiError(
5940 0 : StatusCode::NOT_FOUND,
5941 0 : format!("Node with id {} not found", secondary),
5942 0 : ))?;
5943 :
5944 0 : match node
5945 0 : .with_client_retries(
5946 0 : |client| async move { client.tenant_secondary_status(tenant_shard_id).await },
5947 0 : &self.config.jwt_token,
5948 0 : 1,
5949 0 : 3,
5950 0 : Duration::from_millis(250),
5951 0 : &self.cancel,
5952 0 : )
5953 0 : .await
5954 : {
5955 0 : Some(Ok(status)) => match status.heatmap_mtime {
5956 0 : Some(_) => Ok(Some(status.bytes_total - status.bytes_downloaded)),
5957 0 : None => Ok(None),
5958 : },
5959 0 : Some(Err(e)) => Err(e),
5960 0 : None => Err(mgmt_api::Error::Cancelled),
5961 : }
5962 0 : }
5963 :
5964 : /// Drain a node by moving the shards attached to it as primaries.
5965 : /// This is a long running operation and it should run as a separate Tokio task.
5966 0 : pub(crate) async fn drain_node(
5967 0 : self: &Arc<Self>,
5968 0 : node_id: NodeId,
5969 0 : cancel: CancellationToken,
5970 0 : ) -> Result<(), OperationError> {
5971 0 : const MAX_SECONDARY_LAG_BYTES_DEFAULT: u64 = 256 * 1024 * 1024;
5972 0 : let max_secondary_lag_bytes = self
5973 0 : .config
5974 0 : .max_secondary_lag_bytes
5975 0 : .unwrap_or(MAX_SECONDARY_LAG_BYTES_DEFAULT);
5976 0 :
5977 0 : // By default, live migrations are generous about the wait time for getting
5978 0 : // the secondary location up to speed. When draining, give up earlier in order
5979 0 : // to not stall the operation when a cold secondary is encountered.
5980 0 : const SECONDARY_WARMUP_TIMEOUT: Duration = Duration::from_secs(20);
5981 0 : const SECONDARY_DOWNLOAD_REQUEST_TIMEOUT: Duration = Duration::from_secs(5);
5982 0 : let reconciler_config = ReconcilerConfigBuilder::new()
5983 0 : .secondary_warmup_timeout(SECONDARY_WARMUP_TIMEOUT)
5984 0 : .secondary_download_request_timeout(SECONDARY_DOWNLOAD_REQUEST_TIMEOUT)
5985 0 : .build();
5986 0 :
5987 0 : let mut waiters = Vec::new();
5988 0 :
5989 0 : let mut tid_iter = TenantShardIterator::new({
5990 0 : let service = self.clone();
5991 0 : move |last_inspected_shard: Option<TenantShardId>| {
5992 0 : let locked = &service.inner.read().unwrap();
5993 0 : let tenants = &locked.tenants;
5994 0 : let entry = match last_inspected_shard {
5995 0 : Some(skip_past) => {
5996 0 : // Skip to the last seen tenant shard id
5997 0 : let mut cursor = tenants.iter().skip_while(|(tid, _)| **tid != skip_past);
5998 0 :
5999 0 : // Skip past the last seen
6000 0 : cursor.nth(1)
6001 : }
6002 0 : None => tenants.first_key_value(),
6003 : };
6004 :
6005 0 : entry.map(|(tid, _)| tid).copied()
6006 0 : }
6007 0 : });
6008 :
6009 0 : while !tid_iter.finished() {
6010 0 : if cancel.is_cancelled() {
6011 0 : match self
6012 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
6013 0 : .await
6014 : {
6015 0 : Ok(()) => return Err(OperationError::Cancelled),
6016 0 : Err(err) => {
6017 0 : return Err(OperationError::FinalizeError(
6018 0 : format!(
6019 0 : "Failed to finalise drain cancel of {} by setting scheduling policy to Active: {}",
6020 0 : node_id, err
6021 0 : )
6022 0 : .into(),
6023 0 : ));
6024 : }
6025 : }
6026 0 : }
6027 0 :
6028 0 : drain_utils::validate_node_state(&node_id, self.inner.read().unwrap().nodes.clone())?;
6029 :
6030 0 : while waiters.len() < MAX_RECONCILES_PER_OPERATION {
6031 0 : let tid = match tid_iter.next() {
6032 0 : Some(tid) => tid,
6033 : None => {
6034 0 : break;
6035 : }
6036 : };
6037 :
6038 0 : let tid_drain = TenantShardDrain {
6039 0 : drained_node: node_id,
6040 0 : tenant_shard_id: tid,
6041 0 : };
6042 :
6043 0 : let dest_node_id = {
6044 0 : let locked = self.inner.read().unwrap();
6045 0 :
6046 0 : match tid_drain
6047 0 : .tenant_shard_eligible_for_drain(&locked.tenants, &locked.scheduler)
6048 : {
6049 0 : Some(node_id) => node_id,
6050 : None => {
6051 0 : continue;
6052 : }
6053 : }
6054 : };
6055 :
6056 0 : match self.secondary_lag(&dest_node_id, tid).await {
6057 0 : Ok(Some(lag)) if lag <= max_secondary_lag_bytes => {
6058 0 : // The secondary is reasonably up to date.
