Line data Source code
1 : use std::{
2 : borrow::Cow,
3 : cmp::Ordering,
4 : collections::{BTreeMap, HashMap, HashSet},
5 : str::FromStr,
6 : sync::Arc,
7 : time::{Duration, Instant},
8 : };
9 :
10 : use crate::{
11 : background_node_operations::{
12 : Drain, Fill, Operation, OperationError, OperationHandler, MAX_RECONCILES_PER_OPERATION,
13 : },
14 : compute_hook::NotifyError,
15 : id_lock_map::{trace_exclusive_lock, trace_shared_lock, IdLockMap, WrappedWriteGuard},
16 : persistence::{AbortShardSplitStatus, TenantFilter},
17 : reconciler::{ReconcileError, ReconcileUnits},
18 : scheduler::{MaySchedule, ScheduleContext, ScheduleMode},
19 : tenant_shard::{
20 : MigrateAttachment, ReconcileNeeded, ReconcilerStatus, ScheduleOptimization,
21 : ScheduleOptimizationAction,
22 : },
23 : };
24 : use anyhow::Context;
25 : use control_plane::storage_controller::{
26 : AttachHookRequest, AttachHookResponse, InspectRequest, InspectResponse,
27 : };
28 : use diesel::result::DatabaseErrorKind;
29 : use futures::{stream::FuturesUnordered, StreamExt};
30 : use itertools::Itertools;
31 : use pageserver_api::{
32 : controller_api::{
33 : NodeAvailability, NodeRegisterRequest, NodeSchedulingPolicy, PlacementPolicy,
34 : ShardSchedulingPolicy, TenantCreateResponse, TenantCreateResponseShard,
35 : TenantDescribeResponse, TenantDescribeResponseShard, TenantLocateResponse,
36 : TenantPolicyRequest, TenantShardMigrateRequest, TenantShardMigrateResponse,
37 : UtilizationScore,
38 : },
39 : models::{SecondaryProgress, TenantConfigRequest, TopTenantShardsRequest},
40 : };
41 : use reqwest::StatusCode;
42 : use tracing::{instrument, Instrument};
43 :
44 : use crate::pageserver_client::PageserverClient;
45 : use pageserver_api::{
46 : models::{
47 : self, LocationConfig, LocationConfigListResponse, LocationConfigMode,
48 : PageserverUtilization, ShardParameters, TenantConfig, TenantCreateRequest,
49 : TenantLocationConfigRequest, TenantLocationConfigResponse, TenantShardLocation,
50 : TenantShardSplitRequest, TenantShardSplitResponse, TenantTimeTravelRequest,
51 : TimelineCreateRequest, TimelineInfo,
52 : },
53 : shard::{ShardCount, ShardIdentity, ShardNumber, ShardStripeSize, TenantShardId},
54 : upcall_api::{
55 : ReAttachRequest, ReAttachResponse, ReAttachResponseTenant, ValidateRequest,
56 : ValidateResponse, ValidateResponseTenant,
57 : },
58 : };
59 : use pageserver_client::mgmt_api;
60 : use tokio::sync::mpsc::error::TrySendError;
61 : use tokio_util::sync::CancellationToken;
62 : use utils::{
63 : completion::Barrier,
64 : failpoint_support,
65 : generation::Generation,
66 : http::error::ApiError,
67 : id::{NodeId, TenantId, TimelineId},
68 : sync::gate::Gate,
69 : };
70 :
71 : use crate::{
72 : compute_hook::ComputeHook,
73 : heartbeater::{Heartbeater, PageserverState},
74 : node::{AvailabilityTransition, Node},
75 : persistence::{split_state::SplitState, DatabaseError, Persistence, TenantShardPersistence},
76 : reconciler::attached_location_conf,
77 : scheduler::Scheduler,
78 : tenant_shard::{
79 : IntentState, ObservedState, ObservedStateLocation, ReconcileResult, ReconcileWaitError,
80 : ReconcilerWaiter, TenantShard,
81 : },
82 : };
83 :
84 : // For operations that should be quick, like attaching a new tenant
85 : const SHORT_RECONCILE_TIMEOUT: Duration = Duration::from_secs(5);
86 :
87 : // For operations that might be slow, like migrating a tenant with
88 : // some data in it.
89 : pub const RECONCILE_TIMEOUT: Duration = Duration::from_secs(30);
90 :
91 : // If we receive a call using Secondary mode initially, it will omit generation. We will initialize
92 : // tenant shards into this generation, and as long as it remains in this generation, we will accept
93 : // input generation from future requests as authoritative.
94 : const INITIAL_GENERATION: Generation = Generation::new(0);
95 :
96 : /// How long [`Service::startup_reconcile`] is allowed to take before it should give
97 : /// up on unresponsive pageservers and proceed.
98 : pub(crate) const STARTUP_RECONCILE_TIMEOUT: Duration = Duration::from_secs(30);
99 :
100 : /// How long a node may be unresponsive to heartbeats before we declare it offline.
101 : /// This must be long enough to cover node restarts as well as normal operations: in future
102 : /// it should be separated into distinct timeouts for startup vs. normal operation
103 : /// (`<https://github.com/neondatabase/neon/issues/7552>`)
104 : pub const MAX_UNAVAILABLE_INTERVAL_DEFAULT: Duration = Duration::from_secs(300);
105 :
106 0 : #[derive(Clone, strum_macros::Display)]
107 : enum TenantOperations {
108 : Create,
109 : LocationConfig,
110 : ConfigSet,
111 : TimeTravelRemoteStorage,
112 : Delete,
113 : UpdatePolicy,
114 : ShardSplit,
115 : SecondaryDownload,
116 : TimelineCreate,
117 : TimelineDelete,
118 : }
119 :
120 0 : #[derive(Clone, strum_macros::Display)]
121 : enum NodeOperations {
122 : Register,
123 : Configure,
124 : }
125 :
126 : pub const RECONCILER_CONCURRENCY_DEFAULT: usize = 128;
127 :
128 : // Depth of the channel used to enqueue shards for reconciliation when they can't do it immediately.
129 : // This channel is finite-size to avoid using excessive memory if we get into a state where reconciles are finishing more slowly
130 : // than they're being pushed onto the queue.
131 : const MAX_DELAYED_RECONCILES: usize = 10000;
132 :
133 : // Top level state available to all HTTP handlers
134 : struct ServiceState {
135 : tenants: BTreeMap<TenantShardId, TenantShard>,
136 :
137 : nodes: Arc<HashMap<NodeId, Node>>,
138 :
139 : scheduler: Scheduler,
140 :
141 : /// Ongoing background operation on the cluster if any is running.
142 : /// Note that only one such operation may run at any given time,
143 : /// hence the type choice.
144 : ongoing_operation: Option<OperationHandler>,
145 :
146 : /// Queue of tenants who are waiting for concurrency limits to permit them to reconcile
147 : delayed_reconcile_rx: tokio::sync::mpsc::Receiver<TenantShardId>,
148 : }
149 :
150 : /// Transform an error from a pageserver into an error to return to callers of a storage
151 : /// controller API.
152 0 : fn passthrough_api_error(node: &Node, e: mgmt_api::Error) -> ApiError {
153 0 : match e {
154 0 : mgmt_api::Error::ReceiveErrorBody(str) => {
155 0 : // Presume errors receiving body are connectivity/availability issues
156 0 : ApiError::ResourceUnavailable(
157 0 : format!("{node} error receiving error body: {str}").into(),
158 0 : )
159 : }
160 0 : mgmt_api::Error::ReceiveBody(str) => {
161 0 : // Presume errors receiving body are connectivity/availability issues
162 0 : ApiError::ResourceUnavailable(format!("{node} error receiving body: {str}").into())
163 : }
164 0 : mgmt_api::Error::ApiError(StatusCode::NOT_FOUND, msg) => {
165 0 : ApiError::NotFound(anyhow::anyhow!(format!("{node}: {msg}")).into())
166 : }
167 0 : mgmt_api::Error::ApiError(StatusCode::SERVICE_UNAVAILABLE, msg) => {
168 0 : ApiError::ResourceUnavailable(format!("{node}: {msg}").into())
169 : }
170 0 : mgmt_api::Error::ApiError(status @ StatusCode::UNAUTHORIZED, msg)
171 0 : | mgmt_api::Error::ApiError(status @ StatusCode::FORBIDDEN, msg) => {
172 : // Auth errors talking to a pageserver are not auth errors for the caller: they are
173 : // internal server errors, showing that something is wrong with the pageserver or
174 : // storage controller's auth configuration.
175 0 : ApiError::InternalServerError(anyhow::anyhow!("{node} {status}: {msg}"))
176 : }
177 0 : mgmt_api::Error::ApiError(status, msg) => {
178 0 : // Presume general case of pageserver API errors is that we tried to do something
179 0 : // that can't be done right now.
180 0 : ApiError::Conflict(format!("{node} {status}: {status} {msg}"))
181 : }
182 0 : mgmt_api::Error::Cancelled => ApiError::ShuttingDown,
183 : }
184 0 : }
185 :
186 : impl ServiceState {
187 0 : fn new(
188 0 : nodes: HashMap<NodeId, Node>,
189 0 : tenants: BTreeMap<TenantShardId, TenantShard>,
190 0 : scheduler: Scheduler,
191 0 : delayed_reconcile_rx: tokio::sync::mpsc::Receiver<TenantShardId>,
192 0 : ) -> Self {
193 0 : Self {
194 0 : tenants,
195 0 : nodes: Arc::new(nodes),
196 0 : scheduler,
197 0 : ongoing_operation: None,
198 0 : delayed_reconcile_rx,
199 0 : }
200 0 : }
201 :
202 0 : fn parts_mut(
203 0 : &mut self,
204 0 : ) -> (
205 0 : &mut Arc<HashMap<NodeId, Node>>,
206 0 : &mut BTreeMap<TenantShardId, TenantShard>,
207 0 : &mut Scheduler,
208 0 : ) {
209 0 : (&mut self.nodes, &mut self.tenants, &mut self.scheduler)
210 0 : }
211 : }
212 :
213 : #[derive(Clone)]
214 : pub struct Config {
215 : // All pageservers managed by one instance of this service must have
216 : // the same public key. This JWT token will be used to authenticate
217 : // this service to the pageservers it manages.
218 : pub jwt_token: Option<String>,
219 :
220 : // This JWT token will be used to authenticate this service to the control plane.
221 : pub control_plane_jwt_token: Option<String>,
222 :
223 : /// Where the compute hook should send notifications of pageserver attachment locations
224 : /// (this URL points to the control plane in prod). If this is None, the compute hook will
225 : /// assume it is running in a test environment and try to update neon_local.
226 : pub compute_hook_url: Option<String>,
227 :
228 : /// Grace period within which a pageserver does not respond to heartbeats, but is still
229 : /// considered active. Once the grace period elapses, the next heartbeat failure will
230 : /// mark the pagseserver offline.
231 : pub max_unavailable_interval: Duration,
232 :
233 : /// How many Reconcilers may be spawned concurrently
234 : pub reconciler_concurrency: usize,
235 :
236 : /// How large must a shard grow in bytes before we split it?
237 : /// None disables auto-splitting.
238 : pub split_threshold: Option<u64>,
239 : }
240 :
241 : impl From<DatabaseError> for ApiError {
242 0 : fn from(err: DatabaseError) -> ApiError {
243 0 : match err {
244 0 : DatabaseError::Query(e) => ApiError::InternalServerError(e.into()),
245 : // FIXME: ApiError doesn't have an Unavailable variant, but ShuttingDown maps to 503.
246 : DatabaseError::Connection(_) | DatabaseError::ConnectionPool(_) => {
247 0 : ApiError::ShuttingDown
248 : }
249 0 : DatabaseError::Logical(reason) => {
250 0 : ApiError::InternalServerError(anyhow::anyhow!(reason))
251 : }
252 : }
253 0 : }
254 : }
255 :
256 : pub struct Service {
257 : inner: Arc<std::sync::RwLock<ServiceState>>,
258 : config: Config,
259 : persistence: Arc<Persistence>,
260 : compute_hook: Arc<ComputeHook>,
261 : result_tx: tokio::sync::mpsc::UnboundedSender<ReconcileResult>,
262 :
263 : heartbeater: Heartbeater,
264 :
265 : // Channel for background cleanup from failed operations that require cleanup, such as shard split
266 : abort_tx: tokio::sync::mpsc::UnboundedSender<TenantShardSplitAbort>,
267 :
268 : // Locking on a tenant granularity (covers all shards in the tenant):
269 : // - Take exclusively for rare operations that mutate the tenant's persistent state (e.g. create/delete/split)
270 : // - Take in shared mode for operations that need the set of shards to stay the same to complete reliably (e.g. timeline CRUD)
271 : tenant_op_locks: IdLockMap<TenantId, TenantOperations>,
272 :
273 : // Locking for node-mutating operations: take exclusively for operations that modify the node's persistent state, or
274 : // that transition it to/from Active.
275 : node_op_locks: IdLockMap<NodeId, NodeOperations>,
276 :
277 : // Limit how many Reconcilers we will spawn concurrently
278 : reconciler_concurrency: Arc<tokio::sync::Semaphore>,
279 :
280 : /// Queue of tenants who are waiting for concurrency limits to permit them to reconcile
281 : /// Send into this queue to promptly attempt to reconcile this shard next time units are available.
282 : ///
283 : /// Note that this state logically lives inside ServiceInner, but carrying Sender here makes the code simpler
284 : /// by avoiding needing a &mut ref to something inside the ServiceInner. This could be optimized to
285 : /// use a VecDeque instead of a channel to reduce synchronization overhead, at the cost of some code complexity.
286 : delayed_reconcile_tx: tokio::sync::mpsc::Sender<TenantShardId>,
287 :
288 : // Process shutdown will fire this token
289 : cancel: CancellationToken,
290 :
291 : // Background tasks will hold this gate
292 : gate: Gate,
293 :
294 : /// This waits for initial reconciliation with pageservers to complete. Until this barrier
295 : /// passes, it isn't safe to do any actions that mutate tenants.
296 : pub(crate) startup_complete: Barrier,
297 : }
298 :
299 : impl From<ReconcileWaitError> for ApiError {
300 0 : fn from(value: ReconcileWaitError) -> Self {
301 0 : match value {
302 0 : ReconcileWaitError::Shutdown => ApiError::ShuttingDown,
303 0 : e @ ReconcileWaitError::Timeout(_) => ApiError::Timeout(format!("{e}").into()),
304 0 : e @ ReconcileWaitError::Failed(..) => ApiError::InternalServerError(anyhow::anyhow!(e)),
305 : }
306 0 : }
307 : }
308 :
309 : impl From<OperationError> for ApiError {
310 0 : fn from(value: OperationError) -> Self {
311 0 : match value {
312 0 : OperationError::NodeStateChanged(err) | OperationError::FinalizeError(err) => {
313 0 : ApiError::InternalServerError(anyhow::anyhow!(err))
314 : }
315 0 : OperationError::Cancelled => ApiError::Conflict("Operation was cancelled".into()),
316 : }
317 0 : }
318 : }
319 :
320 : #[allow(clippy::large_enum_variant)]
321 : enum TenantCreateOrUpdate {
322 : Create(TenantCreateRequest),
323 : Update(Vec<ShardUpdate>),
324 : }
325 :
326 : struct ShardSplitParams {
327 : old_shard_count: ShardCount,
328 : new_shard_count: ShardCount,
329 : new_stripe_size: Option<ShardStripeSize>,
330 : targets: Vec<ShardSplitTarget>,
331 : policy: PlacementPolicy,
332 : config: TenantConfig,
333 : shard_ident: ShardIdentity,
334 : }
335 :
336 : // When preparing for a shard split, we may either choose to proceed with the split,
337 : // or find that the work is already done and return NoOp.
338 : enum ShardSplitAction {
339 : Split(ShardSplitParams),
340 : NoOp(TenantShardSplitResponse),
341 : }
342 :
343 : // A parent shard which will be split
344 : struct ShardSplitTarget {
345 : parent_id: TenantShardId,
346 : node: Node,
347 : child_ids: Vec<TenantShardId>,
348 : }
349 :
350 : /// When we tenant shard split operation fails, we may not be able to clean up immediately, because nodes
351 : /// might not be available. We therefore use a queue of abort operations processed in the background.
352 : struct TenantShardSplitAbort {
353 : tenant_id: TenantId,
354 : /// The target values from the request that failed
355 : new_shard_count: ShardCount,
356 : new_stripe_size: Option<ShardStripeSize>,
357 : /// Until this abort op is complete, no other operations may be done on the tenant
358 : _tenant_lock: WrappedWriteGuard<TenantOperations>,
359 : }
360 :
361 0 : #[derive(thiserror::Error, Debug)]
362 : enum TenantShardSplitAbortError {
363 : #[error(transparent)]
364 : Database(#[from] DatabaseError),
365 : #[error(transparent)]
366 : Remote(#[from] mgmt_api::Error),
367 : #[error("Unavailable")]
368 : Unavailable,
369 : }
370 :
371 : struct ShardUpdate {
372 : tenant_shard_id: TenantShardId,
373 : placement_policy: PlacementPolicy,
374 : tenant_config: TenantConfig,
375 :
376 : /// If this is None, generation is not updated.
377 : generation: Option<Generation>,
378 : }
379 :
380 : impl Service {
381 0 : pub fn get_config(&self) -> &Config {
382 0 : &self.config
383 0 : }
384 :
385 : /// Called once on startup, this function attempts to contact all pageservers to build an up-to-date
386 : /// view of the world, and determine which pageservers are responsive.
387 0 : #[instrument(skip_all)]
388 : async fn startup_reconcile(
389 : self: &Arc<Service>,
390 : bg_compute_notify_result_tx: tokio::sync::mpsc::Sender<
391 : Result<(), (TenantShardId, NotifyError)>,
392 : >,
393 : ) {
394 : // For all tenant shards, a vector of observed states on nodes (where None means
395 : // indeterminate, same as in [`ObservedStateLocation`])
396 : let mut observed: HashMap<TenantShardId, Vec<(NodeId, Option<LocationConfig>)>> =
397 : HashMap::new();
398 :
399 : // Startup reconciliation does I/O to other services: whether they
400 : // are responsive or not, we should aim to finish within our deadline, because:
401 : // - If we don't, a k8s readiness hook watching /ready will kill us.
402 : // - While we're waiting for startup reconciliation, we are not fully
403 : // available for end user operations like creating/deleting tenants and timelines.
404 : //
405 : // We set multiple deadlines to break up the time available between the phases of work: this is
406 : // arbitrary, but avoids a situation where the first phase could burn our entire timeout period.
407 : let start_at = Instant::now();
408 : let node_scan_deadline = start_at
409 : .checked_add(STARTUP_RECONCILE_TIMEOUT / 2)
410 : .expect("Reconcile timeout is a modest constant");
411 :
412 : // Accumulate a list of any tenant locations that ought to be detached
413 : let mut cleanup = Vec::new();
414 :
415 : let node_listings = self.scan_node_locations(node_scan_deadline).await;
416 : // Send initial heartbeat requests to nodes that replied to the location listing above.
417 : let nodes_online = self.initial_heartbeat_round(node_listings.keys()).await;
418 :
419 : for (node_id, list_response) in node_listings {
420 : let tenant_shards = list_response.tenant_shards;
421 : tracing::info!(
422 : "Received {} shard statuses from pageserver {}, setting it to Active",
423 : tenant_shards.len(),
424 : node_id
425 : );
426 :
427 : for (tenant_shard_id, conf_opt) in tenant_shards {
428 : let shard_observations = observed.entry(tenant_shard_id).or_default();
429 : shard_observations.push((node_id, conf_opt));
430 : }
431 : }
432 :
433 : // List of tenants for which we will attempt to notify compute of their location at startup
434 : let mut compute_notifications = Vec::new();
435 :
436 : // Populate intent and observed states for all tenants, based on reported state on pageservers
437 : tracing::info!("Populating tenant shards' states from initial pageserver scan...");
438 : let shard_count = {
439 : let mut locked = self.inner.write().unwrap();
440 : let (nodes, tenants, scheduler) = locked.parts_mut();
441 :
442 : // Mark nodes online if they responded to us: nodes are offline by default after a restart.
443 : let mut new_nodes = (**nodes).clone();
444 : for (node_id, node) in new_nodes.iter_mut() {
445 : if let Some(utilization) = nodes_online.get(node_id) {
446 : node.set_availability(NodeAvailability::Active(UtilizationScore(
447 : utilization.utilization_score,
448 : )));
449 : scheduler.node_upsert(node);
450 : }
451 : }
452 : *nodes = Arc::new(new_nodes);
453 :
454 : for (tenant_shard_id, shard_observations) in observed {
455 : for (node_id, observed_loc) in shard_observations {
456 : let Some(tenant_shard) = tenants.get_mut(&tenant_shard_id) else {
457 : cleanup.push((tenant_shard_id, node_id));
458 : continue;
459 : };
460 : tenant_shard
461 : .observed
462 : .locations
463 : .insert(node_id, ObservedStateLocation { conf: observed_loc });
464 : }
465 : }
466 :
467 : // Populate each tenant's intent state
468 : let mut schedule_context = ScheduleContext::default();
469 : for (tenant_shard_id, tenant_shard) in tenants.iter_mut() {
470 : if tenant_shard_id.shard_number == ShardNumber(0) {
471 : // Reset scheduling context each time we advance to the next Tenant
472 : schedule_context = ScheduleContext::default();
473 : }
474 :
475 : tenant_shard.intent_from_observed(scheduler);
476 : if let Err(e) = tenant_shard.schedule(scheduler, &mut schedule_context) {
477 : // Non-fatal error: we are unable to properly schedule the tenant, perhaps because
478 : // not enough pageservers are available. The tenant may well still be available
479 : // to clients.
480 : tracing::error!("Failed to schedule tenant {tenant_shard_id} at startup: {e}");
481 : } else {
482 : // If we're both intending and observed to be attached at a particular node, we will
483 : // emit a compute notification for this. In the case where our observed state does not
484 : // yet match our intent, we will eventually reconcile, and that will emit a compute notification.
485 : if let Some(attached_at) = tenant_shard.stably_attached() {
486 : compute_notifications.push((
487 : *tenant_shard_id,
488 : attached_at,
489 : tenant_shard.shard.stripe_size,
490 : ));
491 : }
492 : }
493 : }
494 :
495 : tenants.len()
496 : };
497 :
498 : // TODO: if any tenant's intent now differs from its loaded generation_pageserver, we should clear that
499 : // generation_pageserver in the database.
500 :
501 : // Emit compute hook notifications for all tenants which are already stably attached. Other tenants
502 : // will emit compute hook notifications when they reconcile.
503 : //
504 : // Ordering: our calls to notify_background synchronously establish a relative order for these notifications vs. any later
505 : // calls into the ComputeHook for the same tenant: we can leave these to run to completion in the background and any later
506 : // calls will be correctly ordered wrt these.
507 : //
508 : // Concurrency: we call notify_background for all tenants, which will create O(N) tokio tasks, but almost all of them
509 : // will just wait on the ComputeHook::API_CONCURRENCY semaphore immediately, so very cheap until they get that semaphore
510 : // unit and start doing I/O.
511 : tracing::info!(
512 : "Sending {} compute notifications",
513 : compute_notifications.len()
514 : );
515 : self.compute_hook.notify_background(
516 : compute_notifications,
517 : bg_compute_notify_result_tx.clone(),
518 : &self.cancel,
519 : );
520 :
521 : // Finally, now that the service is up and running, launch reconcile operations for any tenants
522 : // which require it: under normal circumstances this should only include tenants that were in some
523 : // transient state before we restarted, or any tenants whose compute hooks failed above.
524 : tracing::info!("Checking for shards in need of reconciliation...");
525 : let reconcile_tasks = self.reconcile_all();
526 : // We will not wait for these reconciliation tasks to run here: we're now done with startup and
527 : // normal operations may proceed.
528 :
529 : // Clean up any tenants that were found on pageservers but are not known to us. Do this in the
530 : // background because it does not need to complete in order to proceed with other work.
531 : if !cleanup.is_empty() {
532 : tracing::info!("Cleaning up {} locations in the background", cleanup.len());
533 : tokio::task::spawn({
534 : let cleanup_self = self.clone();
535 0 : async move { cleanup_self.cleanup_locations(cleanup).await }
536 : });
537 : }
538 :
539 : tracing::info!("Startup complete, spawned {reconcile_tasks} reconciliation tasks ({shard_count} shards total)");
540 : }
541 :
542 0 : async fn initial_heartbeat_round<'a>(
543 0 : &self,
544 0 : node_ids: impl Iterator<Item = &'a NodeId>,
545 0 : ) -> HashMap<NodeId, PageserverUtilization> {
546 0 : assert!(!self.startup_complete.is_ready());
547 :
548 0 : let all_nodes = {
549 0 : let locked = self.inner.read().unwrap();
550 0 : locked.nodes.clone()
551 0 : };
552 0 :
553 0 : let mut nodes_to_heartbeat = HashMap::new();
554 0 : for node_id in node_ids {
555 0 : match all_nodes.get(node_id) {
556 0 : Some(node) => {
557 0 : nodes_to_heartbeat.insert(*node_id, node.clone());
558 0 : }
559 : None => {
560 0 : tracing::warn!("Node {node_id} was removed during start-up");
561 : }
562 : }
563 : }
564 :
565 0 : tracing::info!("Sending initial heartbeats...");
566 0 : let res = self
567 0 : .heartbeater
568 0 : .heartbeat(Arc::new(nodes_to_heartbeat))
569 0 : .await;
570 :
571 0 : let mut online_nodes = HashMap::new();
572 0 : if let Ok(deltas) = res {
573 0 : for (node_id, status) in deltas.0 {
574 0 : match status {
575 0 : PageserverState::Available { utilization, .. } => {
576 0 : online_nodes.insert(node_id, utilization);
577 0 : }
578 0 : PageserverState::Offline => {}
579 : }
580 : }
581 0 : }
582 :
583 0 : online_nodes
584 0 : }
585 :
586 : /// Used during [`Self::startup_reconcile`]: issue GETs to all nodes concurrently, with a deadline.
587 : ///
588 : /// The result includes only nodes which responded within the deadline
589 0 : async fn scan_node_locations(
590 0 : &self,
591 0 : deadline: Instant,
592 0 : ) -> HashMap<NodeId, LocationConfigListResponse> {
593 0 : let nodes = {
594 0 : let locked = self.inner.read().unwrap();
595 0 : locked.nodes.clone()
596 0 : };
597 0 :
598 0 : let mut node_results = HashMap::new();
599 0 :
600 0 : let mut node_list_futs = FuturesUnordered::new();
601 0 :
602 0 : tracing::info!("Scanning shards on {} nodes...", nodes.len());
603 0 : for node in nodes.values() {
604 0 : node_list_futs.push({
605 0 : async move {
606 0 : tracing::info!("Scanning shards on node {node}...");
607 0 : let timeout = Duration::from_secs(1);
608 0 : let response = node
609 0 : .with_client_retries(
610 0 : |client| async move { client.list_location_config().await },
611 0 : &self.config.jwt_token,
612 0 : 1,
613 0 : 5,
614 0 : timeout,
615 0 : &self.cancel,
616 0 : )
617 0 : .await;
618 0 : (node.get_id(), response)
619 0 : }
620 0 : });
621 0 : }
622 :
623 : loop {
624 0 : let (node_id, result) = tokio::select! {
625 : next = node_list_futs.next() => {
626 : match next {
627 : Some(result) => result,
628 : None =>{
629 : // We got results for all our nodes
630 : break;
631 : }
632 :
633 : }
634 : },
635 : _ = tokio::time::sleep(deadline.duration_since(Instant::now())) => {
636 : // Give up waiting for anyone who hasn't responded: we will yield the results that we have
637 : tracing::info!("Reached deadline while waiting for nodes to respond to location listing requests");
638 : break;
639 : }
640 : };
641 :
642 0 : let Some(list_response) = result else {
643 0 : tracing::info!("Shutdown during startup_reconcile");
644 0 : break;
645 : };
646 :
647 0 : match list_response {
648 0 : Err(e) => {
649 0 : tracing::warn!("Could not scan node {} ({e})", node_id);
650 : }
651 0 : Ok(listing) => {
652 0 : node_results.insert(node_id, listing);
653 0 : }
654 : }
655 : }
656 :
657 0 : node_results
658 0 : }
659 :
660 : /// Used during [`Self::startup_reconcile`]: detach a list of unknown-to-us tenants from pageservers.
661 : ///
662 : /// This is safe to run in the background, because if we don't have this TenantShardId in our map of
663 : /// tenants, then it is probably something incompletely deleted before: we will not fight with any
664 : /// other task trying to attach it.
665 0 : #[instrument(skip_all)]
666 : async fn cleanup_locations(&self, cleanup: Vec<(TenantShardId, NodeId)>) {
667 : let nodes = self.inner.read().unwrap().nodes.clone();
668 :
669 : for (tenant_shard_id, node_id) in cleanup {
670 : // A node reported a tenant_shard_id which is unknown to us: detach it.
671 : let Some(node) = nodes.get(&node_id) else {
672 : // This is legitimate; we run in the background and [`Self::startup_reconcile`] might have identified
673 : // a location to clean up on a node that has since been removed.
674 : tracing::info!(
675 : "Not cleaning up location {node_id}/{tenant_shard_id}: node not found"
676 : );
677 : continue;
678 : };
679 :
680 : if self.cancel.is_cancelled() {
681 : break;
682 : }
683 :
684 : let client = PageserverClient::new(
685 : node.get_id(),
686 : node.base_url(),
687 : self.config.jwt_token.as_deref(),
688 : );
689 : match client
690 : .location_config(
691 : tenant_shard_id,
692 : LocationConfig {
693 : mode: LocationConfigMode::Detached,
694 : generation: None,
695 : secondary_conf: None,
696 : shard_number: tenant_shard_id.shard_number.0,
697 : shard_count: tenant_shard_id.shard_count.literal(),
698 : shard_stripe_size: 0,
699 : tenant_conf: models::TenantConfig::default(),
700 : },
701 : None,
702 : false,
703 : )
704 : .await
705 : {
706 : Ok(()) => {
707 : tracing::info!(
708 : "Detached unknown shard {tenant_shard_id} on pageserver {node_id}"
709 : );
710 : }
711 : Err(e) => {
712 : // Non-fatal error: leaving a tenant shard behind that we are not managing shouldn't
713 : // break anything.
714 : tracing::error!(
715 : "Failed to detach unknkown shard {tenant_shard_id} on pageserver {node_id}: {e}"
716 : );
717 : }
718 : }
719 : }
720 : }
721 :
722 : /// Long running background task that periodically wakes up and looks for shards that need
723 : /// reconciliation. Reconciliation is fallible, so any reconciliation tasks that fail during
724 : /// e.g. a tenant create/attach/migrate must eventually be retried: this task is responsible
725 : /// for those retries.
726 0 : #[instrument(skip_all)]
727 : async fn background_reconcile(self: &Arc<Self>) {
728 : self.startup_complete.clone().wait().await;
729 :
730 : const BACKGROUND_RECONCILE_PERIOD: Duration = Duration::from_secs(20);
731 :
732 : let mut interval = tokio::time::interval(BACKGROUND_RECONCILE_PERIOD);
733 : while !self.cancel.is_cancelled() {
734 : tokio::select! {
735 : _ = interval.tick() => {
736 : let reconciles_spawned = self.reconcile_all();
737 : if reconciles_spawned == 0 {
738 : // Run optimizer only when we didn't find any other work to do
739 : let optimizations = self.optimize_all().await;
740 : if optimizations == 0 {
741 : // Run new splits only when no optimizations are pending
742 : self.autosplit_tenants().await;
743 : }
744 : }
745 : }
746 : _ = self.cancel.cancelled() => return
747 : }
748 : }
749 : }
750 0 : #[instrument(skip_all)]
751 : async fn spawn_heartbeat_driver(&self) {
752 : self.startup_complete.clone().wait().await;
753 :
754 : const HEARTBEAT_INTERVAL: Duration = Duration::from_secs(5);
755 :
756 : let mut interval = tokio::time::interval(HEARTBEAT_INTERVAL);
757 : while !self.cancel.is_cancelled() {
758 : tokio::select! {
759 : _ = interval.tick() => { }
760 : _ = self.cancel.cancelled() => return
761 : };
762 :
763 : let nodes = {
764 : let locked = self.inner.read().unwrap();
765 : locked.nodes.clone()
766 : };
767 :
768 : let res = self.heartbeater.heartbeat(nodes).await;
769 : if let Ok(deltas) = res {
770 : for (node_id, state) in deltas.0 {
771 : let (new_node, new_availability) = match state {
772 : PageserverState::Available {
773 : utilization, new, ..
774 : } => (
775 : new,
776 : NodeAvailability::Active(UtilizationScore(
777 : utilization.utilization_score,
778 : )),
779 : ),
780 : PageserverState::Offline => (false, NodeAvailability::Offline),
781 : };
782 :
783 : if new_node {
784 : // When the heartbeats detect a newly added node, we don't wish
785 : // to attempt to reconcile the shards assigned to it. The node
786 : // is likely handling it's re-attach response, so reconciling now
787 : // would be counterproductive.
788 : //
789 : // Instead, update the in-memory state with the details learned about the
790 : // node.
791 : let mut locked = self.inner.write().unwrap();
792 : let (nodes, _tenants, scheduler) = locked.parts_mut();
793 :
794 : let mut new_nodes = (**nodes).clone();
795 :
796 : if let Some(node) = new_nodes.get_mut(&node_id) {
797 : node.set_availability(new_availability);
798 : scheduler.node_upsert(node);
799 : }
800 :
801 : locked.nodes = Arc::new(new_nodes);
802 : } else {
803 : // This is the code path for geniune availability transitions (i.e node
804 : // goes unavailable and/or comes back online).
805 : let res = self
806 : .node_configure(node_id, Some(new_availability), None)
807 : .await;
808 :
809 : match res {
810 : Ok(()) => {}
811 : Err(ApiError::NotFound(_)) => {
812 : // This should be rare, but legitimate since the heartbeats are done
813 : // on a snapshot of the nodes.
814 : tracing::info!(
815 : "Node {} was not found after heartbeat round",
816 : node_id
817 : );
818 : }
819 : Err(err) => {
820 : tracing::error!(
821 : "Failed to update node {} after heartbeat round: {}",
822 : node_id,
823 : err
824 : );
825 : }
826 : }
827 : }
828 : }
829 : }
830 : }
831 : }
832 :
833 : /// Apply the contents of a [`ReconcileResult`] to our in-memory state: if the reconciliation
834 : /// was successful and intent hasn't changed since the Reconciler was spawned, this will update
835 : /// the observed state of the tenant such that subsequent calls to [`TenantShard::get_reconcile_needed`]
836 : /// will indicate that reconciliation is not needed.
837 0 : #[instrument(skip_all, fields(
838 : tenant_id=%result.tenant_shard_id.tenant_id, shard_id=%result.tenant_shard_id.shard_slug(),
839 : sequence=%result.sequence
840 0 : ))]
841 : fn process_result(&self, result: ReconcileResult) {
842 : let mut locked = self.inner.write().unwrap();
843 : let Some(tenant) = locked.tenants.get_mut(&result.tenant_shard_id) else {
844 : // A reconciliation result might race with removing a tenant: drop results for
845 : // tenants that aren't in our map.
846 : return;
847 : };
848 :
849 : // Usually generation should only be updated via this path, so the max() isn't
850 : // needed, but it is used to handle out-of-band updates via. e.g. test hook.
851 : tenant.generation = std::cmp::max(tenant.generation, result.generation);
852 :
853 : // If the reconciler signals that it failed to notify compute, set this state on
854 : // the shard so that a future [`TenantShard::maybe_reconcile`] will try again.
855 : tenant.pending_compute_notification = result.pending_compute_notification;
856 :
857 : // Let the TenantShard know it is idle.
858 : tenant.reconcile_complete(result.sequence);
859 :
860 : match result.result {
861 : Ok(()) => {
862 : for (node_id, loc) in &result.observed.locations {
863 : if let Some(conf) = &loc.conf {
864 : tracing::info!("Updating observed location {}: {:?}", node_id, conf);
865 : } else {
866 : tracing::info!("Setting observed location {} to None", node_id,)
867 : }
868 : }
869 : tenant.observed = result.observed;
870 : tenant.waiter.advance(result.sequence);
871 : }
872 : Err(e) => {
873 : match e {
874 : ReconcileError::Cancel => {
875 : tracing::info!("Reconciler was cancelled");
876 : }
877 : ReconcileError::Remote(mgmt_api::Error::Cancelled) => {
878 : // This might be due to the reconciler getting cancelled, or it might
879 : // be due to the `Node` being marked offline.
880 : tracing::info!("Reconciler cancelled during pageserver API call");
881 : }
882 : _ => {
883 : tracing::warn!("Reconcile error: {}", e);
884 : }
885 : }
886 :
887 : // Ordering: populate last_error before advancing error_seq,
888 : // so that waiters will see the correct error after waiting.
889 : tenant.set_last_error(result.sequence, e);
890 :
891 : for (node_id, o) in result.observed.locations {
892 : tenant.observed.locations.insert(node_id, o);
893 : }
894 : }
895 : }
896 :
897 : // Maybe some other work can proceed now that this job finished.
898 : if self.reconciler_concurrency.available_permits() > 0 {
899 : while let Ok(tenant_shard_id) = locked.delayed_reconcile_rx.try_recv() {
900 : let (nodes, tenants, _scheduler) = locked.parts_mut();
901 : if let Some(shard) = tenants.get_mut(&tenant_shard_id) {
902 : shard.delayed_reconcile = false;
903 : self.maybe_reconcile_shard(shard, nodes);
904 : }
905 :
906 : if self.reconciler_concurrency.available_permits() == 0 {
907 : break;
908 : }
909 : }
910 : }
911 : }
912 :
913 0 : async fn process_results(
914 0 : &self,
915 0 : mut result_rx: tokio::sync::mpsc::UnboundedReceiver<ReconcileResult>,
916 0 : mut bg_compute_hook_result_rx: tokio::sync::mpsc::Receiver<
917 0 : Result<(), (TenantShardId, NotifyError)>,
918 0 : >,
919 0 : ) {
920 0 : loop {
921 0 : // Wait for the next result, or for cancellation
922 0 : tokio::select! {
923 : r = result_rx.recv() => {
924 : match r {
925 : Some(result) => {self.process_result(result);},
926 : None => {break;}
927 : }
928 : }
929 0 : _ = async{
930 0 : match bg_compute_hook_result_rx.recv().await {
931 0 : Some(result) => {
932 0 : if let Err((tenant_shard_id, notify_error)) = result {
933 0 : tracing::warn!("Marking shard {tenant_shard_id} for notification retry, due to error {notify_error}");
934 0 : let mut locked = self.inner.write().unwrap();
935 0 : if let Some(shard) = locked.tenants.get_mut(&tenant_shard_id) {
936 0 : shard.pending_compute_notification = true;
937 0 : }
938 :
939 0 : }
940 : },
941 : None => {
942 : // This channel is dead, but we don't want to terminate the outer loop{}: just wait for shutdown
943 0 : self.cancel.cancelled().await;
944 : }
945 : }
946 0 : } => {},
947 : _ = self.cancel.cancelled() => {
948 : break;
949 : }
950 0 : };
951 0 : }
952 :
953 : // We should only fall through on shutdown
954 0 : assert!(self.cancel.is_cancelled());
955 0 : }
956 :
957 0 : async fn process_aborts(
958 0 : &self,
959 0 : mut abort_rx: tokio::sync::mpsc::UnboundedReceiver<TenantShardSplitAbort>,
960 0 : ) {
961 : loop {
962 : // Wait for the next result, or for cancellation
963 0 : let op = tokio::select! {
964 : r = abort_rx.recv() => {
965 : match r {
966 : Some(op) => {op},
967 : None => {break;}
968 : }
969 : }
970 : _ = self.cancel.cancelled() => {
971 : break;
972 : }
973 : };
974 :
975 : // Retry until shutdown: we must keep this request object alive until it is properly
976 : // processed, as it holds a lock guard that prevents other operations trying to do things
977 : // to the tenant while it is in a weird part-split state.
978 0 : while !self.cancel.is_cancelled() {
979 0 : match self.abort_tenant_shard_split(&op).await {
980 0 : Ok(_) => break,
981 0 : Err(e) => {
982 0 : tracing::warn!(
983 0 : "Failed to abort shard split on {}, will retry: {e}",
984 : op.tenant_id
985 : );
986 :
987 : // If a node is unavailable, we hope that it has been properly marked Offline
988 : // when we retry, so that the abort op will succeed. If the abort op is failing
989 : // for some other reason, we will keep retrying forever, or until a human notices
990 : // and does something about it (either fixing a pageserver or restarting the controller).
991 0 : tokio::time::timeout(Duration::from_secs(5), self.cancel.cancelled())
992 0 : .await
993 0 : .ok();
994 : }
995 : }
996 : }
997 : }
998 0 : }
999 :
1000 0 : pub async fn spawn(config: Config, persistence: Arc<Persistence>) -> anyhow::Result<Arc<Self>> {
1001 0 : let (result_tx, result_rx) = tokio::sync::mpsc::unbounded_channel();
1002 0 : let (abort_tx, abort_rx) = tokio::sync::mpsc::unbounded_channel();
1003 0 :
1004 0 : tracing::info!("Loading nodes from database...");
1005 0 : let nodes = persistence
1006 0 : .list_nodes()
1007 0 : .await?
1008 0 : .into_iter()
1009 0 : .map(Node::from_persistent)
1010 0 : .collect::<Vec<_>>();
1011 0 : let nodes: HashMap<NodeId, Node> = nodes.into_iter().map(|n| (n.get_id(), n)).collect();
1012 0 : tracing::info!("Loaded {} nodes from database.", nodes.len());
1013 :
1014 0 : tracing::info!("Loading shards from database...");
1015 0 : let mut tenant_shard_persistence = persistence.list_tenant_shards().await?;
1016 0 : tracing::info!(
1017 0 : "Loaded {} shards from database.",
1018 0 : tenant_shard_persistence.len()
1019 : );
1020 :
1021 : // If any shard splits were in progress, reset the database state to abort them
1022 0 : let mut tenant_shard_count_min_max: HashMap<TenantId, (ShardCount, ShardCount)> =
1023 0 : HashMap::new();
1024 0 : for tsp in &mut tenant_shard_persistence {
1025 0 : let shard = tsp.get_shard_identity()?;
1026 0 : let tenant_shard_id = tsp.get_tenant_shard_id()?;
1027 0 : let entry = tenant_shard_count_min_max
1028 0 : .entry(tenant_shard_id.tenant_id)
1029 0 : .or_insert_with(|| (shard.count, shard.count));
1030 0 : entry.0 = std::cmp::min(entry.0, shard.count);
1031 0 : entry.1 = std::cmp::max(entry.1, shard.count);
1032 : }
1033 :
1034 0 : for (tenant_id, (count_min, count_max)) in tenant_shard_count_min_max {
1035 0 : if count_min != count_max {
1036 : // Aborting the split in the database and dropping the child shards is sufficient: the reconciliation in
1037 : // [`Self::startup_reconcile`] will implicitly drop the child shards on remote pageservers, or they'll
1038 : // be dropped later in [`Self::node_activate_reconcile`] if it isn't available right now.
1039 0 : tracing::info!("Aborting shard split {tenant_id} {count_min:?} -> {count_max:?}");
1040 0 : let abort_status = persistence.abort_shard_split(tenant_id, count_max).await?;
1041 :
1042 : // We may never see the Complete status here: if the split was complete, we wouldn't have
1043 : // identified this tenant has having mismatching min/max counts.
1044 0 : assert!(matches!(abort_status, AbortShardSplitStatus::Aborted));
1045 :
1046 : // Clear the splitting status in-memory, to reflect that we just aborted in the database
1047 0 : tenant_shard_persistence.iter_mut().for_each(|tsp| {
1048 0 : // Set idle split state on those shards that we will retain.
1049 0 : let tsp_tenant_id = TenantId::from_str(tsp.tenant_id.as_str()).unwrap();
1050 0 : if tsp_tenant_id == tenant_id
1051 0 : && tsp.get_shard_identity().unwrap().count == count_min
1052 0 : {
1053 0 : tsp.splitting = SplitState::Idle;
1054 0 : } else if tsp_tenant_id == tenant_id {
1055 : // Leave the splitting state on the child shards: this will be used next to
1056 : // drop them.
1057 0 : tracing::info!(
1058 0 : "Shard {tsp_tenant_id} will be dropped after shard split abort",
1059 : );
1060 0 : }
1061 0 : });
1062 0 :
1063 0 : // Drop shards for this tenant which we didn't just mark idle (i.e. child shards of the aborted split)
1064 0 : tenant_shard_persistence.retain(|tsp| {
1065 0 : TenantId::from_str(tsp.tenant_id.as_str()).unwrap() != tenant_id
1066 0 : || tsp.splitting == SplitState::Idle
1067 0 : });
1068 0 : }
1069 : }
1070 :
1071 0 : let mut tenants = BTreeMap::new();
1072 0 :
1073 0 : let mut scheduler = Scheduler::new(nodes.values());
1074 0 :
1075 0 : #[cfg(feature = "testing")]
1076 0 : {
1077 0 : // Hack: insert scheduler state for all nodes referenced by shards, as compatibility
1078 0 : // tests only store the shards, not the nodes. The nodes will be loaded shortly
1079 0 : // after when pageservers start up and register.
1080 0 : let mut node_ids = HashSet::new();
1081 0 : for tsp in &tenant_shard_persistence {
1082 0 : if let Some(node_id) = tsp.generation_pageserver {
1083 0 : node_ids.insert(node_id);
1084 0 : }
1085 : }
1086 0 : for node_id in node_ids {
1087 0 : tracing::info!("Creating node {} in scheduler for tests", node_id);
1088 0 : let node = Node::new(
1089 0 : NodeId(node_id as u64),
1090 0 : "".to_string(),
1091 0 : 123,
1092 0 : "".to_string(),
1093 0 : 123,
1094 0 : );
1095 0 :
1096 0 : scheduler.node_upsert(&node);
1097 : }
1098 : }
1099 0 : for tsp in tenant_shard_persistence {
1100 0 : let tenant_shard_id = tsp.get_tenant_shard_id()?;
1101 :
1102 : // We will populate intent properly later in [`Self::startup_reconcile`], initially populate
1103 : // it with what we can infer: the node for which a generation was most recently issued.
1104 0 : let mut intent = IntentState::new();
1105 0 : if let Some(generation_pageserver) = tsp.generation_pageserver {
1106 0 : intent.set_attached(&mut scheduler, Some(NodeId(generation_pageserver as u64)));
1107 0 : }
1108 0 : let new_tenant = TenantShard::from_persistent(tsp, intent)?;
1109 :
1110 0 : tenants.insert(tenant_shard_id, new_tenant);
1111 : }
1112 :
1113 0 : let (startup_completion, startup_complete) = utils::completion::channel();
1114 0 :
1115 0 : // This channel is continuously consumed by process_results, so doesn't need to be very large.
1116 0 : let (bg_compute_notify_result_tx, bg_compute_notify_result_rx) =
1117 0 : tokio::sync::mpsc::channel(512);
1118 0 :
1119 0 : let (delayed_reconcile_tx, delayed_reconcile_rx) =
1120 0 : tokio::sync::mpsc::channel(MAX_DELAYED_RECONCILES);
1121 0 :
1122 0 : let cancel = CancellationToken::new();
1123 0 : let heartbeater = Heartbeater::new(
1124 0 : config.jwt_token.clone(),
1125 0 : config.max_unavailable_interval,
1126 0 : cancel.clone(),
1127 0 : );
1128 0 : let this = Arc::new(Self {
1129 0 : inner: Arc::new(std::sync::RwLock::new(ServiceState::new(
1130 0 : nodes,
1131 0 : tenants,
1132 0 : scheduler,
1133 0 : delayed_reconcile_rx,
1134 0 : ))),
1135 0 : config: config.clone(),
1136 0 : persistence,
1137 0 : compute_hook: Arc::new(ComputeHook::new(config.clone())),
1138 0 : result_tx,
1139 0 : heartbeater,
1140 0 : reconciler_concurrency: Arc::new(tokio::sync::Semaphore::new(
1141 0 : config.reconciler_concurrency,
1142 0 : )),
1143 0 : delayed_reconcile_tx,
1144 0 : abort_tx,
1145 0 : startup_complete: startup_complete.clone(),
1146 0 : cancel,
1147 0 : gate: Gate::default(),
1148 0 : tenant_op_locks: Default::default(),
1149 0 : node_op_locks: Default::default(),
1150 0 : });
1151 0 :
1152 0 : let result_task_this = this.clone();
1153 0 : tokio::task::spawn(async move {
1154 : // Block shutdown until we're done (we must respect self.cancel)
1155 0 : if let Ok(_gate) = result_task_this.gate.enter() {
1156 0 : result_task_this
1157 0 : .process_results(result_rx, bg_compute_notify_result_rx)
1158 0 : .await
1159 0 : }
1160 0 : });
1161 0 :
1162 0 : tokio::task::spawn({
1163 0 : let this = this.clone();
1164 0 : async move {
1165 : // Block shutdown until we're done (we must respect self.cancel)
1166 0 : if let Ok(_gate) = this.gate.enter() {
1167 0 : this.process_aborts(abort_rx).await
1168 0 : }
1169 0 : }
1170 0 : });
1171 0 :
1172 0 : tokio::task::spawn({
1173 0 : let this = this.clone();
1174 0 : async move {
1175 0 : if let Ok(_gate) = this.gate.enter() {
1176 0 : loop {
1177 0 : tokio::select! {
1178 : _ = this.cancel.cancelled() => {
1179 : break;
1180 : },
1181 : _ = tokio::time::sleep(Duration::from_secs(60)) => {}
1182 0 : };
1183 0 : this.tenant_op_locks.housekeeping();
1184 0 : }
1185 0 : }
1186 0 : }
1187 0 : });
1188 0 :
1189 0 : tokio::task::spawn({
1190 0 : let this = this.clone();
1191 0 : // We will block the [`Service::startup_complete`] barrier until [`Self::startup_reconcile`]
1192 0 : // is done.