6059 0 : // Migrate to it
6060 0 : }
6061 0 : Ok(Some(lag)) => {
6062 0 : tracing::info!(
6063 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
6064 0 : "Secondary on node {dest_node_id} is lagging by {lag}. Skipping reconcile."
6065 : );
6066 0 : continue;
6067 : }
6068 : Ok(None) => {
6069 0 : tracing::info!(
6070 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
6071 0 : "Could not determine lag for secondary on node {dest_node_id}. Skipping reconcile."
6072 : );
6073 0 : continue;
6074 : }
6075 0 : Err(err) => {
6076 0 : tracing::warn!(
6077 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
6078 0 : "Failed to get secondary lag from node {dest_node_id}. Skipping reconcile: {err}"
6079 : );
6080 0 : continue;
6081 : }
6082 : }
6083 :
6084 : {
6085 0 : let mut locked = self.inner.write().unwrap();
6086 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
6087 0 : let rescheduled = tid_drain.reschedule_to_secondary(
6088 0 : dest_node_id,
6089 0 : tenants,
6090 0 : scheduler,
6091 0 : nodes,
6092 0 : )?;
6093 :
6094 0 : if let Some(tenant_shard) = rescheduled {
6095 0 : let waiter = self.maybe_configured_reconcile_shard(
6096 0 : tenant_shard,
6097 0 : nodes,
6098 0 : reconciler_config,
6099 0 : );
6100 0 : if let Some(some) = waiter {
6101 0 : waiters.push(some);
6102 0 : }
6103 0 : }
6104 : }
6105 : }
6106 :
6107 0 : waiters = self
6108 0 : .await_waiters_remainder(waiters, SHORT_RECONCILE_TIMEOUT)
6109 0 : .await;
6110 :
6111 0 : failpoint_support::sleep_millis_async!("sleepy-drain-loop", &cancel);
6112 : }
6113 :
6114 0 : while !waiters.is_empty() {
6115 0 : if cancel.is_cancelled() {
6116 0 : match self
6117 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
6118 0 : .await
6119 : {
6120 0 : Ok(()) => return Err(OperationError::Cancelled),
6121 0 : Err(err) => {
6122 0 : return Err(OperationError::FinalizeError(
6123 0 : format!(
6124 0 : "Failed to finalise drain cancel of {} by setting scheduling policy to Active: {}",
6125 0 : node_id, err
6126 0 : )
6127 0 : .into(),
6128 0 : ));
6129 : }
6130 : }
6131 0 : }
6132 0 :
6133 0 : tracing::info!("Awaiting {} pending drain reconciliations", waiters.len());
6134 :
6135 0 : waiters = self
6136 0 : .await_waiters_remainder(waiters, SHORT_RECONCILE_TIMEOUT)
6137 0 : .await;
6138 : }
6139 :
6140 : // At this point we have done the best we could to drain shards from this node.
6141 : // Set the node scheduling policy to `[NodeSchedulingPolicy::PauseForRestart]`
6142 : // to complete the drain.
6143 0 : if let Err(err) = self
6144 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::PauseForRestart))
6145 0 : .await
6146 : {
6147 : // This is not fatal. Anything that is polling the node scheduling policy to detect
6148 : // the end of the drain operations will hang, but all such places should enforce an
6149 : // overall timeout. The scheduling policy will be updated upon node re-attach and/or
6150 : // by the counterpart fill operation.
6151 0 : return Err(OperationError::FinalizeError(
6152 0 : format!(
6153 0 : "Failed to finalise drain of {node_id} by setting scheduling policy to PauseForRestart: {err}"
6154 0 : )
6155 0 : .into(),
6156 0 : ));
6157 0 : }
6158 0 :
6159 0 : Ok(())
6160 0 : }
6161 :
6162 : /// Create a node fill plan (pick secondaries to promote) that meets the following requirements:
6163 : /// 1. The node should be filled until it reaches the expected cluster average of
6164 : /// attached shards. If there are not enough secondaries on the node, the plan stops early.