1193 0 : let startup_completion = startup_completion.clone();
1194 0 : async move {
1195 : // Block shutdown until we're done (we must respect self.cancel)
1196 0 : let Ok(_gate) = this.gate.enter() else {
1197 0 : return;
1198 : };
1199 :
1200 0 : this.startup_reconcile(bg_compute_notify_result_tx).await;
1201 0 : drop(startup_completion);
1202 0 : }
1203 0 : });
1204 0 :
1205 0 : tokio::task::spawn({
1206 0 : let this = this.clone();
1207 0 : let startup_complete = startup_complete.clone();
1208 0 : async move {
1209 0 : startup_complete.wait().await;
1210 0 : this.background_reconcile().await;
1211 0 : }
1212 0 : });
1213 0 :
1214 0 : tokio::task::spawn({
1215 0 : let this = this.clone();
1216 0 : let startup_complete = startup_complete.clone();
1217 0 : async move {
1218 0 : startup_complete.wait().await;
1219 0 : this.spawn_heartbeat_driver().await;
1220 0 : }
1221 0 : });
1222 0 :
1223 0 : Ok(this)
1224 0 : }
1225 :
1226 0 : pub(crate) async fn attach_hook(
1227 0 : &self,
1228 0 : attach_req: AttachHookRequest,
1229 0 : ) -> anyhow::Result<AttachHookResponse> {
1230 0 : // This is a test hook. To enable using it on tenants that were created directly with
1231 0 : // the pageserver API (not via this service), we will auto-create any missing tenant
1232 0 : // shards with default state.
1233 0 : let insert = {
1234 0 : let locked = self.inner.write().unwrap();
1235 0 : !locked.tenants.contains_key(&attach_req.tenant_shard_id)
1236 0 : };
1237 0 :
1238 0 : if insert {
1239 0 : let tsp = TenantShardPersistence {
1240 0 : tenant_id: attach_req.tenant_shard_id.tenant_id.to_string(),
1241 0 : shard_number: attach_req.tenant_shard_id.shard_number.0 as i32,
1242 0 : shard_count: attach_req.tenant_shard_id.shard_count.literal() as i32,
1243 0 : shard_stripe_size: 0,
1244 0 : generation: attach_req.generation_override.or(Some(0)),
1245 0 : generation_pageserver: None,
1246 0 : placement_policy: serde_json::to_string(&PlacementPolicy::Attached(0)).unwrap(),
1247 0 : config: serde_json::to_string(&TenantConfig::default()).unwrap(),
1248 0 : splitting: SplitState::default(),
1249 0 : scheduling_policy: serde_json::to_string(&ShardSchedulingPolicy::default())
1250 0 : .unwrap(),
1251 0 : };
1252 0 :
1253 0 : match self.persistence.insert_tenant_shards(vec![tsp]).await {
1254 0 : Err(e) => match e {
1255 : DatabaseError::Query(diesel::result::Error::DatabaseError(
1256 : DatabaseErrorKind::UniqueViolation,
1257 : _,
1258 : )) => {
1259 0 : tracing::info!(
1260 0 : "Raced with another request to insert tenant {}",
1261 : attach_req.tenant_shard_id
1262 : )
1263 : }
1264 0 : _ => return Err(e.into()),
1265 : },
1266 : Ok(()) => {
1267 0 : tracing::info!("Inserted shard {} in database", attach_req.tenant_shard_id);
1268 :
1269 0 : let mut locked = self.inner.write().unwrap();
1270 0 : locked.tenants.insert(
1271 0 : attach_req.tenant_shard_id,
1272 0 : TenantShard::new(
1273 0 : attach_req.tenant_shard_id,
1274 0 : ShardIdentity::unsharded(),
1275 0 : PlacementPolicy::Attached(0),
1276 0 : ),
1277 0 : );
1278 0 : tracing::info!("Inserted shard {} in memory", attach_req.tenant_shard_id);
1279 : }
1280 : }
1281 0 : }
1282 :
1283 0 : let new_generation = if let Some(req_node_id) = attach_req.node_id {
1284 0 : let maybe_tenant_conf = {
1285 0 : let locked = self.inner.write().unwrap();
1286 0 : locked
1287 0 : .tenants
1288 0 : .get(&attach_req.tenant_shard_id)
1289 0 : .map(|t| t.config.clone())
1290 0 : };
1291 0 :
1292 0 : match maybe_tenant_conf {
1293 0 : Some(conf) => {
1294 0 : let new_generation = self
1295 0 : .persistence
1296 0 : .increment_generation(attach_req.tenant_shard_id, req_node_id)
1297 0 : .await?;
1298 :
1299 : // Persist the placement policy update. This is required
1300 : // when we reattaching a detached tenant.
1301 0 : self.persistence
1302 0 : .update_tenant_shard(
1303 0 : TenantFilter::Shard(attach_req.tenant_shard_id),
1304 0 : Some(PlacementPolicy::Attached(0)),
1305 0 : Some(conf),
1306 0 : None,
1307 0 : None,
1308 0 : )
1309 0 : .await?;
1310 0 : Some(new_generation)
1311 : }
1312 : None => {
1313 0 : anyhow::bail!("Attach hook handling raced with tenant removal")
1314 : }
1315 : }
1316 : } else {
1317 0 : self.persistence.detach(attach_req.tenant_shard_id).await?;
1318 0 : None
1319 : };
1320 :
1321 0 : let mut locked = self.inner.write().unwrap();
1322 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
1323 0 :
1324 0 : let tenant_shard = tenants
1325 0 : .get_mut(&attach_req.tenant_shard_id)
1326 0 : .expect("Checked for existence above");
1327 :
1328 0 : if let Some(new_generation) = new_generation {
1329 0 : tenant_shard.generation = Some(new_generation);
1330 0 : tenant_shard.policy = PlacementPolicy::Attached(0);
1331 0 : } else {
1332 : // This is a detach notification. We must update placement policy to avoid re-attaching
1333 : // during background scheduling/reconciliation, or during storage controller restart.
1334 0 : assert!(attach_req.node_id.is_none());
1335 0 : tenant_shard.policy = PlacementPolicy::Detached;
1336 : }
1337 :
1338 0 : if let Some(attaching_pageserver) = attach_req.node_id.as_ref() {
1339 0 : tracing::info!(
1340 : tenant_id = %attach_req.tenant_shard_id,
1341 : ps_id = %attaching_pageserver,
1342 : generation = ?tenant_shard.generation,
1343 0 : "issuing",
1344 : );
1345 0 : } else if let Some(ps_id) = tenant_shard.intent.get_attached() {
1346 0 : tracing::info!(
1347 : tenant_id = %attach_req.tenant_shard_id,
1348 : %ps_id,
1349 : generation = ?tenant_shard.generation,
1350 0 : "dropping",
1351 : );
1352 : } else {
1353 0 : tracing::info!(
1354 : tenant_id = %attach_req.tenant_shard_id,
1355 0 : "no-op: tenant already has no pageserver");
1356 : }
1357 0 : tenant_shard
1358 0 : .intent
1359 0 : .set_attached(scheduler, attach_req.node_id);
1360 0 :
1361 0 : tracing::info!(
1362 0 : "attach_hook: tenant {} set generation {:?}, pageserver {}",
1363 0 : attach_req.tenant_shard_id,
1364 0 : tenant_shard.generation,
1365 0 : // TODO: this is an odd number of 0xf's
1366 0 : attach_req.node_id.unwrap_or(utils::id::NodeId(0xfffffff))
1367 : );
1368 :
1369 : // Trick the reconciler into not doing anything for this tenant: this helps
1370 : // tests that manually configure a tenant on the pagesrever, and then call this
1371 : // attach hook: they don't want background reconciliation to modify what they
1372 : // did to the pageserver.
1373 : #[cfg(feature = "testing")]
1374 : {
1375 0 : if let Some(node_id) = attach_req.node_id {
1376 0 : tenant_shard.observed.locations = HashMap::from([(
1377 0 : node_id,
1378 0 : ObservedStateLocation {
1379 0 : conf: Some(attached_location_conf(
1380 0 : tenant_shard.generation.unwrap(),
1381 0 : &tenant_shard.shard,
1382 0 : &tenant_shard.config,
1383 0 : false,
1384 0 : )),
1385 0 : },
1386 0 : )]);
1387 0 : } else {
1388 0 : tenant_shard.observed.locations.clear();
1389 0 : }
1390 : }
1391 :
1392 0 : Ok(AttachHookResponse {
1393 0 : gen: attach_req
1394 0 : .node_id
1395 0 : .map(|_| tenant_shard.generation.expect("Test hook, not used on tenants that are mid-onboarding with a NULL generation").into().unwrap()),
1396 0 : })
1397 0 : }
1398 :
1399 0 : pub(crate) fn inspect(&self, inspect_req: InspectRequest) -> InspectResponse {
1400 0 : let locked = self.inner.read().unwrap();
1401 0 :
1402 0 : let tenant_shard = locked.tenants.get(&inspect_req.tenant_shard_id);
1403 0 :
1404 0 : InspectResponse {
1405 0 : attachment: tenant_shard.and_then(|s| {
1406 0 : s.intent
1407 0 : .get_attached()
1408 0 : .map(|ps| (s.generation.expect("Test hook, not used on tenants that are mid-onboarding with a NULL generation").into().unwrap(), ps))
1409 0 : }),
1410 0 : }
1411 0 : }
1412 :
1413 : // When the availability state of a node transitions to active, we must do a full reconciliation
1414 : // of LocationConfigs on that node. This is because while a node was offline:
1415 : // - we might have proceeded through startup_reconcile without checking for extraneous LocationConfigs on this node
1416 : // - aborting a tenant shard split might have left rogue child shards behind on this node.
1417 : //
1418 : // This function must complete _before_ setting a `Node` to Active: once it is set to Active, other
1419 : // Reconcilers might communicate with the node, and these must not overlap with the work we do in
1420 : // this function.
1421 : //
1422 : // The reconciliation logic in here is very similar to what [`Self::startup_reconcile`] does, but
1423 : // for written for a single node rather than as a batch job for all nodes.
1424 0 : #[tracing::instrument(skip_all, fields(node_id=%node.get_id()))]
1425 : async fn node_activate_reconcile(
1426 : &self,
1427 : mut node: Node,
1428 : _lock: &WrappedWriteGuard<NodeOperations>,
1429 : ) -> Result<(), ApiError> {
1430 : // This Node is a mutable local copy: we will set it active so that we can use its
1431 : // API client to reconcile with the node. The Node in [`Self::nodes`] will get updated
1432 : // later.
1433 : node.set_availability(NodeAvailability::Active(UtilizationScore::worst()));
1434 :
1435 : let configs = match node
1436 : .with_client_retries(
1437 0 : |client| async move { client.list_location_config().await },
1438 : &self.config.jwt_token,
1439 : 1,
1440 : 5,
1441 : SHORT_RECONCILE_TIMEOUT,
1442 : &self.cancel,
1443 : )
1444 : .await
1445 : {
1446 : None => {
1447 : // We're shutting down (the Node's cancellation token can't have fired, because
1448 : // we're the only scope that has a reference to it, and we didn't fire it).
1449 : return Err(ApiError::ShuttingDown);
1450 : }
1451 : Some(Err(e)) => {
1452 : // This node didn't succeed listing its locations: it may not proceed to active state
1453 : // as it is apparently unavailable.
1454 : return Err(ApiError::PreconditionFailed(
1455 : format!("Failed to query node location configs, cannot activate ({e})").into(),
1456 : ));
1457 : }
1458 : Some(Ok(configs)) => configs,
1459 : };
1460 : tracing::info!("Loaded {} LocationConfigs", configs.tenant_shards.len());
1461 :
1462 : let mut cleanup = Vec::new();
1463 : {
1464 : let mut locked = self.inner.write().unwrap();
1465 :
1466 : for (tenant_shard_id, observed_loc) in configs.tenant_shards {
1467 : let Some(tenant_shard) = locked.tenants.get_mut(&tenant_shard_id) else {
1468 : cleanup.push(tenant_shard_id);
1469 : continue;
1470 : };
1471 : tenant_shard
1472 : .observed
1473 : .locations
1474 : .insert(node.get_id(), ObservedStateLocation { conf: observed_loc });
1475 : }
1476 : }
1477 :
1478 : for tenant_shard_id in cleanup {
1479 : tracing::info!("Detaching {tenant_shard_id}");
1480 : match node
1481 : .with_client_retries(
1482 0 : |client| async move {
1483 0 : let config = LocationConfig {
1484 0 : mode: LocationConfigMode::Detached,
1485 0 : generation: None,
1486 0 : secondary_conf: None,
1487 0 : shard_number: tenant_shard_id.shard_number.0,
1488 0 : shard_count: tenant_shard_id.shard_count.literal(),
1489 0 : shard_stripe_size: 0,
1490 0 : tenant_conf: models::TenantConfig::default(),
1491 0 : };
1492 0 : client
1493 0 : .location_config(tenant_shard_id, config, None, false)
1494 0 : .await
1495 0 : },
1496 : &self.config.jwt_token,
1497 : 1,
1498 : 5,
1499 : SHORT_RECONCILE_TIMEOUT,
1500 : &self.cancel,
1501 : )
1502 : .await
1503 : {
1504 : None => {
1505 : // We're shutting down (the Node's cancellation token can't have fired, because
1506 : // we're the only scope that has a reference to it, and we didn't fire it).
1507 : return Err(ApiError::ShuttingDown);
1508 : }
1509 : Some(Err(e)) => {
1510 : // Do not let the node proceed to Active state if it is not responsive to requests
1511 : // to detach. This could happen if e.g. a shutdown bug in the pageserver is preventing
1512 : // detach completing: we should not let this node back into the set of nodes considered
1513 : // okay for scheduling.
1514 : return Err(ApiError::Conflict(format!(
1515 : "Node {node} failed to detach {tenant_shard_id}: {e}"
1516 : )));
1517 : }
1518 : Some(Ok(_)) => {}
1519 : };
1520 : }
1521 :
1522 : Ok(())
1523 : }
1524 :
1525 0 : pub(crate) async fn re_attach(
1526 0 : &self,
1527 0 : reattach_req: ReAttachRequest,
1528 0 : ) -> Result<ReAttachResponse, ApiError> {
1529 0 : if let Some(register_req) = reattach_req.register {
1530 0 : self.node_register(register_req).await?;
1531 0 : }
1532 :
1533 : // Ordering: we must persist generation number updates before making them visible in the in-memory state
1534 0 : let incremented_generations = self.persistence.re_attach(reattach_req.node_id).await?;
1535 :
1536 0 : tracing::info!(
1537 : node_id=%reattach_req.node_id,
1538 0 : "Incremented {} tenant shards' generations",
1539 0 : incremented_generations.len()
1540 : );
1541 :
1542 : // Apply the updated generation to our in-memory state, and
1543 : // gather discover secondary locations.
1544 0 : let mut locked = self.inner.write().unwrap();
1545 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
1546 0 :
1547 0 : let mut response = ReAttachResponse {
1548 0 : tenants: Vec::new(),
1549 0 : };
1550 :
1551 : // TODO: cancel/restart any running reconciliation for this tenant, it might be trying
1552 : // to call location_conf API with an old generation. Wait for cancellation to complete
1553 : // before responding to this request. Requires well implemented CancellationToken logic
1554 : // all the way to where we call location_conf. Even then, there can still be a location_conf
1555 : // request in flight over the network: TODO handle that by making location_conf API refuse
1556 : // to go backward in generations.
1557 :
1558 : // Scan through all shards, applying updates for ones where we updated generation
1559 : // and identifying shards that intend to have a secondary location on this node.
1560 0 : for (tenant_shard_id, shard) in tenants {
1561 0 : if let Some(new_gen) = incremented_generations.get(tenant_shard_id) {
1562 0 : let new_gen = *new_gen;
1563 0 : response.tenants.push(ReAttachResponseTenant {
1564 0 : id: *tenant_shard_id,
1565 0 : gen: Some(new_gen.into().unwrap()),
1566 0 : // A tenant is only put into multi or stale modes in the middle of a [`Reconciler::live_migrate`]
1567 0 : // execution. If a pageserver is restarted during that process, then the reconcile pass will
1568 0 : // fail, and start from scratch, so it doesn't make sense for us to try and preserve
1569 0 : // the stale/multi states at this point.
1570 0 : mode: LocationConfigMode::AttachedSingle,
1571 0 : });
1572 0 :
1573 0 : shard.generation = std::cmp::max(shard.generation, Some(new_gen));
1574 0 : if let Some(observed) = shard.observed.locations.get_mut(&reattach_req.node_id) {
1575 : // Why can we update `observed` even though we're not sure our response will be received
1576 : // by the pageserver? Because the pageserver will not proceed with startup until
1577 : // it has processed response: if it loses it, we'll see another request and increment
1578 : // generation again, avoiding any uncertainty about dirtiness of tenant's state.
1579 0 : if let Some(conf) = observed.conf.as_mut() {
1580 0 : conf.generation = new_gen.into();
1581 0 : }
1582 0 : } else {
1583 0 : // This node has no observed state for the shard: perhaps it was offline
1584 0 : // when the pageserver restarted. Insert a None, so that the Reconciler
1585 0 : // will be prompted to learn the location's state before it makes changes.
1586 0 : shard
1587 0 : .observed
1588 0 : .locations
1589 0 : .insert(reattach_req.node_id, ObservedStateLocation { conf: None });
1590 0 : }
1591 0 : } else if shard.intent.get_secondary().contains(&reattach_req.node_id) {
1592 0 : // Ordering: pageserver will not accept /location_config requests until it has
1593 0 : // finished processing the response from re-attach. So we can update our in-memory state
1594 0 : // now, and be confident that we are not stamping on the result of some later location config.
1595 0 : // TODO: however, we are not strictly ordered wrt ReconcileResults queue,
1596 0 : // so we might update observed state here, and then get over-written by some racing
1597 0 : // ReconcileResult. The impact is low however, since we have set state on pageserver something
1598 0 : // that matches intent, so worst case if we race then we end up doing a spurious reconcile.
1599 0 :
1600 0 : response.tenants.push(ReAttachResponseTenant {
1601 0 : id: *tenant_shard_id,
1602 0 : gen: None,
1603 0 : mode: LocationConfigMode::Secondary,
1604 0 : });
1605 0 :
1606 0 : // We must not update observed, because we have no guarantee that our
1607 0 : // response will be received by the pageserver. This could leave it
1608 0 : // falsely dirty, but the resulting reconcile should be idempotent.
1609 0 : }
1610 : }
1611 :
1612 : // We consider a node Active once we have composed a re-attach response, but we
1613 : // do not call [`Self::node_activate_reconcile`]: the handling of the re-attach response
1614 : // implicitly synchronizes the LocationConfigs on the node.
1615 : //
1616 : // Setting a node active unblocks any Reconcilers that might write to the location config API,
1617 : // but those requests will not be accepted by the node until it has finished processing
1618 : // the re-attach response.
1619 : //
1620 : // Additionally, reset the nodes scheduling policy to match the conditional update done
1621 : // in [`Persistence::re_attach`].
1622 0 : if let Some(node) = nodes.get(&reattach_req.node_id) {
1623 0 : let reset_scheduling = matches!(
1624 0 : node.get_scheduling(),
1625 : NodeSchedulingPolicy::PauseForRestart
1626 : | NodeSchedulingPolicy::Draining
1627 : | NodeSchedulingPolicy::Filling
1628 : );
1629 :
1630 0 : if !node.is_available() || reset_scheduling {
1631 0 : let mut new_nodes = (**nodes).clone();
1632 0 : if let Some(node) = new_nodes.get_mut(&reattach_req.node_id) {
1633 0 : if !node.is_available() {
1634 0 : node.set_availability(NodeAvailability::Active(UtilizationScore::worst()));
1635 0 : }
1636 :
1637 0 : if reset_scheduling {
1638 0 : node.set_scheduling(NodeSchedulingPolicy::Active);
1639 0 : }
1640 :
1641 0 : scheduler.node_upsert(node);
1642 0 : let new_nodes = Arc::new(new_nodes);
1643 0 : *nodes = new_nodes;
1644 0 : }
1645 0 : }
1646 0 : }
1647 :
1648 0 : Ok(response)
1649 0 : }
1650 :
1651 0 : pub(crate) fn validate(&self, validate_req: ValidateRequest) -> ValidateResponse {
1652 0 : let locked = self.inner.read().unwrap();
1653 0 :
1654 0 : let mut response = ValidateResponse {
1655 0 : tenants: Vec::new(),
1656 0 : };
1657 :
1658 0 : for req_tenant in validate_req.tenants {
1659 0 : if let Some(tenant_shard) = locked.tenants.get(&req_tenant.id) {
1660 0 : let valid = tenant_shard.generation == Some(Generation::new(req_tenant.gen));
1661 0 : tracing::info!(
1662 0 : "handle_validate: {}(gen {}): valid={valid} (latest {:?})",
1663 : req_tenant.id,
1664 : req_tenant.gen,
1665 : tenant_shard.generation
1666 : );
1667 0 : response.tenants.push(ValidateResponseTenant {
1668 0 : id: req_tenant.id,
1669 0 : valid,
1670 0 : });
1671 0 : } else {
1672 0 : // After tenant deletion, we may approve any validation. This avoids
1673 0 : // spurious warnings on the pageserver if it has pending LSN updates
1674 0 : // at the point a deletion happens.
1675 0 : response.tenants.push(ValidateResponseTenant {
1676 0 : id: req_tenant.id,
1677 0 : valid: true,
1678 0 : });
1679 0 : }
1680 : }
1681 0 : response
1682 0 : }
1683 :
1684 0 : pub(crate) async fn tenant_create(
1685 0 : &self,
1686 0 : create_req: TenantCreateRequest,
1687 0 : ) -> Result<TenantCreateResponse, ApiError> {
1688 0 : let tenant_id = create_req.new_tenant_id.tenant_id;
1689 :
1690 : // Exclude any concurrent attempts to create/access the same tenant ID
1691 0 : let _tenant_lock = trace_exclusive_lock(
1692 0 : &self.tenant_op_locks,
1693 0 : create_req.new_tenant_id.tenant_id,
1694 0 : TenantOperations::Create,
1695 0 : )
1696 0 : .await;
1697 0 : let (response, waiters) = self.do_tenant_create(create_req).await?;
1698 :
1699 0 : if let Err(e) = self.await_waiters(waiters, RECONCILE_TIMEOUT).await {
1700 : // Avoid deadlock: reconcile may fail while notifying compute, if the cloud control plane refuses to
1701 : // accept compute notifications while it is in the process of creating. Reconciliation will
1702 : // be retried in the background.
1703 0 : tracing::warn!(%tenant_id, "Reconcile not done yet while creating tenant ({e})");
1704 0 : }
1705 0 : Ok(response)
1706 0 : }
1707 :
1708 0 : pub(crate) async fn do_tenant_create(
1709 0 : &self,
1710 0 : create_req: TenantCreateRequest,
1711 0 : ) -> Result<(TenantCreateResponse, Vec<ReconcilerWaiter>), ApiError> {
1712 0 : let placement_policy = create_req
1713 0 : .placement_policy
1714 0 : .clone()
1715 0 : // As a default, zero secondaries is convenient for tests that don't choose a policy.
1716 0 : .unwrap_or(PlacementPolicy::Attached(0));
1717 :
1718 : // This service expects to handle sharding itself: it is an error to try and directly create
1719 : // a particular shard here.
1720 0 : let tenant_id = if !create_req.new_tenant_id.is_unsharded() {
1721 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
1722 0 : "Attempted to create a specific shard, this API is for creating the whole tenant"
1723 0 : )));
1724 : } else {
1725 0 : create_req.new_tenant_id.tenant_id
1726 0 : };
1727 0 :
1728 0 : tracing::info!(
1729 0 : "Creating tenant {}, shard_count={:?}",
1730 : create_req.new_tenant_id,
1731 : create_req.shard_parameters.count,
1732 : );
1733 :
1734 0 : let create_ids = (0..create_req.shard_parameters.count.count())
1735 0 : .map(|i| TenantShardId {
1736 0 : tenant_id,
1737 0 : shard_number: ShardNumber(i),
1738 0 : shard_count: create_req.shard_parameters.count,
1739 0 : })
1740 0 : .collect::<Vec<_>>();
1741 :
1742 : // If the caller specifies a None generation, it means "start from default". This is different
1743 : // to [`Self::tenant_location_config`], where a None generation is used to represent
1744 : // an incompletely-onboarded tenant.
1745 0 : let initial_generation = if matches!(placement_policy, PlacementPolicy::Secondary) {
1746 0 : tracing::info!(
1747 0 : "tenant_create: secondary mode, generation is_some={}",
1748 0 : create_req.generation.is_some()
1749 : );
1750 0 : create_req.generation.map(Generation::new)
1751 : } else {
1752 0 : tracing::info!(
1753 0 : "tenant_create: not secondary mode, generation is_some={}",
1754 0 : create_req.generation.is_some()
1755 : );
1756 0 : Some(
1757 0 : create_req
1758 0 : .generation
1759 0 : .map(Generation::new)
1760 0 : .unwrap_or(INITIAL_GENERATION),
1761 0 : )
1762 : };
1763 :
1764 : // Ordering: we persist tenant shards before creating them on the pageserver. This enables a caller
1765 : // to clean up after themselves by issuing a tenant deletion if something goes wrong and we restart
1766 : // during the creation, rather than risking leaving orphan objects in S3.
1767 0 : let persist_tenant_shards = create_ids
1768 0 : .iter()
1769 0 : .map(|tenant_shard_id| TenantShardPersistence {
1770 0 : tenant_id: tenant_shard_id.tenant_id.to_string(),
1771 0 : shard_number: tenant_shard_id.shard_number.0 as i32,
1772 0 : shard_count: tenant_shard_id.shard_count.literal() as i32,
1773 0 : shard_stripe_size: create_req.shard_parameters.stripe_size.0 as i32,
1774 0 : generation: initial_generation.map(|g| g.into().unwrap() as i32),
1775 0 : // The pageserver is not known until scheduling happens: we will set this column when
1776 0 : // incrementing the generation the first time we attach to a pageserver.
1777 0 : generation_pageserver: None,
1778 0 : placement_policy: serde_json::to_string(&placement_policy).unwrap(),
1779 0 : config: serde_json::to_string(&create_req.config).unwrap(),
1780 0 : splitting: SplitState::default(),
1781 0 : scheduling_policy: serde_json::to_string(&ShardSchedulingPolicy::default())
1782 0 : .unwrap(),
1783 0 : })
1784 0 : .collect();
1785 0 :
1786 0 : match self
1787 0 : .persistence
1788 0 : .insert_tenant_shards(persist_tenant_shards)
1789 0 : .await
1790 : {
1791 0 : Ok(_) => {}
1792 : Err(DatabaseError::Query(diesel::result::Error::DatabaseError(
1793 : DatabaseErrorKind::UniqueViolation,
1794 : _,
1795 : ))) => {
1796 : // Unique key violation: this is probably a retry. Because the shard count is part of the unique key,
1797 : // if we see a unique key violation it means that the creation request's shard count matches the previous
1798 : // creation's shard count.
1799 0 : tracing::info!("Tenant shards already present in database, proceeding with idempotent creation...");
1800 : }
1801 : // Any other database error is unexpected and a bug.
1802 0 : Err(e) => return Err(ApiError::InternalServerError(anyhow::anyhow!(e))),
1803 : };
1804 :
1805 0 : let mut schedule_context = ScheduleContext::default();
1806 :
1807 0 : let (waiters, response_shards) = {
1808 0 : let mut locked = self.inner.write().unwrap();
1809 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
1810 0 :
1811 0 : let mut response_shards = Vec::new();
1812 0 : let mut schcedule_error = None;
1813 :
1814 0 : for tenant_shard_id in create_ids {
1815 0 : tracing::info!("Creating shard {tenant_shard_id}...");
1816 :
1817 : use std::collections::btree_map::Entry;
1818 0 : match tenants.entry(tenant_shard_id) {
1819 0 : Entry::Occupied(mut entry) => {
1820 0 : tracing::info!(
1821 0 : "Tenant shard {tenant_shard_id} already exists while creating"
1822 : );
1823 :
1824 : // TODO: schedule() should take an anti-affinity expression that pushes
1825 : // attached and secondary locations (independently) away frorm those
1826 : // pageservers also holding a shard for this tenant.
1827 :
1828 0 : entry
1829 0 : .get_mut()
1830 0 : .schedule(scheduler, &mut schedule_context)
1831 0 : .map_err(|e| {
1832 0 : ApiError::Conflict(format!(
1833 0 : "Failed to schedule shard {tenant_shard_id}: {e}"
1834 0 : ))
1835 0 : })?;
1836 :
1837 0 : if let Some(node_id) = entry.get().intent.get_attached() {
1838 0 : let generation = entry
1839 0 : .get()
1840 0 : .generation
1841 0 : .expect("Generation is set when in attached mode");
1842 0 : response_shards.push(TenantCreateResponseShard {
1843 0 : shard_id: tenant_shard_id,
1844 0 : node_id: *node_id,
1845 0 : generation: generation.into().unwrap(),
1846 0 : });
1847 0 : }
1848 :
1849 0 : continue;
1850 : }
1851 0 : Entry::Vacant(entry) => {
1852 0 : let state = entry.insert(TenantShard::new(
1853 0 : tenant_shard_id,
1854 0 : ShardIdentity::from_params(
1855 0 : tenant_shard_id.shard_number,
1856 0 : &create_req.shard_parameters,
1857 0 : ),
1858 0 : placement_policy.clone(),
1859 0 : ));
1860 0 :
1861 0 : state.generation = initial_generation;
1862 0 : state.config = create_req.config.clone();
1863 0 : if let Err(e) = state.schedule(scheduler, &mut schedule_context) {
1864 0 : schcedule_error = Some(e);
1865 0 : }
1866 :
1867 : // Only include shards in result if we are attaching: the purpose
1868 : // of the response is to tell the caller where the shards are attached.
1869 0 : if let Some(node_id) = state.intent.get_attached() {
1870 0 : let generation = state
1871 0 : .generation
1872 0 : .expect("Generation is set when in attached mode");
1873 0 : response_shards.push(TenantCreateResponseShard {
1874 0 : shard_id: tenant_shard_id,
1875 0 : node_id: *node_id,
1876 0 : generation: generation.into().unwrap(),
1877 0 : });
1878 0 : }
1879 : }
1880 : };
1881 : }
1882 :
1883 : // If we failed to schedule shards, then they are still created in the controller,
1884 : // but we return an error to the requester to avoid a silent failure when someone
1885 : // tries to e.g. create a tenant whose placement policy requires more nodes than
1886 : // are present in the system. We do this here rather than in the above loop, to
1887 : // avoid situations where we only create a subset of shards in the tenant.
1888 0 : if let Some(e) = schcedule_error {
1889 0 : return Err(ApiError::Conflict(format!(
1890 0 : "Failed to schedule shard(s): {e}"
1891 0 : )));
1892 0 : }
1893 0 :
1894 0 : let waiters = tenants
1895 0 : .range_mut(TenantShardId::tenant_range(tenant_id))
1896 0 : .filter_map(|(_shard_id, shard)| self.maybe_reconcile_shard(shard, nodes))
1897 0 : .collect::<Vec<_>>();
1898 0 : (waiters, response_shards)
1899 0 : };
1900 0 :
1901 0 : Ok((
1902 0 : TenantCreateResponse {
1903 0 : shards: response_shards,
1904 0 : },
1905 0 : waiters,
1906 0 : ))
1907 0 : }
1908 :
1909 : /// Helper for functions that reconcile a number of shards, and would like to do a timeout-bounded
1910 : /// wait for reconciliation to complete before responding.
1911 0 : async fn await_waiters(
1912 0 : &self,
1913 0 : waiters: Vec<ReconcilerWaiter>,
1914 0 : timeout: Duration,
1915 0 : ) -> Result<(), ReconcileWaitError> {
1916 0 : let deadline = Instant::now().checked_add(timeout).unwrap();
1917 0 : for waiter in waiters {
1918 0 : let timeout = deadline.duration_since(Instant::now());
1919 0 : waiter.wait_timeout(timeout).await?;
1920 : }
1921 :
1922 0 : Ok(())
1923 0 : }
1924 :
1925 : /// Same as [`Service::await_waiters`], but returns the waiters which are still
1926 : /// in progress
1927 0 : async fn await_waiters_remainder(
1928 0 : &self,
1929 0 : waiters: Vec<ReconcilerWaiter>,
1930 0 : timeout: Duration,
1931 0 : ) -> Vec<ReconcilerWaiter> {
1932 0 : let deadline = Instant::now().checked_add(timeout).unwrap();
1933 0 : for waiter in waiters.iter() {
1934 0 : let timeout = deadline.duration_since(Instant::now());
1935 0 : let _ = waiter.wait_timeout(timeout).await;
1936 : }
1937 :
1938 0 : waiters
1939 0 : .into_iter()
1940 0 : .filter(|waiter| matches!(waiter.get_status(), ReconcilerStatus::InProgress))
1941 0 : .collect::<Vec<_>>()
1942 0 : }
1943 :
1944 : /// Part of [`Self::tenant_location_config`]: dissect an incoming location config request,
1945 : /// and transform it into either a tenant creation of a series of shard updates.
1946 : ///
1947 : /// If the incoming request makes no changes, a [`TenantCreateOrUpdate::Update`] result will
1948 : /// still be returned.
1949 0 : fn tenant_location_config_prepare(
1950 0 : &self,
1951 0 : tenant_id: TenantId,
1952 0 : req: TenantLocationConfigRequest,
1953 0 : ) -> TenantCreateOrUpdate {
1954 0 : let mut updates = Vec::new();
1955 0 : let mut locked = self.inner.write().unwrap();
1956 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
1957 0 : let tenant_shard_id = TenantShardId::unsharded(tenant_id);
1958 :
1959 : // Use location config mode as an indicator of policy.
1960 0 : let placement_policy = match req.config.mode {
1961 0 : LocationConfigMode::Detached => PlacementPolicy::Detached,
1962 0 : LocationConfigMode::Secondary => PlacementPolicy::Secondary,
1963 : LocationConfigMode::AttachedMulti
1964 : | LocationConfigMode::AttachedSingle
1965 : | LocationConfigMode::AttachedStale => {
1966 0 : if nodes.len() > 1 {
1967 0 : PlacementPolicy::Attached(1)
1968 : } else {
1969 : // Convenience for dev/test: if we just have one pageserver, import
1970 : // tenants into non-HA mode so that scheduling will succeed.
1971 0 : PlacementPolicy::Attached(0)
1972 : }
1973 : }
1974 : };
1975 :
1976 0 : let mut create = true;
1977 0 : for (shard_id, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
1978 : // Saw an existing shard: this is not a creation
1979 0 : create = false;
1980 :
1981 : // Shards may have initially been created by a Secondary request, where we
1982 : // would have left generation as None.
1983 : //
1984 : // We only update generation the first time we see an attached-mode request,
1985 : // and if there is no existing generation set. The caller is responsible for
1986 : // ensuring that no non-storage-controller pageserver ever uses a higher
1987 : // generation than they passed in here.
1988 : use LocationConfigMode::*;
1989 0 : let set_generation = match req.config.mode {
1990 0 : AttachedMulti | AttachedSingle | AttachedStale if shard.generation.is_none() => {
1991 0 : req.config.generation.map(Generation::new)
1992 : }
1993 0 : _ => None,
1994 : };
1995 :
1996 0 : updates.push(ShardUpdate {
1997 0 : tenant_shard_id: *shard_id,
1998 0 : placement_policy: placement_policy.clone(),
1999 0 : tenant_config: req.config.tenant_conf.clone(),
2000 0 : generation: set_generation,
2001 0 : });
2002 : }
2003 :
2004 0 : if create {
2005 : use LocationConfigMode::*;
2006 0 : let generation = match req.config.mode {
2007 0 : AttachedMulti | AttachedSingle | AttachedStale => req.config.generation,
2008 : // If a caller provided a generation in a non-attached request, ignore it
2009 : // and leave our generation as None: this enables a subsequent update to set
2010 : // the generation when setting an attached mode for the first time.
2011 0 : _ => None,
2012 : };
2013 :
2014 0 : TenantCreateOrUpdate::Create(
2015 0 : // Synthesize a creation request
2016 0 : TenantCreateRequest {
2017 0 : new_tenant_id: tenant_shard_id,
2018 0 : generation,
2019 0 : shard_parameters: ShardParameters {
2020 0 : count: tenant_shard_id.shard_count,
2021 0 : // We only import un-sharded or single-sharded tenants, so stripe
2022 0 : // size can be made up arbitrarily here.
2023 0 : stripe_size: ShardParameters::DEFAULT_STRIPE_SIZE,
2024 0 : },
2025 0 : placement_policy: Some(placement_policy),
2026 0 : config: req.config.tenant_conf,
2027 0 : },
2028 0 : )
2029 : } else {
2030 0 : assert!(!updates.is_empty());
2031 0 : TenantCreateOrUpdate::Update(updates)
2032 : }
2033 0 : }
2034 :
2035 : /// This API is used by the cloud control plane to migrate unsharded tenants that it created
2036 : /// directly with pageservers into this service.
2037 : ///
2038 : /// Cloud control plane MUST NOT continue issuing GENERATION NUMBERS for this tenant once it
2039 : /// has attempted to call this API. Failure to oblige to this rule may lead to S3 corruption.
2040 : /// Think of the first attempt to call this API as a transfer of absolute authority over the
2041 : /// tenant's source of generation numbers.
2042 : ///
2043 : /// The mode in this request coarse-grained control of tenants:
2044 : /// - Call with mode Attached* to upsert the tenant.
2045 : /// - Call with mode Secondary to either onboard a tenant without attaching it, or
2046 : /// to set an existing tenant to PolicyMode::Secondary
2047 : /// - Call with mode Detached to switch to PolicyMode::Detached
2048 0 : pub(crate) async fn tenant_location_config(
2049 0 : &self,
2050 0 : tenant_shard_id: TenantShardId,
2051 0 : req: TenantLocationConfigRequest,
2052 0 : ) -> Result<TenantLocationConfigResponse, ApiError> {
2053 : // We require an exclusive lock, because we are updating both persistent and in-memory state
2054 0 : let _tenant_lock = trace_exclusive_lock(
2055 0 : &self.tenant_op_locks,
2056 0 : tenant_shard_id.tenant_id,
2057 0 : TenantOperations::LocationConfig,
2058 0 : )
2059 0 : .await;
2060 :
2061 0 : if !tenant_shard_id.is_unsharded() {
2062 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
2063 0 : "This API is for importing single-sharded or unsharded tenants"
2064 0 : )));
2065 0 : }
2066 0 :
2067 0 : // First check if this is a creation or an update
2068 0 : let create_or_update = self.tenant_location_config_prepare(tenant_shard_id.tenant_id, req);
2069 0 :
2070 0 : let mut result = TenantLocationConfigResponse {
2071 0 : shards: Vec::new(),
2072 0 : stripe_size: None,
2073 0 : };
2074 0 : let waiters = match create_or_update {
2075 0 : TenantCreateOrUpdate::Create(create_req) => {
2076 0 : let (create_resp, waiters) = self.do_tenant_create(create_req).await?;
2077 0 : result.shards = create_resp
2078 0 : .shards
2079 0 : .into_iter()
2080 0 : .map(|s| TenantShardLocation {
2081 0 : node_id: s.node_id,
2082 0 : shard_id: s.shard_id,
2083 0 : })
2084 0 : .collect();
2085 0 : waiters
2086 : }
2087 0 : TenantCreateOrUpdate::Update(updates) => {
2088 0 : // Persist updates
2089 0 : // Ordering: write to the database before applying changes in-memory, so that
2090 0 : // we will not appear time-travel backwards on a restart.
2091 0 : let mut schedule_context = ScheduleContext::default();
2092 : for ShardUpdate {
2093 0 : tenant_shard_id,
2094 0 : placement_policy,
2095 0 : tenant_config,
2096 0 : generation,
2097 0 : } in &updates
2098 : {
2099 0 : self.persistence
2100 0 : .update_tenant_shard(
2101 0 : TenantFilter::Shard(*tenant_shard_id),
2102 0 : Some(placement_policy.clone()),
2103 0 : Some(tenant_config.clone()),
2104 0 : *generation,
2105 0 : None,
2106 0 : )
2107 0 : .await?;
2108 : }
2109 :
2110 : // Apply updates in-memory
2111 0 : let mut waiters = Vec::new();
2112 0 : {
2113 0 : let mut locked = self.inner.write().unwrap();
2114 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
2115 :
2116 : for ShardUpdate {
2117 0 : tenant_shard_id,
2118 0 : placement_policy,
2119 0 : tenant_config,
2120 0 : generation: update_generation,
2121 0 : } in updates
2122 : {
2123 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
2124 0 : tracing::warn!("Shard {tenant_shard_id} removed while updating");
2125 0 : continue;
2126 : };
2127 :
2128 : // Update stripe size
2129 0 : if result.stripe_size.is_none() && shard.shard.count.count() > 1 {
2130 0 : result.stripe_size = Some(shard.shard.stripe_size);
2131 0 : }
2132 :
2133 0 : shard.policy = placement_policy;
2134 0 : shard.config = tenant_config;
2135 0 : if let Some(generation) = update_generation {
2136 0 : shard.generation = Some(generation);
2137 0 : }
2138 :
2139 0 : shard.schedule(scheduler, &mut schedule_context)?;
2140 :
2141 0 : let maybe_waiter = self.maybe_reconcile_shard(shard, nodes);
2142 0 : if let Some(waiter) = maybe_waiter {
2143 0 : waiters.push(waiter);
2144 0 : }
2145 :
2146 0 : if let Some(node_id) = shard.intent.get_attached() {
2147 0 : result.shards.push(TenantShardLocation {
2148 0 : shard_id: tenant_shard_id,
2149 0 : node_id: *node_id,
2150 0 : })
2151 0 : }
2152 : }
2153 : }
2154 0 : waiters
2155 : }
2156 : };
2157 :
2158 0 : if let Err(e) = self.await_waiters(waiters, SHORT_RECONCILE_TIMEOUT).await {
2159 : // Do not treat a reconcile error as fatal: we have already applied any requested
2160 : // Intent changes, and the reconcile can fail for external reasons like unavailable
2161 : // compute notification API. In these cases, it is important that we do not
2162 : // cause the cloud control plane to retry forever on this API.
2163 0 : tracing::warn!(
2164 0 : "Failed to reconcile after /location_config: {e}, returning success anyway"
2165 : );
2166 0 : }
2167 :
2168 : // Logging the full result is useful because it lets us cross-check what the cloud control
2169 : // plane's tenant_shards table should contain.
2170 0 : tracing::info!("Complete, returning {result:?}");
2171 :
2172 0 : Ok(result)
2173 0 : }
2174 :
2175 0 : pub(crate) async fn tenant_config_set(&self, req: TenantConfigRequest) -> Result<(), ApiError> {
2176 : // We require an exclusive lock, because we are updating persistent and in-memory state
2177 0 : let _tenant_lock = trace_exclusive_lock(
2178 0 : &self.tenant_op_locks,
2179 0 : req.tenant_id,
2180 0 : TenantOperations::ConfigSet,
2181 0 : )
2182 0 : .await;
2183 :
2184 0 : let tenant_id = req.tenant_id;
2185 0 : let config = req.config;
2186 0 :
2187 0 : self.persistence
2188 0 : .update_tenant_shard(
2189 0 : TenantFilter::Tenant(req.tenant_id),
2190 0 : None,
2191 0 : Some(config.clone()),
2192 0 : None,
2193 0 : None,
2194 0 : )
2195 0 : .await?;
2196 :
2197 0 : let waiters = {
2198 0 : let mut waiters = Vec::new();
2199 0 : let mut locked = self.inner.write().unwrap();
2200 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
2201 0 : for (_shard_id, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
2202 0 : shard.config = config.clone();
2203 0 : if let Some(waiter) = self.maybe_reconcile_shard(shard, nodes) {
2204 0 : waiters.push(waiter);
2205 0 : }
2206 : }
2207 0 : waiters
2208 : };
2209 :
2210 0 : if let Err(e) = self.await_waiters(waiters, SHORT_RECONCILE_TIMEOUT).await {
2211 : // Treat this as success because we have stored the configuration. If e.g.
2212 : // a node was unavailable at this time, it should not stop us accepting a
2213 : // configuration change.
2214 0 : tracing::warn!(%tenant_id, "Accepted configuration update but reconciliation failed: {e}");
2215 0 : }
2216 :
2217 0 : Ok(())
2218 0 : }
2219 :
2220 0 : pub(crate) fn tenant_config_get(
2221 0 : &self,
2222 0 : tenant_id: TenantId,
2223 0 : ) -> Result<HashMap<&str, serde_json::Value>, ApiError> {
2224 0 : let config = {
2225 0 : let locked = self.inner.read().unwrap();
2226 0 :
2227 0 : match locked
2228 0 : .tenants
2229 0 : .range(TenantShardId::tenant_range(tenant_id))
2230 0 : .next()
2231 : {
2232 0 : Some((_tenant_shard_id, shard)) => shard.config.clone(),
2233 : None => {
2234 0 : return Err(ApiError::NotFound(
2235 0 : anyhow::anyhow!("Tenant not found").into(),
2236 0 : ))
2237 : }
2238 : }
2239 : };
2240 :
2241 : // Unlike the pageserver, we do not have a set of global defaults: the config is
2242 : // entirely per-tenant. Therefore the distinction between `tenant_specific_overrides`
2243 : // and `effective_config` in the response is meaningless, but we retain that syntax
2244 : // in order to remain compatible with the pageserver API.
2245 :
2246 0 : let response = HashMap::from([
2247 : (
2248 : "tenant_specific_overrides",
2249 0 : serde_json::to_value(&config)
2250 0 : .context("serializing tenant specific overrides")
2251 0 : .map_err(ApiError::InternalServerError)?,
2252 : ),
2253 : (
2254 0 : "effective_config",
2255 0 : serde_json::to_value(&config)
2256 0 : .context("serializing effective config")
2257 0 : .map_err(ApiError::InternalServerError)?,
2258 : ),
2259 : ]);
2260 :
2261 0 : Ok(response)
2262 0 : }
2263 :
2264 0 : pub(crate) async fn tenant_time_travel_remote_storage(
2265 0 : &self,
2266 0 : time_travel_req: &TenantTimeTravelRequest,
2267 0 : tenant_id: TenantId,
2268 0 : timestamp: Cow<'_, str>,
2269 0 : done_if_after: Cow<'_, str>,
2270 0 : ) -> Result<(), ApiError> {
2271 0 : let _tenant_lock = trace_exclusive_lock(
2272 0 : &self.tenant_op_locks,
2273 0 : tenant_id,
2274 0 : TenantOperations::TimeTravelRemoteStorage,
2275 0 : )
2276 0 : .await;
2277 :
2278 0 : let node = {
2279 0 : let locked = self.inner.read().unwrap();
2280 : // Just a sanity check to prevent misuse: the API expects that the tenant is fully
2281 : // detached everywhere, and nothing writes to S3 storage. Here, we verify that,
2282 : // but only at the start of the process, so it's really just to prevent operator
2283 : // mistakes.
2284 0 : for (shard_id, shard) in locked.tenants.range(TenantShardId::tenant_range(tenant_id)) {
2285 0 : if shard.intent.get_attached().is_some() || !shard.intent.get_secondary().is_empty()
2286 : {
2287 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
2288 0 : "We want tenant to be attached in shard with tenant_shard_id={shard_id}"
2289 0 : )));
2290 0 : }
2291 0 : let maybe_attached = shard
2292 0 : .observed
2293 0 : .locations
2294 0 : .iter()
2295 0 : .filter_map(|(node_id, observed_location)| {
2296 0 : observed_location
2297 0 : .conf
2298 0 : .as_ref()
2299 0 : .map(|loc| (node_id, observed_location, loc.mode))
2300 0 : })
2301 0 : .find(|(_, _, mode)| *mode != LocationConfigMode::Detached);
2302 0 : if let Some((node_id, _observed_location, mode)) = maybe_attached {
2303 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!("We observed attached={mode:?} tenant in node_id={node_id} shard with tenant_shard_id={shard_id}")));
2304 0 : }
2305 : }
2306 0 : let scheduler = &locked.scheduler;
2307 : // Right now we only perform the operation on a single node without parallelization
2308 : // TODO fan out the operation to multiple nodes for better performance
2309 0 : let node_id = scheduler.schedule_shard(&[], &ScheduleContext::default())?;
2310 0 : let node = locked
2311 0 : .nodes
2312 0 : .get(&node_id)
2313 0 : .expect("Pageservers may not be deleted while lock is active");
2314 0 : node.clone()
2315 0 : };
2316 0 :
2317 0 : // The shard count is encoded in the remote storage's URL, so we need to handle all historically used shard counts
2318 0 : let mut counts = time_travel_req
2319 0 : .shard_counts
2320 0 : .iter()
2321 0 : .copied()
2322 0 : .collect::<HashSet<_>>()
2323 0 : .into_iter()
2324 0 : .collect::<Vec<_>>();
2325 0 : counts.sort_unstable();
2326 :
2327 0 : for count in counts {
2328 0 : let shard_ids = (0..count.count())
2329 0 : .map(|i| TenantShardId {
2330 0 : tenant_id,
2331 0 : shard_number: ShardNumber(i),
2332 0 : shard_count: count,
2333 0 : })
2334 0 : .collect::<Vec<_>>();
2335 0 : for tenant_shard_id in shard_ids {
2336 0 : let client = PageserverClient::new(
2337 0 : node.get_id(),
2338 0 : node.base_url(),
2339 0 : self.config.jwt_token.as_deref(),
2340 0 : );
2341 0 :
2342 0 : tracing::info!("Doing time travel recovery for shard {tenant_shard_id}",);
2343 :
2344 0 : client
2345 0 : .tenant_time_travel_remote_storage(
2346 0 : tenant_shard_id,
2347 0 : ×tamp,
2348 0 : &done_if_after,
2349 0 : )
2350 0 : .await
2351 0 : .map_err(|e| {
2352 0 : ApiError::InternalServerError(anyhow::anyhow!(
2353 0 : "Error doing time travel recovery for shard {tenant_shard_id} on node {}: {e}",
2354 0 : node
2355 0 : ))
2356 0 : })?;
2357 : }
2358 : }
2359 0 : Ok(())
2360 0 : }
2361 :
2362 0 : pub(crate) async fn tenant_secondary_download(
2363 0 : &self,
2364 0 : tenant_id: TenantId,
2365 0 : wait: Option<Duration>,
2366 0 : ) -> Result<(StatusCode, SecondaryProgress), ApiError> {
2367 0 : let _tenant_lock = trace_shared_lock(
2368 0 : &self.tenant_op_locks,
2369 0 : tenant_id,
2370 0 : TenantOperations::SecondaryDownload,
2371 0 : )
2372 0 : .await;
2373 :
2374 : // Acquire lock and yield the collection of shard-node tuples which we will send requests onward to
2375 0 : let targets = {
2376 0 : let locked = self.inner.read().unwrap();
2377 0 : let mut targets = Vec::new();
2378 :
2379 0 : for (tenant_shard_id, shard) in
2380 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
2381 : {
2382 0 : for node_id in shard.intent.get_secondary() {
2383 0 : let node = locked
2384 0 : .nodes
2385 0 : .get(node_id)
2386 0 : .expect("Pageservers may not be deleted while referenced");
2387 0 :
2388 0 : targets.push((*tenant_shard_id, node.clone()));
2389 0 : }
2390 : }
2391 0 : targets
2392 0 : };
2393 0 :
2394 0 : // Issue concurrent requests to all shards' locations
2395 0 : let mut futs = FuturesUnordered::new();
2396 0 : for (tenant_shard_id, node) in targets {
2397 0 : let client = PageserverClient::new(
2398 0 : node.get_id(),
2399 0 : node.base_url(),
2400 0 : self.config.jwt_token.as_deref(),
2401 0 : );
2402 0 : futs.push(async move {
2403 0 : let result = client
2404 0 : .tenant_secondary_download(tenant_shard_id, wait)
2405 0 : .await;
2406 0 : (result, node, tenant_shard_id)
2407 0 : })
2408 : }
2409 :
2410 : // Handle any errors returned by pageservers. This includes cases like this request racing with
2411 : // a scheduling operation, such that the tenant shard we're calling doesn't exist on that pageserver any more, as
2412 : // well as more general cases like 503s, 500s, or timeouts.