6165 : /// 2. Select tenant shards to promote such that the number of attached shards is balanced
6166 : /// throughout the cluster. We achieve this by picking tenant shards from each node,
6167 : /// starting from the ones with the largest number of attached shards, until the node
6168 : /// reaches the expected cluster average.
6169 : /// 3. Avoid promoting more shards of the same tenant than required. The upper bound
6170 : /// for the number of tenants from the same shard promoted to the node being filled is:
6171 : /// shard count for the tenant divided by the number of nodes in the cluster.
6172 0 : fn fill_node_plan(&self, node_id: NodeId) -> Vec<TenantShardId> {
6173 0 : let mut locked = self.inner.write().unwrap();
6174 0 : let fill_requirement = locked.scheduler.compute_fill_requirement(node_id);
6175 0 :
6176 0 : let mut tids_by_node = locked
6177 0 : .tenants
6178 0 : .iter_mut()
6179 0 : .filter_map(|(tid, tenant_shard)| {
6180 0 : if tenant_shard.intent.get_secondary().contains(&node_id) {
6181 0 : if let Some(primary) = tenant_shard.intent.get_attached() {
6182 0 : return Some((*primary, *tid));
6183 0 : }
6184 0 : }
6185 :
6186 0 : None
6187 0 : })
6188 0 : .into_group_map();
6189 0 :
6190 0 : let expected_attached = locked.scheduler.expected_attached_shard_count();
6191 0 : let nodes_by_load = locked.scheduler.nodes_by_attached_shard_count();
6192 0 :
6193 0 : let mut promoted_per_tenant: HashMap<TenantId, usize> = HashMap::new();
6194 0 : let mut plan = Vec::new();
6195 :
6196 0 : for (node_id, attached) in nodes_by_load {
6197 0 : let available = locked
6198 0 : .nodes
6199 0 : .get(&node_id)
6200 0 : .map_or(false, |n| n.is_available());
6201 0 : if !available {
6202 0 : continue;
6203 0 : }
6204 0 :
6205 0 : if plan.len() >= fill_requirement
6206 0 : || tids_by_node.is_empty()
6207 0 : || attached <= expected_attached
6208 : {
6209 0 : break;
6210 0 : }
6211 0 :
6212 0 : let can_take = attached - expected_attached;
6213 0 : let needed = fill_requirement - plan.len();
6214 0 : let mut take = std::cmp::min(can_take, needed);
6215 0 :
6216 0 : let mut remove_node = false;
6217 0 : while take > 0 {
6218 0 : match tids_by_node.get_mut(&node_id) {
6219 0 : Some(tids) => match tids.pop() {
6220 0 : Some(tid) => {
6221 0 : let max_promote_for_tenant = std::cmp::max(
6222 0 : tid.shard_count.count() as usize / locked.nodes.len(),
6223 0 : 1,
6224 0 : );
6225 0 : let promoted = promoted_per_tenant.entry(tid.tenant_id).or_default();
6226 0 : if *promoted < max_promote_for_tenant {
6227 0 : plan.push(tid);
6228 0 : *promoted += 1;
6229 0 : take -= 1;
6230 0 : }
6231 : }
6232 : None => {
6233 0 : remove_node = true;
6234 0 : break;
6235 : }
6236 : },
6237 : None => {
6238 0 : break;
6239 : }
6240 : }
6241 : }
6242 :
6243 0 : if remove_node {
6244 0 : tids_by_node.remove(&node_id);
6245 0 : }
6246 : }
6247 :
6248 0 : plan
6249 0 : }
6250 :
6251 : /// Fill a node by promoting its secondaries until the cluster is balanced
6252 : /// with regards to attached shard counts. Note that this operation only
6253 : /// makes sense as a counterpart to the drain implemented in [`Service::drain_node`].
6254 : /// This is a long running operation and it should run as a separate Tokio task.
6255 0 : pub(crate) async fn fill_node(
6256 0 : &self,
6257 0 : node_id: NodeId,
6258 0 : cancel: CancellationToken,
6259 0 : ) -> Result<(), OperationError> {
6260 0 : // TODO(vlad): Currently this operates on the assumption that all
6261 0 : // secondaries are warm. This is not always true (e.g. we just migrated the
6262 0 : // tenant). Take that into consideration by checking the secondary status.