2413 0 : let mut aggregate_progress = SecondaryProgress::default();
2414 0 : let mut aggregate_status: Option<StatusCode> = None;
2415 0 : let mut error: Option<mgmt_api::Error> = None;
2416 0 : while let Some((result, node, tenant_shard_id)) = futs.next().await {
2417 0 : match result {
2418 0 : Err(e) => {
2419 0 : // Secondary downloads are always advisory: if something fails, we nevertheless report success, so that whoever
2420 0 : // is calling us will proceed with whatever migration they're doing, albeit with a slightly less warm cache
2421 0 : // than they had hoped for.
2422 0 : tracing::warn!("Secondary download error from pageserver {node}: {e}",);
2423 0 : error = Some(e)
2424 : }
2425 0 : Ok((status_code, progress)) => {
2426 0 : tracing::info!(%tenant_shard_id, "Shard status={status_code} progress: {progress:?}");
2427 0 : aggregate_progress.layers_downloaded += progress.layers_downloaded;
2428 0 : aggregate_progress.layers_total += progress.layers_total;
2429 0 : aggregate_progress.bytes_downloaded += progress.bytes_downloaded;
2430 0 : aggregate_progress.bytes_total += progress.bytes_total;
2431 0 : aggregate_progress.heatmap_mtime =
2432 0 : std::cmp::max(aggregate_progress.heatmap_mtime, progress.heatmap_mtime);
2433 0 : aggregate_status = match aggregate_status {
2434 0 : None => Some(status_code),
2435 0 : Some(StatusCode::OK) => Some(status_code),
2436 0 : Some(cur) => {
2437 0 : // Other status codes (e.g. 202) -- do not overwrite.
2438 0 : Some(cur)
2439 : }
2440 : };
2441 : }
2442 : }
2443 : }
2444 :
2445 : // If any of the shards return 202, indicate our result as 202.
2446 0 : match aggregate_status {
2447 : None => {
2448 0 : match error {
2449 0 : Some(e) => {
2450 0 : // No successes, and an error: surface it
2451 0 : Err(ApiError::Conflict(format!("Error from pageserver: {e}")))
2452 : }
2453 : None => {
2454 : // No shards found
2455 0 : Err(ApiError::NotFound(
2456 0 : anyhow::anyhow!("Tenant {} not found", tenant_id).into(),
2457 0 : ))
2458 : }
2459 : }
2460 : }
2461 0 : Some(aggregate_status) => Ok((aggregate_status, aggregate_progress)),
2462 : }
2463 0 : }
2464 :
2465 0 : pub(crate) async fn tenant_delete(&self, tenant_id: TenantId) -> Result<StatusCode, ApiError> {
2466 0 : let _tenant_lock =
2467 0 : trace_exclusive_lock(&self.tenant_op_locks, tenant_id, TenantOperations::Delete).await;
2468 :
2469 : // Detach all shards
2470 0 : let (detach_waiters, shard_ids, node) = {
2471 0 : let mut shard_ids = Vec::new();
2472 0 : let mut detach_waiters = Vec::new();
2473 0 : let mut locked = self.inner.write().unwrap();
2474 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
2475 0 : for (tenant_shard_id, shard) in
2476 0 : tenants.range_mut(TenantShardId::tenant_range(tenant_id))
2477 : {
2478 0 : shard_ids.push(*tenant_shard_id);
2479 0 :
2480 0 : // Update the tenant's intent to remove all attachments
2481 0 : shard.policy = PlacementPolicy::Detached;
2482 0 : shard
2483 0 : .schedule(scheduler, &mut ScheduleContext::default())
2484 0 : .expect("De-scheduling is infallible");
2485 0 : debug_assert!(shard.intent.get_attached().is_none());
2486 0 : debug_assert!(shard.intent.get_secondary().is_empty());
2487 :
2488 0 : if let Some(waiter) = self.maybe_reconcile_shard(shard, nodes) {
2489 0 : detach_waiters.push(waiter);
2490 0 : }
2491 : }
2492 :
2493 : // Pick an arbitrary node to use for remote deletions (does not have to be where the tenant
2494 : // was attached, just has to be able to see the S3 content)
2495 0 : let node_id = scheduler.schedule_shard(&[], &ScheduleContext::default())?;
2496 0 : let node = nodes
2497 0 : .get(&node_id)
2498 0 : .expect("Pageservers may not be deleted while lock is active");
2499 0 : (detach_waiters, shard_ids, node.clone())
2500 0 : };
2501 0 :
2502 0 : // This reconcile wait can fail in a few ways:
2503 0 : // A there is a very long queue for the reconciler semaphore
2504 0 : // B some pageserver is failing to handle a detach promptly
2505 0 : // C some pageserver goes offline right at the moment we send it a request.
2506 0 : //
2507 0 : // A and C are transient: the semaphore will eventually become available, and once a node is marked offline
2508 0 : // the next attempt to reconcile will silently skip detaches for an offline node and succeed. If B happens,
2509 0 : // it's a bug, and needs resolving at the pageserver level (we shouldn't just leave attachments behind while
2510 0 : // deleting the underlying data).
2511 0 : self.await_waiters(detach_waiters, RECONCILE_TIMEOUT)
2512 0 : .await?;
2513 :
2514 0 : let locations = shard_ids
2515 0 : .into_iter()
2516 0 : .map(|s| (s, node.clone()))
2517 0 : .collect::<Vec<_>>();
2518 0 : let results = self.tenant_for_shards_api(
2519 0 : locations,
2520 0 : |tenant_shard_id, client| async move { client.tenant_delete(tenant_shard_id).await },
2521 0 : 1,
2522 0 : 3,
2523 0 : RECONCILE_TIMEOUT,
2524 0 : &self.cancel,
2525 0 : )
2526 0 : .await;
2527 0 : for result in results {
2528 0 : match result {
2529 : Ok(StatusCode::ACCEPTED) => {
2530 : // This should never happen: we waited for detaches to finish above
2531 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
2532 0 : "Unexpectedly still attached on {}",
2533 0 : node
2534 0 : )));
2535 : }
2536 0 : Ok(_) => {}
2537 : Err(mgmt_api::Error::Cancelled) => {
2538 0 : return Err(ApiError::ShuttingDown);
2539 : }
2540 0 : Err(e) => {
2541 0 : // This is unexpected: remote deletion should be infallible, unless the object store
2542 0 : // at large is unavailable.
2543 0 : tracing::error!("Error deleting via node {}: {e}", node);
2544 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(e)));
2545 : }
2546 : }
2547 : }
2548 :
2549 : // Fall through: deletion of the tenant on pageservers is complete, we may proceed to drop
2550 : // our in-memory state and database state.
2551 :
2552 : // Ordering: we delete persistent state first: if we then
2553 : // crash, we will drop the in-memory state.
2554 :
2555 : // Drop persistent state.
2556 0 : self.persistence.delete_tenant(tenant_id).await?;
2557 :
2558 : // Drop in-memory state
2559 : {
2560 0 : let mut locked = self.inner.write().unwrap();
2561 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
2562 :
2563 : // Dereference Scheduler from shards before dropping them
2564 0 : for (_tenant_shard_id, shard) in
2565 0 : tenants.range_mut(TenantShardId::tenant_range(tenant_id))
2566 0 : {
2567 0 : shard.intent.clear(scheduler);
2568 0 : }
2569 :
2570 0 : tenants.retain(|tenant_shard_id, _shard| tenant_shard_id.tenant_id != tenant_id);
2571 0 : tracing::info!(
2572 0 : "Deleted tenant {tenant_id}, now have {} tenants",
2573 0 : locked.tenants.len()
2574 : );
2575 : };
2576 :
2577 : // Success is represented as 404, to imitate the existing pageserver deletion API
2578 0 : Ok(StatusCode::NOT_FOUND)
2579 0 : }
2580 :
2581 : /// Naming: this configures the storage controller's policies for a tenant, whereas [`Self::tenant_config_set`] is "set the TenantConfig"
2582 : /// for a tenant. The TenantConfig is passed through to pageservers, whereas this function modifies
2583 : /// the tenant's policies (configuration) within the storage controller
2584 0 : pub(crate) async fn tenant_update_policy(
2585 0 : &self,
2586 0 : tenant_id: TenantId,
2587 0 : req: TenantPolicyRequest,
2588 0 : ) -> Result<(), ApiError> {
2589 : // We require an exclusive lock, because we are updating persistent and in-memory state
2590 0 : let _tenant_lock = trace_exclusive_lock(
2591 0 : &self.tenant_op_locks,
2592 0 : tenant_id,
2593 0 : TenantOperations::UpdatePolicy,
2594 0 : )
2595 0 : .await;
2596 :
2597 0 : failpoint_support::sleep_millis_async!("tenant-update-policy-exclusive-lock");
2598 :
2599 : let TenantPolicyRequest {
2600 0 : placement,
2601 0 : scheduling,
2602 0 : } = req;
2603 0 :
2604 0 : self.persistence
2605 0 : .update_tenant_shard(
2606 0 : TenantFilter::Tenant(tenant_id),
2607 0 : placement.clone(),
2608 0 : None,
2609 0 : None,
2610 0 : scheduling,
2611 0 : )
2612 0 : .await?;
2613 :
2614 0 : let mut schedule_context = ScheduleContext::default();
2615 0 : let mut locked = self.inner.write().unwrap();
2616 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
2617 0 : for (shard_id, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
2618 0 : if let Some(placement) = &placement {
2619 0 : shard.policy = placement.clone();
2620 0 :
2621 0 : tracing::info!(tenant_id=%shard_id.tenant_id, shard_id=%shard_id.shard_slug(),
2622 0 : "Updated placement policy to {placement:?}");
2623 0 : }
2624 :
2625 0 : if let Some(scheduling) = &scheduling {
2626 0 : shard.set_scheduling_policy(*scheduling);
2627 0 :
2628 0 : tracing::info!(tenant_id=%shard_id.tenant_id, shard_id=%shard_id.shard_slug(),
2629 0 : "Updated scheduling policy to {scheduling:?}");
2630 0 : }
2631 :
2632 : // In case scheduling is being switched back on, try it now.
2633 0 : shard.schedule(scheduler, &mut schedule_context).ok();
2634 0 : self.maybe_reconcile_shard(shard, nodes);
2635 : }
2636 :
2637 0 : Ok(())
2638 0 : }
2639 :
2640 0 : pub(crate) async fn tenant_timeline_create(
2641 0 : &self,
2642 0 : tenant_id: TenantId,
2643 0 : mut create_req: TimelineCreateRequest,
2644 0 : ) -> Result<TimelineInfo, ApiError> {
2645 0 : tracing::info!(
2646 0 : "Creating timeline {}/{}",
2647 : tenant_id,
2648 : create_req.new_timeline_id,
2649 : );
2650 :
2651 0 : let _tenant_lock = trace_shared_lock(
2652 0 : &self.tenant_op_locks,
2653 0 : tenant_id,
2654 0 : TenantOperations::TimelineCreate,
2655 0 : )
2656 0 : .await;
2657 :
2658 0 : self.ensure_attached_wait(tenant_id).await?;
2659 :
2660 0 : let mut targets = {
2661 0 : let locked = self.inner.read().unwrap();
2662 0 : let mut targets = Vec::new();
2663 :
2664 0 : for (tenant_shard_id, shard) in
2665 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
2666 0 : {
2667 0 : let node_id = shard.intent.get_attached().ok_or_else(|| {
2668 0 : ApiError::InternalServerError(anyhow::anyhow!("Shard not scheduled"))
2669 0 : })?;
2670 0 : let node = locked
2671 0 : .nodes
2672 0 : .get(&node_id)
2673 0 : .expect("Pageservers may not be deleted while referenced");
2674 0 :
2675 0 : targets.push((*tenant_shard_id, node.clone()));
2676 : }
2677 0 : targets
2678 0 : };
2679 0 :
2680 0 : if targets.is_empty() {
2681 0 : return Err(ApiError::NotFound(
2682 0 : anyhow::anyhow!("Tenant not found").into(),
2683 0 : ));
2684 0 : };
2685 0 : let shard_zero = targets.remove(0);
2686 :
2687 0 : async fn create_one(
2688 0 : tenant_shard_id: TenantShardId,
2689 0 : node: Node,
2690 0 : jwt: Option<String>,
2691 0 : create_req: TimelineCreateRequest,
2692 0 : ) -> Result<TimelineInfo, ApiError> {
2693 0 : tracing::info!(
2694 0 : "Creating timeline on shard {}/{}, attached to node {node}",
2695 : tenant_shard_id,
2696 : create_req.new_timeline_id,
2697 : );
2698 0 : let client = PageserverClient::new(node.get_id(), node.base_url(), jwt.as_deref());
2699 0 :
2700 0 : client
2701 0 : .timeline_create(tenant_shard_id, &create_req)
2702 0 : .await
2703 0 : .map_err(|e| passthrough_api_error(&node, e))
2704 0 : }
2705 :
2706 : // Because the caller might not provide an explicit LSN, we must do the creation first on a single shard, and then
2707 : // use whatever LSN that shard picked when creating on subsequent shards. We arbitrarily use shard zero as the shard
2708 : // that will get the first creation request, and propagate the LSN to all the >0 shards.
2709 0 : let timeline_info = create_one(
2710 0 : shard_zero.0,
2711 0 : shard_zero.1,
2712 0 : self.config.jwt_token.clone(),
2713 0 : create_req.clone(),
2714 0 : )
2715 0 : .await?;
2716 :
2717 : // Propagate the LSN that shard zero picked, if caller didn't provide one
2718 0 : if create_req.ancestor_timeline_id.is_some() && create_req.ancestor_start_lsn.is_none() {
2719 0 : create_req.ancestor_start_lsn = timeline_info.ancestor_lsn;
2720 0 : }
2721 :
2722 : // Create timeline on remaining shards with number >0
2723 0 : if !targets.is_empty() {
2724 : // If we had multiple shards, issue requests for the remainder now.
2725 0 : let jwt = self.config.jwt_token.clone();
2726 0 : self.tenant_for_shards(targets, |tenant_shard_id: TenantShardId, node: Node| {
2727 0 : let create_req = create_req.clone();
2728 0 : Box::pin(create_one(tenant_shard_id, node, jwt.clone(), create_req))
2729 0 : })
2730 0 : .await?;
2731 0 : }
2732 :
2733 0 : Ok(timeline_info)
2734 0 : }
2735 :
2736 : /// Helper for concurrently calling a pageserver API on a number of shards, such as timeline creation.
2737 : ///
2738 : /// On success, the returned vector contains exactly the same number of elements as the input `locations`.
2739 0 : async fn tenant_for_shards<F, R>(
2740 0 : &self,
2741 0 : locations: Vec<(TenantShardId, Node)>,
2742 0 : mut req_fn: F,
2743 0 : ) -> Result<Vec<R>, ApiError>
2744 0 : where
2745 0 : F: FnMut(
2746 0 : TenantShardId,
2747 0 : Node,
2748 0 : )
2749 0 : -> std::pin::Pin<Box<dyn futures::Future<Output = Result<R, ApiError>> + Send>>,
2750 0 : {
2751 0 : let mut futs = FuturesUnordered::new();
2752 0 : let mut results = Vec::with_capacity(locations.len());
2753 :
2754 0 : for (tenant_shard_id, node) in locations {
2755 0 : futs.push(req_fn(tenant_shard_id, node));
2756 0 : }
2757 :
2758 0 : while let Some(r) = futs.next().await {
2759 0 : results.push(r?);
2760 : }
2761 :
2762 0 : Ok(results)
2763 0 : }
2764 :
2765 : /// Concurrently invoke a pageserver API call on many shards at once
2766 0 : pub(crate) async fn tenant_for_shards_api<T, O, F>(
2767 0 : &self,
2768 0 : locations: Vec<(TenantShardId, Node)>,
2769 0 : op: O,
2770 0 : warn_threshold: u32,
2771 0 : max_retries: u32,
2772 0 : timeout: Duration,
2773 0 : cancel: &CancellationToken,
2774 0 : ) -> Vec<mgmt_api::Result<T>>
2775 0 : where
2776 0 : O: Fn(TenantShardId, PageserverClient) -> F + Copy,
2777 0 : F: std::future::Future<Output = mgmt_api::Result<T>>,
2778 0 : {
2779 0 : let mut futs = FuturesUnordered::new();
2780 0 : let mut results = Vec::with_capacity(locations.len());
2781 :
2782 0 : for (tenant_shard_id, node) in locations {
2783 0 : futs.push(async move {
2784 0 : node.with_client_retries(
2785 0 : |client| op(tenant_shard_id, client),
2786 0 : &self.config.jwt_token,
2787 0 : warn_threshold,
2788 0 : max_retries,
2789 0 : timeout,
2790 0 : cancel,
2791 0 : )
2792 0 : .await
2793 0 : });
2794 0 : }
2795 :
2796 0 : while let Some(r) = futs.next().await {
2797 0 : let r = r.unwrap_or(Err(mgmt_api::Error::Cancelled));
2798 0 : results.push(r);
2799 0 : }
2800 :
2801 0 : results
2802 0 : }
2803 :
2804 0 : pub(crate) async fn tenant_timeline_delete(
2805 0 : &self,
2806 0 : tenant_id: TenantId,
2807 0 : timeline_id: TimelineId,
2808 0 : ) -> Result<StatusCode, ApiError> {
2809 0 : tracing::info!("Deleting timeline {}/{}", tenant_id, timeline_id,);
2810 0 : let _tenant_lock = trace_shared_lock(
2811 0 : &self.tenant_op_locks,
2812 0 : tenant_id,
2813 0 : TenantOperations::TimelineDelete,
2814 0 : )
2815 0 : .await;
2816 :
2817 0 : self.ensure_attached_wait(tenant_id).await?;
2818 :
2819 0 : let mut targets = {
2820 0 : let locked = self.inner.read().unwrap();
2821 0 : let mut targets = Vec::new();
2822 :
2823 0 : for (tenant_shard_id, shard) in
2824 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
2825 0 : {
2826 0 : let node_id = shard.intent.get_attached().ok_or_else(|| {
2827 0 : ApiError::InternalServerError(anyhow::anyhow!("Shard not scheduled"))
2828 0 : })?;
2829 0 : let node = locked
2830 0 : .nodes
2831 0 : .get(&node_id)
2832 0 : .expect("Pageservers may not be deleted while referenced");
2833 0 :
2834 0 : targets.push((*tenant_shard_id, node.clone()));
2835 : }
2836 0 : targets
2837 0 : };
2838 0 :
2839 0 : if targets.is_empty() {
2840 0 : return Err(ApiError::NotFound(
2841 0 : anyhow::anyhow!("Tenant not found").into(),
2842 0 : ));
2843 0 : }
2844 0 : let shard_zero = targets.remove(0);
2845 :
2846 0 : async fn delete_one(
2847 0 : tenant_shard_id: TenantShardId,
2848 0 : timeline_id: TimelineId,
2849 0 : node: Node,
2850 0 : jwt: Option<String>,
2851 0 : ) -> Result<StatusCode, ApiError> {
2852 0 : tracing::info!(
2853 0 : "Deleting timeline on shard {tenant_shard_id}/{timeline_id}, attached to node {node}",
2854 : );
2855 :
2856 0 : let client = PageserverClient::new(node.get_id(), node.base_url(), jwt.as_deref());
2857 0 : client
2858 0 : .timeline_delete(tenant_shard_id, timeline_id)
2859 0 : .await
2860 0 : .map_err(|e| {
2861 0 : ApiError::InternalServerError(anyhow::anyhow!(
2862 0 : "Error deleting timeline {timeline_id} on {tenant_shard_id} on node {node}: {e}",
2863 0 : ))
2864 0 : })
2865 0 : }
2866 :
2867 0 : let statuses = self
2868 0 : .tenant_for_shards(targets, |tenant_shard_id: TenantShardId, node: Node| {
2869 0 : Box::pin(delete_one(
2870 0 : tenant_shard_id,
2871 0 : timeline_id,
2872 0 : node,
2873 0 : self.config.jwt_token.clone(),
2874 0 : ))
2875 0 : })
2876 0 : .await?;
2877 :
2878 : // If any shards >0 haven't finished deletion yet, don't start deletion on shard zero
2879 0 : if statuses.iter().any(|s| s != &StatusCode::NOT_FOUND) {
2880 0 : return Ok(StatusCode::ACCEPTED);
2881 0 : }
2882 :
2883 : // Delete shard zero last: this is not strictly necessary, but since a caller's GET on a timeline will be routed
2884 : // to shard zero, it gives a more obvious behavior that a GET returns 404 once the deletion is done.
2885 0 : let shard_zero_status = delete_one(
2886 0 : shard_zero.0,
2887 0 : timeline_id,
2888 0 : shard_zero.1,
2889 0 : self.config.jwt_token.clone(),
2890 0 : )
2891 0 : .await?;
2892 :
2893 0 : Ok(shard_zero_status)
2894 0 : }
2895 :
2896 : /// When you need to send an HTTP request to the pageserver that holds shard0 of a tenant, this
2897 : /// function looks up and returns node. If the tenant isn't found, returns Err(ApiError::NotFound)
2898 0 : pub(crate) fn tenant_shard0_node(
2899 0 : &self,
2900 0 : tenant_id: TenantId,
2901 0 : ) -> Result<(Node, TenantShardId), ApiError> {
2902 0 : let locked = self.inner.read().unwrap();
2903 0 : let Some((tenant_shard_id, shard)) = locked
2904 0 : .tenants
2905 0 : .range(TenantShardId::tenant_range(tenant_id))
2906 0 : .next()
2907 : else {
2908 0 : return Err(ApiError::NotFound(
2909 0 : anyhow::anyhow!("Tenant {tenant_id} not found").into(),
2910 0 : ));
2911 : };
2912 :
2913 : // TODO: should use the ID last published to compute_hook, rather than the intent: the intent might
2914 : // point to somewhere we haven't attached yet.
2915 0 : let Some(node_id) = shard.intent.get_attached() else {
2916 0 : tracing::warn!(
2917 0 : tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(),
2918 0 : "Shard not scheduled (policy {:?}), cannot generate pass-through URL",
2919 : shard.policy
2920 : );
2921 0 : return Err(ApiError::Conflict(
2922 0 : "Cannot call timeline API on non-attached tenant".to_string(),
2923 0 : ));
2924 : };
2925 :
2926 0 : let Some(node) = locked.nodes.get(node_id) else {
2927 : // This should never happen
2928 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
2929 0 : "Shard refers to nonexistent node"
2930 0 : )));
2931 : };
2932 :
2933 0 : Ok((node.clone(), *tenant_shard_id))
2934 0 : }
2935 :
2936 0 : pub(crate) fn tenant_locate(
2937 0 : &self,
2938 0 : tenant_id: TenantId,
2939 0 : ) -> Result<TenantLocateResponse, ApiError> {
2940 0 : let locked = self.inner.read().unwrap();
2941 0 : tracing::info!("Locating shards for tenant {tenant_id}");
2942 :
2943 0 : let mut result = Vec::new();
2944 0 : let mut shard_params: Option<ShardParameters> = None;
2945 :
2946 0 : for (tenant_shard_id, shard) in locked.tenants.range(TenantShardId::tenant_range(tenant_id))
2947 : {
2948 0 : let node_id =
2949 0 : shard
2950 0 : .intent
2951 0 : .get_attached()
2952 0 : .ok_or(ApiError::BadRequest(anyhow::anyhow!(
2953 0 : "Cannot locate a tenant that is not attached"
2954 0 : )))?;
2955 :
2956 0 : let node = locked
2957 0 : .nodes
2958 0 : .get(&node_id)
2959 0 : .expect("Pageservers may not be deleted while referenced");
2960 0 :
2961 0 : result.push(node.shard_location(*tenant_shard_id));
2962 0 :
2963 0 : match &shard_params {
2964 0 : None => {
2965 0 : shard_params = Some(ShardParameters {
2966 0 : stripe_size: shard.shard.stripe_size,
2967 0 : count: shard.shard.count,
2968 0 : });
2969 0 : }
2970 0 : Some(params) => {
2971 0 : if params.stripe_size != shard.shard.stripe_size {
2972 : // This should never happen. We enforce at runtime because it's simpler than
2973 : // adding an extra per-tenant data structure to store the things that should be the same
2974 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
2975 0 : "Inconsistent shard stripe size parameters!"
2976 0 : )));
2977 0 : }
2978 : }
2979 : }
2980 : }
2981 :
2982 0 : if result.is_empty() {
2983 0 : return Err(ApiError::NotFound(
2984 0 : anyhow::anyhow!("No shards for this tenant ID found").into(),
2985 0 : ));
2986 0 : }
2987 0 : let shard_params = shard_params.expect("result is non-empty, therefore this is set");
2988 0 : tracing::info!(
2989 0 : "Located tenant {} with params {:?} on shards {}",
2990 0 : tenant_id,
2991 0 : shard_params,
2992 0 : result
2993 0 : .iter()
2994 0 : .map(|s| format!("{:?}", s))
2995 0 : .collect::<Vec<_>>()
2996 0 : .join(",")
2997 : );
2998 :
2999 0 : Ok(TenantLocateResponse {
3000 0 : shards: result,
3001 0 : shard_params,
3002 0 : })
3003 0 : }
3004 :
3005 : /// Returns None if the input iterator of shards does not include a shard with number=0
3006 0 : fn tenant_describe_impl<'a>(
3007 0 : &self,
3008 0 : shards: impl Iterator<Item = &'a TenantShard>,
3009 0 : ) -> Option<TenantDescribeResponse> {
3010 0 : let mut shard_zero = None;
3011 0 : let mut describe_shards = Vec::new();
3012 :
3013 0 : for shard in shards {
3014 0 : if shard.tenant_shard_id.is_shard_zero() {
3015 0 : shard_zero = Some(shard);
3016 0 : }
3017 :
3018 0 : describe_shards.push(TenantDescribeResponseShard {
3019 0 : tenant_shard_id: shard.tenant_shard_id,
3020 0 : node_attached: *shard.intent.get_attached(),
3021 0 : node_secondary: shard.intent.get_secondary().to_vec(),
3022 0 : last_error: shard
3023 0 : .last_error
3024 0 : .lock()
3025 0 : .unwrap()
3026 0 : .as_ref()
3027 0 : .map(|e| format!("{e}"))
3028 0 : .unwrap_or("".to_string())
3029 0 : .clone(),
3030 0 : is_reconciling: shard.reconciler.is_some(),
3031 0 : is_pending_compute_notification: shard.pending_compute_notification,
3032 0 : is_splitting: matches!(shard.splitting, SplitState::Splitting),
3033 0 : scheduling_policy: *shard.get_scheduling_policy(),
3034 : })
3035 : }
3036 :
3037 0 : let shard_zero = shard_zero?;
3038 :
3039 0 : Some(TenantDescribeResponse {
3040 0 : tenant_id: shard_zero.tenant_shard_id.tenant_id,
3041 0 : shards: describe_shards,
3042 0 : stripe_size: shard_zero.shard.stripe_size,
3043 0 : policy: shard_zero.policy.clone(),
3044 0 : config: shard_zero.config.clone(),
3045 0 : })
3046 0 : }
3047 :
3048 0 : pub(crate) fn tenant_describe(
3049 0 : &self,
3050 0 : tenant_id: TenantId,
3051 0 : ) -> Result<TenantDescribeResponse, ApiError> {
3052 0 : let locked = self.inner.read().unwrap();
3053 0 :
3054 0 : self.tenant_describe_impl(
3055 0 : locked
3056 0 : .tenants
3057 0 : .range(TenantShardId::tenant_range(tenant_id))
3058 0 : .map(|(_k, v)| v),
3059 0 : )
3060 0 : .ok_or_else(|| ApiError::NotFound(anyhow::anyhow!("Tenant {tenant_id} not found").into()))
3061 0 : }
3062 :
3063 0 : pub(crate) fn tenant_list(&self) -> Vec<TenantDescribeResponse> {
3064 0 : let locked = self.inner.read().unwrap();
3065 0 :
3066 0 : let mut result = Vec::new();
3067 0 : for (_tenant_id, tenant_shards) in
3068 0 : &locked.tenants.iter().group_by(|(id, _shard)| id.tenant_id)
3069 0 : {
3070 0 : result.push(
3071 0 : self.tenant_describe_impl(tenant_shards.map(|(_k, v)| v))
3072 0 : .expect("Groups are always non-empty"),
3073 0 : );
3074 0 : }
3075 :
3076 0 : result
3077 0 : }
3078 :
3079 0 : #[instrument(skip_all, fields(tenant_id=%op.tenant_id))]
3080 : async fn abort_tenant_shard_split(
3081 : &self,
3082 : op: &TenantShardSplitAbort,
3083 : ) -> Result<(), TenantShardSplitAbortError> {
3084 : // Cleaning up a split:
3085 : // - Parent shards are not destroyed during a split, just detached.
3086 : // - Failed pageserver split API calls can leave the remote node with just the parent attached,
3087 : // just the children attached, or both.
3088 : //
3089 : // Therefore our work to do is to:
3090 : // 1. Clean up storage controller's internal state to just refer to parents, no children
3091 : // 2. Call out to pageservers to ensure that children are detached
3092 : // 3. Call out to pageservers to ensure that parents are attached.
3093 : //
3094 : // Crash safety:
3095 : // - If the storage controller stops running during this cleanup *after* clearing the splitting state
3096 : // from our database, then [`Self::startup_reconcile`] will regard child attachments as garbage
3097 : // and detach them.
3098 : // - TODO: If the storage controller stops running during this cleanup *before* clearing the splitting state
3099 : // from our database, then we will re-enter this cleanup routine on startup.
3100 :
3101 : let TenantShardSplitAbort {
3102 : tenant_id,
3103 : new_shard_count,
3104 : new_stripe_size,
3105 : ..
3106 : } = op;
3107 :
3108 : // First abort persistent state, if any exists.
3109 : match self
3110 : .persistence
3111 : .abort_shard_split(*tenant_id, *new_shard_count)
3112 : .await?
3113 : {
3114 : AbortShardSplitStatus::Aborted => {
3115 : // Proceed to roll back any child shards created on pageservers
3116 : }
3117 : AbortShardSplitStatus::Complete => {
3118 : // The split completed (we might hit that path if e.g. our database transaction
3119 : // to write the completion landed in the database, but we dropped connection
3120 : // before seeing the result).
3121 : //
3122 : // We must update in-memory state to reflect the successful split.
3123 : self.tenant_shard_split_commit_inmem(
3124 : *tenant_id,
3125 : *new_shard_count,
3126 : *new_stripe_size,
3127 : );
3128 : return Ok(());
3129 : }
3130 : }
3131 :
3132 : // Clean up in-memory state, and accumulate the list of child locations that need detaching
3133 : let detach_locations: Vec<(Node, TenantShardId)> = {
3134 : let mut detach_locations = Vec::new();
3135 : let mut locked = self.inner.write().unwrap();
3136 : let (nodes, tenants, scheduler) = locked.parts_mut();
3137 :
3138 : for (tenant_shard_id, shard) in
3139 : tenants.range_mut(TenantShardId::tenant_range(op.tenant_id))
3140 : {
3141 : if shard.shard.count == op.new_shard_count {
3142 : // Surprising: the phase of [`Self::do_tenant_shard_split`] which inserts child shards in-memory
3143 : // is infallible, so if we got an error we shouldn't have got that far.
3144 : tracing::warn!(
3145 : "During split abort, child shard {tenant_shard_id} found in-memory"
3146 : );
3147 : continue;
3148 : }
3149 :
3150 : // Add the children of this shard to this list of things to detach
3151 : if let Some(node_id) = shard.intent.get_attached() {
3152 : for child_id in tenant_shard_id.split(*new_shard_count) {
3153 : detach_locations.push((
3154 : nodes
3155 : .get(node_id)
3156 : .expect("Intent references nonexistent node")
3157 : .clone(),
3158 : child_id,
3159 : ));
3160 : }
3161 : } else {
3162 : tracing::warn!(
3163 : "During split abort, shard {tenant_shard_id} has no attached location"
3164 : );
3165 : }
3166 :
3167 : tracing::info!("Restoring parent shard {tenant_shard_id}");
3168 : shard.splitting = SplitState::Idle;
3169 : if let Err(e) = shard.schedule(scheduler, &mut ScheduleContext::default()) {
3170 : // If this shard can't be scheduled now (perhaps due to offline nodes or
3171 : // capacity issues), that must not prevent us rolling back a split. In this
3172 : // case it should be eventually scheduled in the background.
3173 : tracing::warn!("Failed to schedule {tenant_shard_id} during shard abort: {e}")
3174 : }
3175 :
3176 : self.maybe_reconcile_shard(shard, nodes);
3177 : }
3178 :
3179 : // We don't expect any new_shard_count shards to exist here, but drop them just in case
3180 0 : tenants.retain(|_id, s| s.shard.count != *new_shard_count);
3181 :
3182 : detach_locations
3183 : };
3184 :
3185 : for (node, child_id) in detach_locations {
3186 : if !node.is_available() {
3187 : // An unavailable node cannot be cleaned up now: to avoid blocking forever, we will permit this, and
3188 : // rely on the reconciliation that happens when a node transitions to Active to clean up. Since we have
3189 : // removed child shards from our in-memory state and database, the reconciliation will implicitly remove
3190 : // them from the node.
3191 : tracing::warn!("Node {node} unavailable, can't clean up during split abort. It will be cleaned up when it is reactivated.");
3192 : continue;
3193 : }
3194 :
3195 : // Detach the remote child. If the pageserver split API call is still in progress, this call will get
3196 : // a 503 and retry, up to our limit.
3197 : tracing::info!("Detaching {child_id} on {node}...");
3198 : match node
3199 : .with_client_retries(
3200 0 : |client| async move {
3201 0 : let config = LocationConfig {
3202 0 : mode: LocationConfigMode::Detached,
3203 0 : generation: None,
3204 0 : secondary_conf: None,
3205 0 : shard_number: child_id.shard_number.0,
3206 0 : shard_count: child_id.shard_count.literal(),
3207 0 : // Stripe size and tenant config don't matter when detaching
3208 0 : shard_stripe_size: 0,
3209 0 : tenant_conf: TenantConfig::default(),
3210 0 : };
3211 0 :
3212 0 : client.location_config(child_id, config, None, false).await
3213 0 : },
3214 : &self.config.jwt_token,
3215 : 1,
3216 : 10,
3217 : Duration::from_secs(5),
3218 : &self.cancel,
3219 : )
3220 : .await
3221 : {
3222 : Some(Ok(_)) => {}
3223 : Some(Err(e)) => {
3224 : // We failed to communicate with the remote node. This is problematic: we may be
3225 : // leaving it with a rogue child shard.
3226 : tracing::warn!(
3227 : "Failed to detach child {child_id} from node {node} during abort"
3228 : );
3229 : return Err(e.into());
3230 : }
3231 : None => {
3232 : // Cancellation: we were shutdown or the node went offline. Shutdown is fine, we'll
3233 : // clean up on restart. The node going offline requires a retry.
3234 : return Err(TenantShardSplitAbortError::Unavailable);
3235 : }
3236 : };
3237 : }
3238 :
3239 : tracing::info!("Successfully aborted split");
3240 : Ok(())
3241 : }
3242 :
3243 : /// Infallible final stage of [`Self::tenant_shard_split`]: update the contents
3244 : /// of the tenant map to reflect the child shards that exist after the split.
3245 0 : fn tenant_shard_split_commit_inmem(
3246 0 : &self,
3247 0 : tenant_id: TenantId,
3248 0 : new_shard_count: ShardCount,
3249 0 : new_stripe_size: Option<ShardStripeSize>,
3250 0 : ) -> (
3251 0 : TenantShardSplitResponse,
3252 0 : Vec<(TenantShardId, NodeId, ShardStripeSize)>,
3253 0 : Vec<ReconcilerWaiter>,
3254 0 : ) {
3255 0 : let mut response = TenantShardSplitResponse {
3256 0 : new_shards: Vec::new(),
3257 0 : };
3258 0 : let mut child_locations = Vec::new();
3259 0 : let mut waiters = Vec::new();
3260 0 :
3261 0 : {
3262 0 : let mut locked = self.inner.write().unwrap();
3263 0 :
3264 0 : let parent_ids = locked
3265 0 : .tenants
3266 0 : .range(TenantShardId::tenant_range(tenant_id))
3267 0 : .map(|(shard_id, _)| *shard_id)
3268 0 : .collect::<Vec<_>>();
3269 0 :
3270 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
3271 0 : for parent_id in parent_ids {
3272 0 : let child_ids = parent_id.split(new_shard_count);
3273 :
3274 0 : let (pageserver, generation, policy, parent_ident, config) = {
3275 0 : let mut old_state = tenants
3276 0 : .remove(&parent_id)
3277 0 : .expect("It was present, we just split it");
3278 0 :
3279 0 : // A non-splitting state is impossible, because [`Self::tenant_shard_split`] holds
3280 0 : // a TenantId lock and passes it through to [`TenantShardSplitAbort`] in case of cleanup:
3281 0 : // nothing else can clear this.
3282 0 : assert!(matches!(old_state.splitting, SplitState::Splitting));
3283 :
3284 0 : let old_attached = old_state.intent.get_attached().unwrap();
3285 0 : old_state.intent.clear(scheduler);
3286 0 : let generation = old_state.generation.expect("Shard must have been attached");
3287 0 : (
3288 0 : old_attached,
3289 0 : generation,
3290 0 : old_state.policy,
3291 0 : old_state.shard,
3292 0 : old_state.config,
3293 0 : )
3294 0 : };
3295 0 :
3296 0 : let mut schedule_context = ScheduleContext::default();
3297 0 : for child in child_ids {
3298 0 : let mut child_shard = parent_ident;
3299 0 : child_shard.number = child.shard_number;
3300 0 : child_shard.count = child.shard_count;
3301 0 : if let Some(stripe_size) = new_stripe_size {
3302 0 : child_shard.stripe_size = stripe_size;
3303 0 : }
3304 :
3305 0 : let mut child_observed: HashMap<NodeId, ObservedStateLocation> = HashMap::new();
3306 0 : child_observed.insert(
3307 0 : pageserver,
3308 0 : ObservedStateLocation {
3309 0 : conf: Some(attached_location_conf(
3310 0 : generation,
3311 0 : &child_shard,
3312 0 : &config,
3313 0 : matches!(policy, PlacementPolicy::Attached(n) if n > 0),
3314 : )),
3315 : },
3316 : );
3317 :
3318 0 : let mut child_state = TenantShard::new(child, child_shard, policy.clone());
3319 0 : child_state.intent = IntentState::single(scheduler, Some(pageserver));
3320 0 : child_state.observed = ObservedState {
3321 0 : locations: child_observed,
3322 0 : };
3323 0 : child_state.generation = Some(generation);
3324 0 : child_state.config = config.clone();
3325 0 :
3326 0 : // The child's TenantShard::splitting is intentionally left at the default value of Idle,
3327 0 : // as at this point in the split process we have succeeded and this part is infallible:
3328 0 : // we will never need to do any special recovery from this state.
3329 0 :
3330 0 : child_locations.push((child, pageserver, child_shard.stripe_size));
3331 :
3332 0 : if let Err(e) = child_state.schedule(scheduler, &mut schedule_context) {
3333 : // This is not fatal, because we've implicitly already got an attached
3334 : // location for the child shard. Failure here just means we couldn't
3335 : // find a secondary (e.g. because cluster is overloaded).
3336 0 : tracing::warn!("Failed to schedule child shard {child}: {e}");
3337 0 : }
3338 : // In the background, attach secondary locations for the new shards
3339 0 : if let Some(waiter) = self.maybe_reconcile_shard(&mut child_state, nodes) {
3340 0 : waiters.push(waiter);
3341 0 : }
3342 :
3343 0 : tenants.insert(child, child_state);
3344 0 : response.new_shards.push(child);
3345 : }
3346 : }
3347 0 : (response, child_locations, waiters)
3348 0 : }
3349 0 : }
3350 :
3351 0 : async fn tenant_shard_split_start_secondaries(
3352 0 : &self,
3353 0 : tenant_id: TenantId,
3354 0 : waiters: Vec<ReconcilerWaiter>,
3355 0 : ) {
3356 : // Wait for initial reconcile of child shards, this creates the secondary locations
3357 0 : if let Err(e) = self.await_waiters(waiters, RECONCILE_TIMEOUT).await {
3358 : // This is not a failure to split: it's some issue reconciling the new child shards, perhaps
3359 : // their secondaries couldn't be attached.
3360 0 : tracing::warn!("Failed to reconcile after split: {e}");
3361 0 : return;
3362 0 : }
3363 :
3364 : // Take the state lock to discover the attached & secondary intents for all shards
3365 0 : let (attached, secondary) = {
3366 0 : let locked = self.inner.read().unwrap();
3367 0 : let mut attached = Vec::new();
3368 0 : let mut secondary = Vec::new();
3369 :
3370 0 : for (tenant_shard_id, shard) in
3371 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
3372 : {
3373 0 : let Some(node_id) = shard.intent.get_attached() else {
3374 : // Unexpected. Race with a PlacementPolicy change?
3375 0 : tracing::warn!(
3376 0 : "No attached node on {tenant_shard_id} immediately after shard split!"
3377 : );
3378 0 : continue;
3379 : };
3380 :
3381 0 : let Some(secondary_node_id) = shard.intent.get_secondary().first() else {
3382 : // No secondary location. Nothing for us to do.
3383 0 : continue;
3384 : };
3385 :
3386 0 : let attached_node = locked
3387 0 : .nodes
3388 0 : .get(node_id)
3389 0 : .expect("Pageservers may not be deleted while referenced");
3390 0 :
3391 0 : let secondary_node = locked
3392 0 : .nodes
3393 0 : .get(secondary_node_id)
3394 0 : .expect("Pageservers may not be deleted while referenced");
3395 0 :
3396 0 : attached.push((*tenant_shard_id, attached_node.clone()));
3397 0 : secondary.push((*tenant_shard_id, secondary_node.clone()));
3398 : }
3399 0 : (attached, secondary)
3400 0 : };
3401 0 :
3402 0 : if secondary.is_empty() {
3403 : // No secondary locations; nothing for us to do
3404 0 : return;
3405 0 : }
3406 :
3407 0 : for result in self
3408 0 : .tenant_for_shards_api(
3409 0 : attached,
3410 0 : |tenant_shard_id, client| async move {
3411 0 : client.tenant_heatmap_upload(tenant_shard_id).await
3412 0 : },
3413 0 : 1,
3414 0 : 1,
3415 0 : SHORT_RECONCILE_TIMEOUT,
3416 0 : &self.cancel,
3417 0 : )
3418 0 : .await
3419 : {
3420 0 : if let Err(e) = result {
3421 0 : tracing::warn!("Error calling heatmap upload after shard split: {e}");
3422 0 : return;
3423 0 : }
3424 : }
3425 :
3426 0 : for result in self
3427 0 : .tenant_for_shards_api(
3428 0 : secondary,
3429 0 : |tenant_shard_id, client| async move {
3430 0 : client
3431 0 : .tenant_secondary_download(tenant_shard_id, Some(Duration::ZERO))
3432 0 : .await
3433 0 : },
3434 0 : 1,
3435 0 : 1,
3436 0 : SHORT_RECONCILE_TIMEOUT,
3437 0 : &self.cancel,
3438 0 : )
3439 0 : .await
3440 : {
3441 0 : if let Err(e) = result {
3442 0 : tracing::warn!("Error calling secondary download after shard split: {e}");
3443 0 : return;
3444 0 : }
3445 : }
3446 0 : }
3447 :
3448 0 : pub(crate) async fn tenant_shard_split(
3449 0 : &self,
3450 0 : tenant_id: TenantId,
3451 0 : split_req: TenantShardSplitRequest,
3452 0 : ) -> Result<TenantShardSplitResponse, ApiError> {
3453 : // TODO: return 503 if we get stuck waiting for this lock
3454 : // (issue https://github.com/neondatabase/neon/issues/7108)
3455 0 : let _tenant_lock = trace_exclusive_lock(
3456 0 : &self.tenant_op_locks,
3457 0 : tenant_id,
3458 0 : TenantOperations::ShardSplit,
3459 0 : )
3460 0 : .await;
3461 :
3462 0 : let new_shard_count = ShardCount::new(split_req.new_shard_count);
3463 0 : let new_stripe_size = split_req.new_stripe_size;
3464 :
3465 : // Validate the request and construct parameters. This phase is fallible, but does not require
3466 : // rollback on errors, as it does no I/O and mutates no state.
3467 0 : let shard_split_params = match self.prepare_tenant_shard_split(tenant_id, split_req)? {
3468 0 : ShardSplitAction::NoOp(resp) => return Ok(resp),
3469 0 : ShardSplitAction::Split(params) => params,
3470 : };
3471 :
3472 : // Execute this split: this phase mutates state and does remote I/O on pageservers. If it fails,
3473 : // we must roll back.
3474 0 : let r = self
3475 0 : .do_tenant_shard_split(tenant_id, shard_split_params)
3476 0 : .await;
3477 :
3478 0 : let (response, waiters) = match r {
3479 0 : Ok(r) => r,
3480 0 : Err(e) => {
3481 0 : // Split might be part-done, we must do work to abort it.
3482 0 : tracing::warn!("Enqueuing background abort of split on {tenant_id}");
3483 0 : self.abort_tx
3484 0 : .send(TenantShardSplitAbort {
3485 0 : tenant_id,
3486 0 : new_shard_count,
3487 0 : new_stripe_size,
3488 0 : _tenant_lock,
3489 0 : })
3490 0 : // Ignore error sending: that just means we're shutting down: aborts are ephemeral so it's fine to drop it.
3491 0 : .ok();
3492 0 : return Err(e);
3493 : }
3494 : };
3495 :
3496 : // The split is now complete. As an optimization, we will trigger all the child shards to upload
3497 : // a heatmap immediately, and all their secondary locations to start downloading: this avoids waiting
3498 : // for the background heatmap/download interval before secondaries get warm enough to migrate shards
3499 : // in [`Self::optimize_all`]
3500 0 : self.tenant_shard_split_start_secondaries(tenant_id, waiters)
3501 0 : .await;
3502 0 : Ok(response)
3503 0 : }
3504 :
3505 0 : fn prepare_tenant_shard_split(
3506 0 : &self,
3507 0 : tenant_id: TenantId,
3508 0 : split_req: TenantShardSplitRequest,
3509 0 : ) -> Result<ShardSplitAction, ApiError> {
3510 0 : fail::fail_point!("shard-split-validation", |_| Err(ApiError::BadRequest(
3511 0 : anyhow::anyhow!("failpoint")
3512 0 : )));
3513 :
3514 0 : let mut policy = None;
3515 0 : let mut config = None;
3516 0 : let mut shard_ident = None;
3517 : // Validate input, and calculate which shards we will create
3518 0 : let (old_shard_count, targets) =
3519 : {
3520 0 : let locked = self.inner.read().unwrap();
3521 0 :
3522 0 : let pageservers = locked.nodes.clone();
3523 0 :
3524 0 : let mut targets = Vec::new();
3525 0 :
3526 0 : // In case this is a retry, count how many already-split shards we found
3527 0 : let mut children_found = Vec::new();
3528 0 : let mut old_shard_count = None;
3529 :
3530 0 : for (tenant_shard_id, shard) in
3531 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
3532 : {
3533 0 : match shard.shard.count.count().cmp(&split_req.new_shard_count) {
3534 : Ordering::Equal => {
3535 : // Already split this
3536 0 : children_found.push(*tenant_shard_id);
3537 0 : continue;
3538 : }
3539 : Ordering::Greater => {
3540 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
3541 0 : "Requested count {} but already have shards at count {}",
3542 0 : split_req.new_shard_count,
3543 0 : shard.shard.count.count()
3544 0 : )));
3545 : }
3546 0 : Ordering::Less => {
3547 0 : // Fall through: this shard has lower count than requested,
3548 0 : // is a candidate for splitting.
3549 0 : }
3550 0 : }
3551 0 :
3552 0 : match old_shard_count {
3553 0 : None => old_shard_count = Some(shard.shard.count),
3554 0 : Some(old_shard_count) => {
3555 0 : if old_shard_count != shard.shard.count {
3556 : // We may hit this case if a caller asked for two splits to
3557 : // different sizes, before the first one is complete.
3558 : // e.g. 1->2, 2->4, where the 4 call comes while we have a mixture
3559 : // of shard_count=1 and shard_count=2 shards in the map.