6263 0 : let mut tids_to_promote = self.fill_node_plan(node_id);
6264 0 : let mut waiters = Vec::new();
6265 :
6266 : // Execute the plan we've composed above. Before aplying each move from the plan,
6267 : // we validate to ensure that it has not gone stale in the meantime.
6268 0 : while !tids_to_promote.is_empty() {
6269 0 : if cancel.is_cancelled() {
6270 0 : match self
6271 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
6272 0 : .await
6273 : {
6274 0 : Ok(()) => return Err(OperationError::Cancelled),
6275 0 : Err(err) => {
6276 0 : return Err(OperationError::FinalizeError(
6277 0 : format!(
6278 0 : "Failed to finalise drain cancel of {} by setting scheduling policy to Active: {}",
6279 0 : node_id, err
6280 0 : )
6281 0 : .into(),
6282 0 : ));
6283 : }
6284 : }
6285 0 : }
6286 0 :
6287 0 : {
6288 0 : let mut locked = self.inner.write().unwrap();
6289 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
6290 :
6291 0 : let node = nodes.get(&node_id).ok_or(OperationError::NodeStateChanged(
6292 0 : format!("node {node_id} was removed").into(),
6293 0 : ))?;
6294 :
6295 0 : let current_policy = node.get_scheduling();
6296 0 : if !matches!(current_policy, NodeSchedulingPolicy::Filling) {
6297 : // TODO(vlad): maybe cancel pending reconciles before erroring out. need to think
6298 : // about it
6299 0 : return Err(OperationError::NodeStateChanged(
6300 0 : format!("node {node_id} changed state to {current_policy:?}").into(),
6301 0 : ));
6302 0 : }
6303 :
6304 0 : while waiters.len() < MAX_RECONCILES_PER_OPERATION {
6305 0 : if let Some(tid) = tids_to_promote.pop() {
6306 0 : if let Some(tenant_shard) = tenants.get_mut(&tid) {
6307 : // If the node being filled is not a secondary anymore,
6308 : // skip the promotion.
6309 0 : if !tenant_shard.intent.get_secondary().contains(&node_id) {
6310 0 : continue;
6311 0 : }
6312 0 :
6313 0 : let previously_attached_to = *tenant_shard.intent.get_attached();
6314 0 : match tenant_shard.reschedule_to_secondary(Some(node_id), scheduler) {
6315 0 : Err(e) => {
6316 0 : tracing::warn!(
6317 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
6318 0 : "Scheduling error when filling pageserver {} : {e}", node_id
6319 : );
6320 : }
6321 : Ok(()) => {
6322 0 : tracing::info!(
6323 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
6324 0 : "Rescheduled shard while filling node {}: {:?} -> {}",
6325 : node_id,
6326 : previously_attached_to,
6327 : node_id
6328 : );
6329 :
6330 0 : if let Some(waiter) =
6331 0 : self.maybe_reconcile_shard(tenant_shard, nodes)
6332 0 : {
6333 0 : waiters.push(waiter);
6334 0 : }
6335 : }
6336 : }
6337 0 : }
6338 : } else {
6339 0 : break;
6340 : }
6341 : }
6342 : }
6343 :
6344 0 : waiters = self
6345 0 : .await_waiters_remainder(waiters, SHORT_RECONCILE_TIMEOUT)
6346 0 : .await;
6347 : }
6348 :
6349 0 : while !waiters.is_empty() {
6350 0 : if cancel.is_cancelled() {
6351 0 : match self
6352 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
6353 0 : .await
6354 : {
6355 0 : Ok(()) => return Err(OperationError::Cancelled),
6356 0 : Err(err) => {
6357 0 : return Err(OperationError::FinalizeError(
6358 0 : format!(
6359 0 : "Failed to finalise drain cancel of {} by setting scheduling policy to Active: {}",
6360 0 : node_id, err
6361 0 : )
6362 0 : .into(),
6363 0 : ));
6364 : }
6365 : }
6366 0 : }
6367 0 :
6368 0 : tracing::info!("Awaiting {} pending fill reconciliations", waiters.len());
6369 :
6370 0 : waiters = self
6371 0 : .await_waiters_remainder(waiters, SHORT_RECONCILE_TIMEOUT)
6372 0 : .await;
6373 : }
6374 :
6375 0 : if let Err(err) = self
6376 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
6377 0 : .await
6378 : {
6379 : // This isn't a huge issue since the filling process starts upon request. However, it
6380 : // will prevent the next drain from starting. The only case in which this can fail
6381 : // is database unavailability. Such a case will require manual intervention.