3560 0 : return Err(ApiError::Conflict(
3561 0 : "Cannot split, currently mid-split".to_string(),
3562 0 : ));
3563 0 : }
3564 : }
3565 : }
3566 0 : if policy.is_none() {
3567 0 : policy = Some(shard.policy.clone());
3568 0 : }
3569 0 : if shard_ident.is_none() {
3570 0 : shard_ident = Some(shard.shard);
3571 0 : }
3572 0 : if config.is_none() {
3573 0 : config = Some(shard.config.clone());
3574 0 : }
3575 :
3576 0 : if tenant_shard_id.shard_count.count() == split_req.new_shard_count {
3577 0 : tracing::info!(
3578 0 : "Tenant shard {} already has shard count {}",
3579 : tenant_shard_id,
3580 : split_req.new_shard_count
3581 : );
3582 0 : continue;
3583 0 : }
3584 :
3585 0 : let node_id = shard.intent.get_attached().ok_or(ApiError::BadRequest(
3586 0 : anyhow::anyhow!("Cannot split a tenant that is not attached"),
3587 0 : ))?;
3588 :
3589 0 : let node = pageservers
3590 0 : .get(&node_id)
3591 0 : .expect("Pageservers may not be deleted while referenced");
3592 0 :
3593 0 : targets.push(ShardSplitTarget {
3594 0 : parent_id: *tenant_shard_id,
3595 0 : node: node.clone(),
3596 0 : child_ids: tenant_shard_id
3597 0 : .split(ShardCount::new(split_req.new_shard_count)),
3598 0 : });
3599 : }
3600 :
3601 0 : if targets.is_empty() {
3602 0 : if children_found.len() == split_req.new_shard_count as usize {
3603 0 : return Ok(ShardSplitAction::NoOp(TenantShardSplitResponse {
3604 0 : new_shards: children_found,
3605 0 : }));
3606 : } else {
3607 : // No shards found to split, and no existing children found: the
3608 : // tenant doesn't exist at all.
3609 0 : return Err(ApiError::NotFound(
3610 0 : anyhow::anyhow!("Tenant {} not found", tenant_id).into(),
3611 0 : ));
3612 : }
3613 0 : }
3614 0 :
3615 0 : (old_shard_count, targets)
3616 0 : };
3617 0 :
3618 0 : // unwrap safety: we would have returned above if we didn't find at least one shard to split
3619 0 : let old_shard_count = old_shard_count.unwrap();
3620 0 : let shard_ident = if let Some(new_stripe_size) = split_req.new_stripe_size {
3621 : // This ShardIdentity will be used as the template for all children, so this implicitly
3622 : // applies the new stripe size to the children.
3623 0 : let mut shard_ident = shard_ident.unwrap();
3624 0 : if shard_ident.count.count() > 1 && shard_ident.stripe_size != new_stripe_size {
3625 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)));
3626 0 : }
3627 0 :
3628 0 : shard_ident.stripe_size = new_stripe_size;
3629 0 : tracing::info!("applied stripe size {}", shard_ident.stripe_size.0);
3630 0 : shard_ident
3631 : } else {
3632 0 : shard_ident.unwrap()
3633 : };
3634 0 : let policy = policy.unwrap();
3635 0 : let config = config.unwrap();
3636 0 :
3637 0 : Ok(ShardSplitAction::Split(ShardSplitParams {
3638 0 : old_shard_count,
3639 0 : new_shard_count: ShardCount::new(split_req.new_shard_count),
3640 0 : new_stripe_size: split_req.new_stripe_size,
3641 0 : targets,
3642 0 : policy,
3643 0 : config,
3644 0 : shard_ident,
3645 0 : }))
3646 0 : }
3647 :
3648 0 : async fn do_tenant_shard_split(
3649 0 : &self,
3650 0 : tenant_id: TenantId,
3651 0 : params: ShardSplitParams,
3652 0 : ) -> Result<(TenantShardSplitResponse, Vec<ReconcilerWaiter>), ApiError> {
3653 0 : // FIXME: we have dropped self.inner lock, and not yet written anything to the database: another
3654 0 : // request could occur here, deleting or mutating the tenant. begin_shard_split checks that the
3655 0 : // parent shards exist as expected, but it would be neater to do the above pre-checks within the
3656 0 : // same database transaction rather than pre-check in-memory and then maybe-fail the database write.
3657 0 : // (https://github.com/neondatabase/neon/issues/6676)
3658 0 :
3659 0 : let ShardSplitParams {
3660 0 : old_shard_count,
3661 0 : new_shard_count,
3662 0 : new_stripe_size,
3663 0 : mut targets,
3664 0 : policy,
3665 0 : config,
3666 0 : shard_ident,
3667 0 : } = params;
3668 :
3669 : // Drop any secondary locations: pageservers do not support splitting these, and in any case the
3670 : // end-state for a split tenant will usually be to have secondary locations on different nodes.
3671 : // The reconciliation calls in this block also implicitly cancel+barrier wrt any ongoing reconciliation
3672 : // at the time of split.
3673 0 : let waiters = {
3674 0 : let mut locked = self.inner.write().unwrap();
3675 0 : let mut waiters = Vec::new();
3676 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
3677 0 : for target in &mut targets {
3678 0 : let Some(shard) = tenants.get_mut(&target.parent_id) else {
3679 : // Paranoia check: this shouldn't happen: we have the oplock for this tenant ID.
3680 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
3681 0 : "Shard {} not found",
3682 0 : target.parent_id
3683 0 : )));
3684 : };
3685 :
3686 0 : if shard.intent.get_attached() != &Some(target.node.get_id()) {
3687 : // Paranoia check: this shouldn't happen: we have the oplock for this tenant ID.
3688 0 : return Err(ApiError::Conflict(format!(
3689 0 : "Shard {} unexpectedly rescheduled during split",
3690 0 : target.parent_id
3691 0 : )));
3692 0 : }
3693 0 :
3694 0 : // Irrespective of PlacementPolicy, clear secondary locations from intent
3695 0 : shard.intent.clear_secondary(scheduler);
3696 :
3697 : // Run Reconciler to execute detach fo secondary locations.
3698 0 : if let Some(waiter) = self.maybe_reconcile_shard(shard, nodes) {
3699 0 : waiters.push(waiter);
3700 0 : }
3701 : }
3702 0 : waiters
3703 0 : };
3704 0 : self.await_waiters(waiters, RECONCILE_TIMEOUT).await?;
3705 :
3706 : // Before creating any new child shards in memory or on the pageservers, persist them: this
3707 : // enables us to ensure that we will always be able to clean up if something goes wrong. This also
3708 : // acts as the protection against two concurrent attempts to split: one of them will get a database
3709 : // error trying to insert the child shards.
3710 0 : let mut child_tsps = Vec::new();
3711 0 : for target in &targets {
3712 0 : let mut this_child_tsps = Vec::new();
3713 0 : for child in &target.child_ids {
3714 0 : let mut child_shard = shard_ident;
3715 0 : child_shard.number = child.shard_number;
3716 0 : child_shard.count = child.shard_count;
3717 0 :
3718 0 : tracing::info!(
3719 0 : "Create child shard persistence with stripe size {}",
3720 : shard_ident.stripe_size.0
3721 : );
3722 :
3723 0 : this_child_tsps.push(TenantShardPersistence {
3724 0 : tenant_id: child.tenant_id.to_string(),
3725 0 : shard_number: child.shard_number.0 as i32,
3726 0 : shard_count: child.shard_count.literal() as i32,
3727 0 : shard_stripe_size: shard_ident.stripe_size.0 as i32,
3728 0 : // Note: this generation is a placeholder, [`Persistence::begin_shard_split`] will
3729 0 : // populate the correct generation as part of its transaction, to protect us
3730 0 : // against racing with changes in the state of the parent.
3731 0 : generation: None,
3732 0 : generation_pageserver: Some(target.node.get_id().0 as i64),
3733 0 : placement_policy: serde_json::to_string(&policy).unwrap(),
3734 0 : config: serde_json::to_string(&config).unwrap(),
3735 0 : splitting: SplitState::Splitting,
3736 0 :
3737 0 : // Scheduling policies do not carry through to children
3738 0 : scheduling_policy: serde_json::to_string(&ShardSchedulingPolicy::default())
3739 0 : .unwrap(),
3740 0 : });
3741 : }
3742 :
3743 0 : child_tsps.push((target.parent_id, this_child_tsps));
3744 : }
3745 :
3746 0 : if let Err(e) = self
3747 0 : .persistence
3748 0 : .begin_shard_split(old_shard_count, tenant_id, child_tsps)
3749 0 : .await
3750 : {
3751 0 : match e {
3752 : DatabaseError::Query(diesel::result::Error::DatabaseError(
3753 : DatabaseErrorKind::UniqueViolation,
3754 : _,
3755 : )) => {
3756 : // Inserting a child shard violated a unique constraint: we raced with another call to
3757 : // this function
3758 0 : tracing::warn!("Conflicting attempt to split {tenant_id}: {e}");
3759 0 : return Err(ApiError::Conflict("Tenant is already splitting".into()));
3760 : }
3761 0 : _ => return Err(ApiError::InternalServerError(e.into())),
3762 : }
3763 0 : }
3764 0 : fail::fail_point!("shard-split-post-begin", |_| Err(
3765 0 : ApiError::InternalServerError(anyhow::anyhow!("failpoint"))
3766 0 : ));
3767 :
3768 : // Now that I have persisted the splitting state, apply it in-memory. This is infallible, so
3769 : // callers may assume that if splitting is set in memory, then it was persisted, and if splitting
3770 : // is not set in memory, then it was not persisted.
3771 : {
3772 0 : let mut locked = self.inner.write().unwrap();
3773 0 : for target in &targets {
3774 0 : if let Some(parent_shard) = locked.tenants.get_mut(&target.parent_id) {
3775 0 : parent_shard.splitting = SplitState::Splitting;
3776 0 : // Put the observed state to None, to reflect that it is indeterminate once we start the
3777 0 : // split operation.
3778 0 : parent_shard
3779 0 : .observed
3780 0 : .locations
3781 0 : .insert(target.node.get_id(), ObservedStateLocation { conf: None });
3782 0 : }
3783 : }
3784 : }
3785 :
3786 : // TODO: issue split calls concurrently (this only matters once we're splitting
3787 : // N>1 shards into M shards -- initially we're usually splitting 1 shard into N).
3788 :
3789 0 : for target in &targets {
3790 : let ShardSplitTarget {
3791 0 : parent_id,
3792 0 : node,
3793 0 : child_ids,
3794 0 : } = target;
3795 0 : let client = PageserverClient::new(
3796 0 : node.get_id(),
3797 0 : node.base_url(),
3798 0 : self.config.jwt_token.as_deref(),
3799 0 : );
3800 0 : let response = client
3801 0 : .tenant_shard_split(
3802 0 : *parent_id,
3803 0 : TenantShardSplitRequest {
3804 0 : new_shard_count: new_shard_count.literal(),
3805 0 : new_stripe_size,
3806 0 : },
3807 0 : )
3808 0 : .await
3809 0 : .map_err(|e| ApiError::Conflict(format!("Failed to split {}: {}", parent_id, e)))?;
3810 :
3811 0 : fail::fail_point!("shard-split-post-remote", |_| Err(ApiError::Conflict(
3812 0 : "failpoint".to_string()
3813 0 : )));
3814 :
3815 0 : tracing::info!(
3816 0 : "Split {} into {}",
3817 0 : parent_id,
3818 0 : response
3819 0 : .new_shards
3820 0 : .iter()
3821 0 : .map(|s| format!("{:?}", s))
3822 0 : .collect::<Vec<_>>()
3823 0 : .join(",")
3824 : );
3825 :
3826 0 : if &response.new_shards != child_ids {
3827 : // This should never happen: the pageserver should agree with us on how shard splits work.
3828 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
3829 0 : "Splitting shard {} resulted in unexpected IDs: {:?} (expected {:?})",
3830 0 : parent_id,
3831 0 : response.new_shards,
3832 0 : child_ids
3833 0 : )));
3834 0 : }
3835 : }
3836 :
3837 : // TODO: if the pageserver restarted concurrently with our split API call,
3838 : // the actual generation of the child shard might differ from the generation
3839 : // we expect it to have. In order for our in-database generation to end up
3840 : // correct, we should carry the child generation back in the response and apply it here
3841 : // in complete_shard_split (and apply the correct generation in memory)
3842 : // (or, we can carry generation in the request and reject the request if
3843 : // it doesn't match, but that requires more retry logic on this side)
3844 :
3845 0 : self.persistence
3846 0 : .complete_shard_split(tenant_id, old_shard_count)
3847 0 : .await?;
3848 :
3849 0 : fail::fail_point!("shard-split-post-complete", |_| Err(
3850 0 : ApiError::InternalServerError(anyhow::anyhow!("failpoint"))
3851 0 : ));
3852 :
3853 : // Replace all the shards we just split with their children: this phase is infallible.
3854 0 : let (response, child_locations, waiters) =
3855 0 : self.tenant_shard_split_commit_inmem(tenant_id, new_shard_count, new_stripe_size);
3856 0 :
3857 0 : // Send compute notifications for all the new shards
3858 0 : let mut failed_notifications = Vec::new();
3859 0 : for (child_id, child_ps, stripe_size) in child_locations {
3860 0 : if let Err(e) = self
3861 0 : .compute_hook
3862 0 : .notify(child_id, child_ps, stripe_size, &self.cancel)
3863 0 : .await
3864 : {
3865 0 : tracing::warn!("Failed to update compute of {}->{} during split, proceeding anyway to complete split ({e})",
3866 : child_id, child_ps);
3867 0 : failed_notifications.push(child_id);
3868 0 : }
3869 : }
3870 :
3871 : // If we failed any compute notifications, make a note to retry later.
3872 0 : if !failed_notifications.is_empty() {
3873 0 : let mut locked = self.inner.write().unwrap();
3874 0 : for failed in failed_notifications {
3875 0 : if let Some(shard) = locked.tenants.get_mut(&failed) {
3876 0 : shard.pending_compute_notification = true;
3877 0 : }
3878 : }
3879 0 : }
3880 :
3881 0 : Ok((response, waiters))
3882 0 : }
3883 :
3884 0 : pub(crate) async fn tenant_shard_migrate(
3885 0 : &self,
3886 0 : tenant_shard_id: TenantShardId,
3887 0 : migrate_req: TenantShardMigrateRequest,
3888 0 : ) -> Result<TenantShardMigrateResponse, ApiError> {
3889 0 : let waiter = {
3890 0 : let mut locked = self.inner.write().unwrap();
3891 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
3892 :
3893 0 : let Some(node) = nodes.get(&migrate_req.node_id) else {
3894 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
3895 0 : "Node {} not found",
3896 0 : migrate_req.node_id
3897 0 : )));
3898 : };
3899 :
3900 0 : if !node.is_available() {
3901 : // Warn but proceed: the caller may intend to manually adjust the placement of
3902 : // a shard even if the node is down, e.g. if intervening during an incident.
3903 0 : tracing::warn!("Migrating to unavailable node {node}");
3904 0 : }
3905 :
3906 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
3907 0 : return Err(ApiError::NotFound(
3908 0 : anyhow::anyhow!("Tenant shard not found").into(),
3909 0 : ));
3910 : };
3911 :
3912 0 : if shard.intent.get_attached() == &Some(migrate_req.node_id) {
3913 : // No-op case: we will still proceed to wait for reconciliation in case it is
3914 : // incomplete from an earlier update to the intent.
3915 0 : tracing::info!("Migrating: intent is unchanged {:?}", shard.intent);
3916 : } else {
3917 0 : let old_attached = *shard.intent.get_attached();
3918 0 :
3919 0 : match shard.policy {
3920 0 : PlacementPolicy::Attached(n) => {
3921 0 : // If our new attached node was a secondary, it no longer should be.
3922 0 : shard.intent.remove_secondary(scheduler, migrate_req.node_id);
3923 :
3924 : // If we were already attached to something, demote that to a secondary
3925 0 : if let Some(old_attached) = old_attached {
3926 0 : if n > 0 {
3927 : // Remove other secondaries to make room for the location we'll demote
3928 0 : while shard.intent.get_secondary().len() >= n {
3929 0 : shard.intent.pop_secondary(scheduler);
3930 0 : }
3931 :
3932 0 : shard.intent.push_secondary(scheduler, old_attached);
3933 0 : }
3934 0 : }
3935 :
3936 0 : shard.intent.set_attached(scheduler, Some(migrate_req.node_id));
3937 : }
3938 0 : PlacementPolicy::Secondary => {
3939 0 : shard.intent.clear(scheduler);
3940 0 : shard.intent.push_secondary(scheduler, migrate_req.node_id);
3941 0 : }
3942 : PlacementPolicy::Detached => {
3943 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
3944 0 : "Cannot migrate a tenant that is PlacementPolicy::Detached: configure it to an attached policy first"
3945 0 : )))
3946 : }
3947 : }
3948 :
3949 0 : tracing::info!("Migrating: new intent {:?}", shard.intent);
3950 0 : shard.sequence = shard.sequence.next();
3951 : }
3952 :
3953 0 : self.maybe_reconcile_shard(shard, nodes)
3954 : };
3955 :
3956 0 : if let Some(waiter) = waiter {
3957 0 : waiter.wait_timeout(RECONCILE_TIMEOUT).await?;
3958 : } else {
3959 0 : tracing::info!("Migration is a no-op");
3960 : }
3961 :
3962 0 : Ok(TenantShardMigrateResponse {})
3963 0 : }
3964 :
3965 : /// This is for debug/support only: we simply drop all state for a tenant, without
3966 : /// detaching or deleting it on pageservers.
3967 0 : pub(crate) async fn tenant_drop(&self, tenant_id: TenantId) -> Result<(), ApiError> {
3968 0 : self.persistence.delete_tenant(tenant_id).await?;
3969 :
3970 0 : let mut locked = self.inner.write().unwrap();
3971 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
3972 0 : let mut shards = Vec::new();
3973 0 : for (tenant_shard_id, _) in tenants.range(TenantShardId::tenant_range(tenant_id)) {
3974 0 : shards.push(*tenant_shard_id);
3975 0 : }
3976 :
3977 0 : for shard_id in shards {
3978 0 : if let Some(mut shard) = tenants.remove(&shard_id) {
3979 0 : shard.intent.clear(scheduler);
3980 0 : }
3981 : }
3982 :
3983 0 : Ok(())
3984 0 : }
3985 :
3986 : /// This is for debug/support only: assuming tenant data is already present in S3, we "create" a
3987 : /// tenant with a very high generation number so that it will see the existing data.
3988 0 : pub(crate) async fn tenant_import(
3989 0 : &self,
3990 0 : tenant_id: TenantId,
3991 0 : ) -> Result<TenantCreateResponse, ApiError> {
3992 0 : // Pick an arbitrary available pageserver to use for scanning the tenant in remote storage
3993 0 : let maybe_node = {
3994 0 : self.inner
3995 0 : .read()
3996 0 : .unwrap()
3997 0 : .nodes
3998 0 : .values()
3999 0 : .find(|n| n.is_available())
4000 0 : .cloned()
4001 : };
4002 0 : let Some(node) = maybe_node else {
4003 0 : return Err(ApiError::BadRequest(anyhow::anyhow!("No nodes available")));
4004 : };
4005 :
4006 0 : let client = PageserverClient::new(
4007 0 : node.get_id(),
4008 0 : node.base_url(),
4009 0 : self.config.jwt_token.as_deref(),
4010 0 : );
4011 :
4012 0 : let scan_result = client
4013 0 : .tenant_scan_remote_storage(tenant_id)
4014 0 : .await
4015 0 : .map_err(|e| passthrough_api_error(&node, e))?;
4016 :
4017 : // A post-split tenant may contain a mixture of shard counts in remote storage: pick the highest count.
4018 0 : let Some(shard_count) = scan_result
4019 0 : .shards
4020 0 : .iter()
4021 0 : .map(|s| s.tenant_shard_id.shard_count)
4022 0 : .max()
4023 : else {
4024 0 : return Err(ApiError::NotFound(
4025 0 : anyhow::anyhow!("No shards found").into(),
4026 0 : ));
4027 : };
4028 :
4029 : // Ideally we would set each newly imported shard's generation independently, but for correctness it is sufficient
4030 : // to
4031 0 : let generation = scan_result
4032 0 : .shards
4033 0 : .iter()
4034 0 : .map(|s| s.generation)
4035 0 : .max()
4036 0 : .expect("We already validated >0 shards");
4037 0 :
4038 0 : // FIXME: we have no way to recover the shard stripe size from contents of remote storage: this will
4039 0 : // only work if they were using the default stripe size.
4040 0 : let stripe_size = ShardParameters::DEFAULT_STRIPE_SIZE;
4041 :
4042 0 : let (response, waiters) = self
4043 0 : .do_tenant_create(TenantCreateRequest {
4044 0 : new_tenant_id: TenantShardId::unsharded(tenant_id),
4045 0 : generation,
4046 0 :
4047 0 : shard_parameters: ShardParameters {
4048 0 : count: shard_count,
4049 0 : stripe_size,
4050 0 : },
4051 0 : placement_policy: Some(PlacementPolicy::Attached(0)), // No secondaries, for convenient debug/hacking
4052 0 :
4053 0 : // There is no way to know what the tenant's config was: revert to defaults
4054 0 : config: TenantConfig::default(),
4055 0 : })
4056 0 : .await?;
4057 :
4058 0 : if let Err(e) = self.await_waiters(waiters, SHORT_RECONCILE_TIMEOUT).await {
4059 : // Since this is a debug/support operation, all kinds of weird issues are possible (e.g. this
4060 : // tenant doesn't exist in the control plane), so don't fail the request if it can't fully
4061 : // reconcile, as reconciliation includes notifying compute.
4062 0 : tracing::warn!(%tenant_id, "Reconcile not done yet while importing tenant ({e})");
4063 0 : }
4064 :
4065 0 : Ok(response)
4066 0 : }
4067 :
4068 : /// For debug/support: a full JSON dump of TenantShards. Returns a response so that
4069 : /// we don't have to make TenantShard clonable in the return path.
4070 0 : pub(crate) fn tenants_dump(&self) -> Result<hyper::Response<hyper::Body>, ApiError> {
4071 0 : let serialized = {
4072 0 : let locked = self.inner.read().unwrap();
4073 0 : let result = locked.tenants.values().collect::<Vec<_>>();
4074 0 : serde_json::to_string(&result).map_err(|e| ApiError::InternalServerError(e.into()))?
4075 : };
4076 :
4077 0 : hyper::Response::builder()
4078 0 : .status(hyper::StatusCode::OK)
4079 0 : .header(hyper::header::CONTENT_TYPE, "application/json")
4080 0 : .body(hyper::Body::from(serialized))
4081 0 : .map_err(|e| ApiError::InternalServerError(e.into()))
4082 0 : }
4083 :
4084 : /// Check the consistency of in-memory state vs. persistent state, and check that the
4085 : /// scheduler's statistics are up to date.
4086 : ///
4087 : /// These consistency checks expect an **idle** system. If changes are going on while
4088 : /// we run, then we can falsely indicate a consistency issue. This is sufficient for end-of-test
4089 : /// checks, but not suitable for running continuously in the background in the field.
4090 0 : pub(crate) async fn consistency_check(&self) -> Result<(), ApiError> {
4091 0 : let (mut expect_nodes, mut expect_shards) = {
4092 0 : let locked = self.inner.read().unwrap();
4093 0 :
4094 0 : locked
4095 0 : .scheduler
4096 0 : .consistency_check(locked.nodes.values(), locked.tenants.values())
4097 0 : .context("Scheduler checks")
4098 0 : .map_err(ApiError::InternalServerError)?;
4099 :
4100 0 : let expect_nodes = locked
4101 0 : .nodes
4102 0 : .values()
4103 0 : .map(|n| n.to_persistent())
4104 0 : .collect::<Vec<_>>();
4105 0 :
4106 0 : let expect_shards = locked
4107 0 : .tenants
4108 0 : .values()
4109 0 : .map(|t| t.to_persistent())
4110 0 : .collect::<Vec<_>>();
4111 :
4112 : // This method can only validate the state of an idle system: if a reconcile is in
4113 : // progress, fail out early to avoid giving false errors on state that won't match
4114 : // between database and memory under a ReconcileResult is processed.
4115 0 : for t in locked.tenants.values() {
4116 0 : if t.reconciler.is_some() {
4117 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
4118 0 : "Shard {} reconciliation in progress",
4119 0 : t.tenant_shard_id
4120 0 : )));
4121 0 : }
4122 : }
4123 :
4124 0 : (expect_nodes, expect_shards)
4125 : };
4126 :
4127 0 : let mut nodes = self.persistence.list_nodes().await?;
4128 0 : expect_nodes.sort_by_key(|n| n.node_id);
4129 0 : nodes.sort_by_key(|n| n.node_id);
4130 0 :
4131 0 : if nodes != expect_nodes {
4132 0 : tracing::error!("Consistency check failed on nodes.");
4133 0 : tracing::error!(
4134 0 : "Nodes in memory: {}",
4135 0 : serde_json::to_string(&expect_nodes)
4136 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
4137 : );
4138 0 : tracing::error!(
4139 0 : "Nodes in database: {}",
4140 0 : serde_json::to_string(&nodes)
4141 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
4142 : );
4143 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
4144 0 : "Node consistency failure"
4145 0 : )));
4146 0 : }
4147 :
4148 0 : let mut shards = self.persistence.list_tenant_shards().await?;
4149 0 : shards.sort_by_key(|tsp| (tsp.tenant_id.clone(), tsp.shard_number, tsp.shard_count));
4150 0 : expect_shards.sort_by_key(|tsp| (tsp.tenant_id.clone(), tsp.shard_number, tsp.shard_count));
4151 0 :
4152 0 : if shards != expect_shards {
4153 0 : tracing::error!("Consistency check failed on shards.");
4154 0 : tracing::error!(
4155 0 : "Shards in memory: {}",
4156 0 : serde_json::to_string(&expect_shards)
4157 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
4158 : );
4159 0 : tracing::error!(
4160 0 : "Shards in database: {}",
4161 0 : serde_json::to_string(&shards)
4162 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
4163 : );
4164 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
4165 0 : "Shard consistency failure"
4166 0 : )));
4167 0 : }
4168 0 :
4169 0 : Ok(())
4170 0 : }
4171 :
4172 : /// For debug/support: a JSON dump of the [`Scheduler`]. Returns a response so that
4173 : /// we don't have to make TenantShard clonable in the return path.