6382 0 : return Err(OperationError::FinalizeError(
6383 0 : format!("Failed to finalise fill of {node_id} by setting scheduling policy to Active: {err}")
6384 0 : .into(),
6385 0 : ));
6386 0 : }
6387 0 :
6388 0 : Ok(())
6389 0 : }
6390 :
6391 : /// Updates scrubber metadata health check results.
6392 0 : pub(crate) async fn metadata_health_update(
6393 0 : &self,
6394 0 : update_req: MetadataHealthUpdateRequest,
6395 0 : ) -> Result<(), ApiError> {
6396 0 : let now = chrono::offset::Utc::now();
6397 0 : let (healthy_records, unhealthy_records) = {
6398 0 : let locked = self.inner.read().unwrap();
6399 0 : let healthy_records = update_req
6400 0 : .healthy_tenant_shards
6401 0 : .into_iter()
6402 0 : // Retain only health records associated with tenant shards managed by storage controller.
6403 0 : .filter(|tenant_shard_id| locked.tenants.contains_key(tenant_shard_id))
6404 0 : .map(|tenant_shard_id| MetadataHealthPersistence::new(tenant_shard_id, true, now))
6405 0 : .collect();
6406 0 : let unhealthy_records = update_req
6407 0 : .unhealthy_tenant_shards
6408 0 : .into_iter()
6409 0 : .filter(|tenant_shard_id| locked.tenants.contains_key(tenant_shard_id))
6410 0 : .map(|tenant_shard_id| MetadataHealthPersistence::new(tenant_shard_id, false, now))
6411 0 : .collect();
6412 0 :
6413 0 : (healthy_records, unhealthy_records)
6414 0 : };
6415 0 :
6416 0 : self.persistence
6417 0 : .update_metadata_health_records(healthy_records, unhealthy_records, now)
6418 0 : .await?;
6419 0 : Ok(())
6420 0 : }
6421 :
6422 : /// Lists the tenant shards that has unhealthy metadata status.
6423 0 : pub(crate) async fn metadata_health_list_unhealthy(
6424 0 : &self,
6425 0 : ) -> Result<Vec<TenantShardId>, ApiError> {
6426 0 : let result = self
6427 0 : .persistence
6428 0 : .list_unhealthy_metadata_health_records()
6429 0 : .await?
6430 0 : .iter()
6431 0 : .map(|p| p.get_tenant_shard_id().unwrap())
6432 0 : .collect();
6433 0 :
6434 0 : Ok(result)
6435 0 : }
6436 :
6437 : /// Lists the tenant shards that have not been scrubbed for some duration.
6438 0 : pub(crate) async fn metadata_health_list_outdated(
6439 0 : &self,
6440 0 : not_scrubbed_for: Duration,
6441 0 : ) -> Result<Vec<MetadataHealthRecord>, ApiError> {
6442 0 : let earlier = chrono::offset::Utc::now() - not_scrubbed_for;
6443 0 : let result = self
6444 0 : .persistence
6445 0 : .list_outdated_metadata_health_records(earlier)
6446 0 : .await?
6447 0 : .into_iter()
6448 0 : .map(|record| record.into())
6449 0 : .collect();
6450 0 : Ok(result)
6451 0 : }
6452 :
6453 0 : pub(crate) fn get_leadership_status(&self) -> LeadershipStatus {
6454 0 : self.inner.read().unwrap().get_leadership_status()
6455 0 : }
6456 :
6457 0 : pub(crate) async fn step_down(&self) -> GlobalObservedState {
6458 0 : tracing::info!("Received step down request from peer");
6459 0 : failpoint_support::sleep_millis_async!("sleep-on-step-down-handling");
6460 :
6461 0 : self.inner.write().unwrap().step_down();
6462 0 : // TODO: would it make sense to have a time-out for this?
6463 0 : self.stop_reconciliations(StopReconciliationsReason::SteppingDown)
6464 0 : .await;
6465 :
6466 0 : let mut global_observed = GlobalObservedState::default();
6467 0 : let locked = self.inner.read().unwrap();
6468 0 : for (tid, tenant_shard) in locked.tenants.iter() {
6469 0 : global_observed
6470 0 : .0
6471 0 : .insert(*tid, tenant_shard.observed.clone());
6472 0 : }
6473 :
6474 0 : global_observed
6475 0 : }
6476 : }
|