4174 0 : pub(crate) fn scheduler_dump(&self) -> Result<hyper::Response<hyper::Body>, ApiError> {
4175 0 : let serialized = {
4176 0 : let locked = self.inner.read().unwrap();
4177 0 : serde_json::to_string(&locked.scheduler)
4178 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
4179 : };
4180 :
4181 0 : hyper::Response::builder()
4182 0 : .status(hyper::StatusCode::OK)
4183 0 : .header(hyper::header::CONTENT_TYPE, "application/json")
4184 0 : .body(hyper::Body::from(serialized))
4185 0 : .map_err(|e| ApiError::InternalServerError(e.into()))
4186 0 : }
4187 :
4188 : /// This is for debug/support only: we simply drop all state for a tenant, without
4189 : /// detaching or deleting it on pageservers. We do not try and re-schedule any
4190 : /// tenants that were on this node.
4191 : ///
4192 : /// TODO: proper node deletion API that unhooks things more gracefully
4193 0 : pub(crate) async fn node_drop(&self, node_id: NodeId) -> Result<(), ApiError> {
4194 0 : self.persistence.delete_node(node_id).await?;
4195 :
4196 0 : let mut locked = self.inner.write().unwrap();
4197 :
4198 0 : for shard in locked.tenants.values_mut() {
4199 0 : shard.deref_node(node_id);
4200 0 : }
4201 :
4202 0 : let mut nodes = (*locked.nodes).clone();
4203 0 : nodes.remove(&node_id);
4204 0 : locked.nodes = Arc::new(nodes);
4205 0 :
4206 0 : locked.scheduler.node_remove(node_id);
4207 0 :
4208 0 : Ok(())
4209 0 : }
4210 :
4211 0 : pub(crate) async fn node_list(&self) -> Result<Vec<Node>, ApiError> {
4212 0 : let nodes = {
4213 0 : self.inner
4214 0 : .read()
4215 0 : .unwrap()
4216 0 : .nodes
4217 0 : .values()
4218 0 : .cloned()
4219 0 : .collect::<Vec<_>>()
4220 0 : };
4221 0 :
4222 0 : Ok(nodes)
4223 0 : }
4224 :
4225 0 : pub(crate) async fn get_node(&self, node_id: NodeId) -> Result<Node, ApiError> {
4226 0 : self.inner
4227 0 : .read()
4228 0 : .unwrap()
4229 0 : .nodes
4230 0 : .get(&node_id)
4231 0 : .cloned()
4232 0 : .ok_or(ApiError::NotFound(
4233 0 : format!("Node {node_id} not registered").into(),
4234 0 : ))
4235 0 : }
4236 :
4237 0 : pub(crate) async fn node_register(
4238 0 : &self,
4239 0 : register_req: NodeRegisterRequest,
4240 0 : ) -> Result<(), ApiError> {
4241 0 : let _node_lock = trace_exclusive_lock(
4242 0 : &self.node_op_locks,
4243 0 : register_req.node_id,
4244 0 : NodeOperations::Register,
4245 0 : )
4246 0 : .await;
4247 :
4248 : {
4249 0 : let locked = self.inner.read().unwrap();
4250 0 : if let Some(node) = locked.nodes.get(®ister_req.node_id) {
4251 : // Note that we do not do a total equality of the struct, because we don't require
4252 : // the availability/scheduling states to agree for a POST to be idempotent.
4253 0 : if node.registration_match(®ister_req) {
4254 0 : tracing::info!(
4255 0 : "Node {} re-registered with matching address",
4256 : register_req.node_id
4257 : );
4258 0 : return Ok(());
4259 : } else {
4260 : // TODO: decide if we want to allow modifying node addresses without removing and re-adding
4261 : // the node. Safest/simplest thing is to refuse it, and usually we deploy with
4262 : // a fixed address through the lifetime of a node.
4263 0 : tracing::warn!(
4264 0 : "Node {} tried to register with different address",
4265 : register_req.node_id
4266 : );
4267 0 : return Err(ApiError::Conflict(
4268 0 : "Node is already registered with different address".to_string(),
4269 0 : ));
4270 : }
4271 0 : }
4272 0 : }
4273 0 :
4274 0 : // We do not require that a node is actually online when registered (it will start life
4275 0 : // with it's availability set to Offline), but we _do_ require that its DNS record exists. We're
4276 0 : // therefore not immune to asymmetric L3 connectivity issues, but we are protected against nodes
4277 0 : // that register themselves with a broken DNS config. We check only the HTTP hostname, because
4278 0 : // the postgres hostname might only be resolvable to clients (e.g. if we're on a different VPC than clients).
4279 0 : if tokio::net::lookup_host(format!(
4280 0 : "{}:{}",
4281 0 : register_req.listen_http_addr, register_req.listen_http_port
4282 0 : ))
4283 0 : .await
4284 0 : .is_err()
4285 : {
4286 : // If we have a transient DNS issue, it's up to the caller to retry their registration. Because
4287 : // we can't robustly distinguish between an intermittent issue and a totally bogus DNS situation,
4288 : // we return a soft 503 error, to encourage callers to retry past transient issues.
4289 0 : return Err(ApiError::ResourceUnavailable(
4290 0 : format!(
4291 0 : "Node {} tried to register with unknown DNS name '{}'",
4292 0 : register_req.node_id, register_req.listen_http_addr
4293 0 : )
4294 0 : .into(),
4295 0 : ));
4296 0 : }
4297 0 :
4298 0 : // Ordering: we must persist the new node _before_ adding it to in-memory state.
4299 0 : // This ensures that before we use it for anything or expose it via any external
4300 0 : // API, it is guaranteed to be available after a restart.
4301 0 : let new_node = Node::new(
4302 0 : register_req.node_id,
4303 0 : register_req.listen_http_addr,
4304 0 : register_req.listen_http_port,
4305 0 : register_req.listen_pg_addr,
4306 0 : register_req.listen_pg_port,
4307 0 : );
4308 0 :
4309 0 : // TODO: idempotency if the node already exists in the database
4310 0 : self.persistence.insert_node(&new_node).await?;
4311 :
4312 0 : let mut locked = self.inner.write().unwrap();
4313 0 : let mut new_nodes = (*locked.nodes).clone();
4314 0 :
4315 0 : locked.scheduler.node_upsert(&new_node);
4316 0 : new_nodes.insert(register_req.node_id, new_node);
4317 0 :
4318 0 : locked.nodes = Arc::new(new_nodes);
4319 0 :
4320 0 : tracing::info!(
4321 0 : "Registered pageserver {}, now have {} pageservers",
4322 0 : register_req.node_id,
4323 0 : locked.nodes.len()
4324 : );
4325 0 : Ok(())
4326 0 : }
4327 :
4328 0 : pub(crate) async fn node_configure(
4329 0 : &self,
4330 0 : node_id: NodeId,
4331 0 : availability: Option<NodeAvailability>,
4332 0 : scheduling: Option<NodeSchedulingPolicy>,
4333 0 : ) -> Result<(), ApiError> {
4334 0 : let _node_lock =
4335 0 : trace_exclusive_lock(&self.node_op_locks, node_id, NodeOperations::Configure).await;
4336 :
4337 0 : if let Some(scheduling) = scheduling {
4338 : // Scheduling is a persistent part of Node: we must write updates to the database before
4339 : // applying them in memory
4340 0 : self.persistence.update_node(node_id, scheduling).await?;
4341 0 : }
4342 :
4343 : // If we're activating a node, then before setting it active we must reconcile any shard locations
4344 : // on that node, in case it is out of sync, e.g. due to being unavailable during controller startup,
4345 : // by calling [`Self::node_activate_reconcile`]
4346 : //
4347 : // The transition we calculate here remains valid later in the function because we hold the op lock on the node:
4348 : // nothing else can mutate its availability while we run.
4349 0 : let availability_transition = if let Some(input_availability) = availability {
4350 0 : let (activate_node, availability_transition) = {
4351 0 : let locked = self.inner.read().unwrap();
4352 0 : let Some(node) = locked.nodes.get(&node_id) else {
4353 0 : return Err(ApiError::NotFound(
4354 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
4355 0 : ));
4356 : };
4357 :
4358 0 : (
4359 0 : node.clone(),
4360 0 : node.get_availability_transition(input_availability),
4361 0 : )
4362 : };
4363 :
4364 0 : if matches!(availability_transition, AvailabilityTransition::ToActive) {
4365 0 : self.node_activate_reconcile(activate_node, &_node_lock)
4366 0 : .await?;
4367 0 : }
4368 0 : availability_transition
4369 : } else {
4370 0 : AvailabilityTransition::Unchanged
4371 : };
4372 :
4373 : // Apply changes from the request to our in-memory state for the Node
4374 0 : let mut locked = self.inner.write().unwrap();
4375 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
4376 0 :
4377 0 : let mut new_nodes = (**nodes).clone();
4378 :
4379 0 : let Some(node) = new_nodes.get_mut(&node_id) else {
4380 0 : return Err(ApiError::NotFound(
4381 0 : anyhow::anyhow!("Node not registered").into(),
4382 0 : ));
4383 : };
4384 :
4385 0 : if let Some(availability) = &availability {
4386 0 : node.set_availability(*availability);
4387 0 : }
4388 :
4389 0 : if let Some(scheduling) = scheduling {
4390 0 : node.set_scheduling(scheduling);
4391 0 : }
4392 :
4393 : // Update the scheduler, in case the elegibility of the node for new shards has changed
4394 0 : scheduler.node_upsert(node);
4395 0 :
4396 0 : let new_nodes = Arc::new(new_nodes);
4397 0 :
4398 0 : // Modify scheduling state for any Tenants that are affected by a change in the node's availability state.
4399 0 : match availability_transition {
4400 : AvailabilityTransition::ToOffline => {
4401 0 : tracing::info!("Node {} transition to offline", node_id);
4402 0 : let mut tenants_affected: usize = 0;
4403 :
4404 0 : for (tenant_shard_id, tenant_shard) in tenants {
4405 0 : if let Some(observed_loc) = tenant_shard.observed.locations.get_mut(&node_id) {
4406 0 : // When a node goes offline, we set its observed configuration to None, indicating unknown: we will
4407 0 : // not assume our knowledge of the node's configuration is accurate until it comes back online
4408 0 : observed_loc.conf = None;
4409 0 : }
4410 :
4411 0 : if new_nodes.len() == 1 {
4412 : // Special case for single-node cluster: there is no point trying to reschedule
4413 : // any tenant shards: avoid doing so, in order to avoid spewing warnings about
4414 : // failures to schedule them.
4415 0 : continue;
4416 0 : }
4417 0 :
4418 0 : if !new_nodes
4419 0 : .values()
4420 0 : .any(|n| matches!(n.may_schedule(), MaySchedule::Yes(_)))
4421 : {
4422 : // Special case for when all nodes are unavailable and/or unschedulable: there is no point
4423 : // trying to reschedule since there's nowhere else to go. Without this
4424 : // branch we incorrectly detach tenants in response to node unavailability.
4425 0 : continue;
4426 0 : }
4427 0 :
4428 0 : if tenant_shard.intent.demote_attached(scheduler, node_id) {
4429 0 : tenant_shard.sequence = tenant_shard.sequence.next();
4430 0 :
4431 0 : // TODO: populate a ScheduleContext including all shards in the same tenant_id (only matters
4432 0 : // for tenants without secondary locations: if they have a secondary location, then this
4433 0 : // schedule() call is just promoting an existing secondary)
4434 0 : let mut schedule_context = ScheduleContext::default();
4435 0 :
4436 0 : match tenant_shard.schedule(scheduler, &mut schedule_context) {
4437 0 : Err(e) => {
4438 0 : // It is possible that some tenants will become unschedulable when too many pageservers
4439 0 : // go offline: in this case there isn't much we can do other than make the issue observable.
4440 0 : // TODO: give TenantShard a scheduling error attribute to be queried later.
4441 0 : tracing::warn!(%tenant_shard_id, "Scheduling error when marking pageserver {} offline: {e}", node_id);
4442 : }
4443 : Ok(()) => {
4444 0 : if self
4445 0 : .maybe_reconcile_shard(tenant_shard, &new_nodes)
4446 0 : .is_some()
4447 0 : {
4448 0 : tenants_affected += 1;
4449 0 : };
4450 : }
4451 : }
4452 0 : }
4453 : }
4454 0 : tracing::info!(
4455 0 : "Launched {} reconciler tasks for tenants affected by node {} going offline",
4456 : tenants_affected,
4457 : node_id
4458 : )
4459 : }
4460 : AvailabilityTransition::ToActive => {
4461 0 : tracing::info!("Node {} transition to active", node_id);
4462 : // When a node comes back online, we must reconcile any tenant that has a None observed
4463 : // location on the node.
4464 0 : for tenant_shard in locked.tenants.values_mut() {
4465 : // If a reconciliation is already in progress, rely on the previous scheduling
4466 : // decision and skip triggering a new reconciliation.
4467 0 : if tenant_shard.reconciler.is_some() {
4468 0 : continue;
4469 0 : }
4470 :
4471 0 : if let Some(observed_loc) = tenant_shard.observed.locations.get_mut(&node_id) {
4472 0 : if observed_loc.conf.is_none() {
4473 0 : self.maybe_reconcile_shard(tenant_shard, &new_nodes);
4474 0 : }
4475 0 : }
4476 : }
4477 :
4478 : // TODO: in the background, we should balance work back onto this pageserver
4479 : }
4480 : AvailabilityTransition::Unchanged => {
4481 0 : tracing::debug!("Node {} no availability change during config", node_id);
4482 : }
4483 : }
4484 :
4485 0 : locked.nodes = new_nodes;
4486 0 :
4487 0 : Ok(())
4488 0 : }
4489 :
4490 0 : pub(crate) async fn start_node_drain(
4491 0 : self: &Arc<Self>,
4492 0 : node_id: NodeId,
4493 0 : ) -> Result<(), ApiError> {
4494 0 : let (ongoing_op, node_available, node_policy, schedulable_nodes_count) = {
4495 0 : let locked = self.inner.read().unwrap();
4496 0 : let nodes = &locked.nodes;
4497 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
4498 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
4499 0 : ))?;
4500 0 : let schedulable_nodes_count = nodes
4501 0 : .iter()
4502 0 : .filter(|(_, n)| matches!(n.may_schedule(), MaySchedule::Yes(_)))
4503 0 : .count();
4504 0 :
4505 0 : (
4506 0 : locked
4507 0 : .ongoing_operation
4508 0 : .as_ref()
4509 0 : .map(|ongoing| ongoing.operation),
4510 0 : node.is_available(),
4511 0 : node.get_scheduling(),
4512 0 : schedulable_nodes_count,
4513 0 : )
4514 : };
4515 :
4516 0 : if let Some(ongoing) = ongoing_op {
4517 0 : return Err(ApiError::PreconditionFailed(
4518 0 : format!("Background operation already ongoing for node: {}", ongoing).into(),
4519 0 : ));
4520 0 : }
4521 0 :
4522 0 : if !node_available {
4523 0 : return Err(ApiError::ResourceUnavailable(
4524 0 : format!("Node {node_id} is currently unavailable").into(),
4525 0 : ));
4526 0 : }
4527 0 :
4528 0 : if schedulable_nodes_count == 0 {
4529 0 : return Err(ApiError::PreconditionFailed(
4530 0 : "No other schedulable nodes to drain to".into(),
4531 0 : ));
4532 0 : }
4533 0 :
4534 0 : match node_policy {
4535 : NodeSchedulingPolicy::Active | NodeSchedulingPolicy::Pause => {
4536 0 : self.node_configure(node_id, None, Some(NodeSchedulingPolicy::Draining))
4537 0 : .await?;
4538 :
4539 0 : let cancel = CancellationToken::new();
4540 0 :
4541 0 : self.inner.write().unwrap().ongoing_operation = Some(OperationHandler {
4542 0 : operation: Operation::Drain(Drain { node_id }),
4543 0 : cancel: cancel.clone(),
4544 0 : });
4545 0 :
4546 0 : tokio::task::spawn({
4547 0 : let service = self.clone();
4548 0 : let cancel = cancel.clone();
4549 0 : async move {
4550 : scopeguard::defer! {
4551 : let prev = service.inner.write().unwrap().ongoing_operation.take();
4552 :
4553 0 : if let Some(Operation::Drain(removed_drain)) = prev.map(|h| h.operation) {
4554 : assert_eq!(removed_drain.node_id, node_id, "We always take the same operation");
4555 : } else {
4556 : panic!("We always remove the same operation")
4557 : }
4558 : }
4559 :
4560 0 : tracing::info!(%node_id, "Drain background operation starting");
4561 0 : let res = service.drain_node(node_id, cancel).await;
4562 0 : match res {
4563 : Ok(()) => {
4564 0 : tracing::info!(%node_id, "Drain background operation completed successfully");
4565 : }
4566 : Err(OperationError::Cancelled) => {
4567 0 : tracing::info!(%node_id, "Drain background operation was cancelled");
4568 : }
4569 0 : Err(err) => {
4570 0 : tracing::error!(%node_id, "Drain background operation encountered: {err}")
4571 : }
4572 : }
4573 0 : }
4574 0 : });
4575 : }
4576 : NodeSchedulingPolicy::Draining => {
4577 0 : return Err(ApiError::Conflict(format!(
4578 0 : "Node {node_id} has drain in progress"
4579 0 : )));
4580 : }
4581 0 : policy => {
4582 0 : return Err(ApiError::PreconditionFailed(
4583 0 : format!("Node {node_id} cannot be drained due to {policy:?} policy").into(),
4584 0 : ));
4585 : }
4586 : }
4587 :
4588 0 : Ok(())
4589 0 : }
4590 :
4591 0 : pub(crate) async fn start_node_fill(self: &Arc<Self>, node_id: NodeId) -> Result<(), ApiError> {
4592 0 : let (ongoing_op, node_available, node_policy, total_nodes_count) = {
4593 0 : let locked = self.inner.read().unwrap();
4594 0 : let nodes = &locked.nodes;
4595 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
4596 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
4597 0 : ))?;
4598 :
4599 0 : (
4600 0 : locked
4601 0 : .ongoing_operation
4602 0 : .as_ref()
4603 0 : .map(|ongoing| ongoing.operation),
4604 0 : node.is_available(),
4605 0 : node.get_scheduling(),
4606 0 : nodes.len(),
4607 0 : )
4608 : };
4609 :
4610 0 : if let Some(ongoing) = ongoing_op {
4611 0 : return Err(ApiError::PreconditionFailed(
4612 0 : format!("Background operation already ongoing for node: {}", ongoing).into(),
4613 0 : ));
4614 0 : }
4615 0 :
4616 0 : if !node_available {
4617 0 : return Err(ApiError::ResourceUnavailable(
4618 0 : format!("Node {node_id} is currently unavailable").into(),
4619 0 : ));
4620 0 : }
4621 0 :
4622 0 : if total_nodes_count <= 1 {
4623 0 : return Err(ApiError::PreconditionFailed(
4624 0 : "No other nodes to fill from".into(),
4625 0 : ));
4626 0 : }
4627 0 :
4628 0 : match node_policy {
4629 : NodeSchedulingPolicy::Active => {
4630 0 : self.node_configure(node_id, None, Some(NodeSchedulingPolicy::Filling))
4631 0 : .await?;
4632 :
4633 0 : let cancel = CancellationToken::new();
4634 0 :
4635 0 : self.inner.write().unwrap().ongoing_operation = Some(OperationHandler {
4636 0 : operation: Operation::Fill(Fill { node_id }),
4637 0 : cancel: cancel.clone(),
4638 0 : });
4639 0 :
4640 0 : tokio::task::spawn({
4641 0 : let service = self.clone();
4642 0 : let cancel = cancel.clone();
4643 0 : async move {
4644 : scopeguard::defer! {
4645 : let prev = service.inner.write().unwrap().ongoing_operation.take();
4646 :
4647 0 : if let Some(Operation::Fill(removed_fill)) = prev.map(|h| h.operation) {
4648 : assert_eq!(removed_fill.node_id, node_id, "We always take the same operation");
4649 : } else {
4650 : panic!("We always remove the same operation")
4651 : }
4652 : }
4653 :
4654 0 : tracing::info!(%node_id, "Fill background operation starting");
4655 0 : let res = service.fill_node(node_id, cancel).await;
4656 0 : match res {
4657 : Ok(()) => {
4658 0 : tracing::info!(%node_id, "Fill background operation completed successfully");
4659 : }
4660 : Err(OperationError::Cancelled) => {
4661 0 : tracing::info!(%node_id, "Fill background operation was cancelled");
4662 : }
4663 0 : Err(err) => {
4664 0 : tracing::error!(%node_id, "Fill background operation encountered: {err}")
4665 : }
4666 : }
4667 0 : }
4668 0 : });
4669 : }
4670 : NodeSchedulingPolicy::Filling => {
4671 0 : return Err(ApiError::Conflict(format!(
4672 0 : "Node {node_id} has fill in progress"
4673 0 : )));
4674 : }
4675 0 : policy => {
4676 0 : return Err(ApiError::PreconditionFailed(
4677 0 : format!("Node {node_id} cannot be filled due to {policy:?} policy").into(),
4678 0 : ));
4679 : }
4680 : }
4681 :
4682 0 : Ok(())
4683 0 : }
4684 :
4685 : /// Helper for methods that will try and call pageserver APIs for
4686 : /// a tenant, such as timeline CRUD: they cannot proceed unless the tenant
4687 : /// is attached somewhere.
4688 0 : fn ensure_attached_schedule(
4689 0 : &self,
4690 0 : mut locked: std::sync::RwLockWriteGuard<'_, ServiceState>,
4691 0 : tenant_id: TenantId,
4692 0 : ) -> Result<Vec<ReconcilerWaiter>, anyhow::Error> {
4693 0 : let mut waiters = Vec::new();
4694 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
4695 0 :
4696 0 : let mut schedule_context = ScheduleContext::default();
4697 0 : for (tenant_shard_id, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
4698 0 : shard.schedule(scheduler, &mut schedule_context)?;
4699 :
4700 : // The shard's policies may not result in an attached location being scheduled: this
4701 : // is an error because our caller needs it attached somewhere.
4702 0 : if shard.intent.get_attached().is_none() {
4703 0 : return Err(anyhow::anyhow!(
4704 0 : "Tenant {tenant_id} not scheduled to be attached"
4705 0 : ));
4706 0 : };
4707 0 :
4708 0 : if shard.stably_attached().is_some() {
4709 : // We do not require the shard to be totally up to date on reconciliation: we just require
4710 : // that it has been attached on the intended node. Other dirty state such as unattached secondary
4711 : // locations, or compute hook notifications can be ignored.
4712 0 : continue;
4713 0 : }
4714 :
4715 0 : if let Some(waiter) = self.maybe_reconcile_shard(shard, nodes) {
4716 0 : tracing::info!("Waiting for shard {tenant_shard_id} to reconcile, in order to ensure it is attached");
4717 0 : waiters.push(waiter);
4718 0 : }
4719 : }
4720 0 : Ok(waiters)
4721 0 : }
4722 :
4723 0 : async fn ensure_attached_wait(&self, tenant_id: TenantId) -> Result<(), ApiError> {
4724 0 : let ensure_waiters = {
4725 0 : let locked = self.inner.write().unwrap();
4726 :
4727 : // Check if the tenant is splitting: in this case, even if it is attached,
4728 : // we must act as if it is not: this blocks e.g. timeline creation/deletion
4729 : // operations during the split.
4730 0 : for (_shard_id, shard) in locked.tenants.range(TenantShardId::tenant_range(tenant_id)) {
4731 0 : if !matches!(shard.splitting, SplitState::Idle) {
4732 0 : return Err(ApiError::ResourceUnavailable(
4733 0 : "Tenant shards are currently splitting".into(),
4734 0 : ));
4735 0 : }
4736 : }
4737 :
4738 0 : self.ensure_attached_schedule(locked, tenant_id)
4739 0 : .map_err(ApiError::InternalServerError)?
4740 : };
4741 :
4742 0 : let deadline = Instant::now().checked_add(Duration::from_secs(5)).unwrap();
4743 0 : for waiter in ensure_waiters {
4744 0 : let timeout = deadline.duration_since(Instant::now());
4745 0 : waiter.wait_timeout(timeout).await?;
4746 : }
4747 :
4748 0 : Ok(())
4749 0 : }
4750 :
4751 : /// Wrap [`TenantShard`] reconciliation methods with acquisition of [`Gate`] and [`ReconcileUnits`],
4752 0 : fn maybe_reconcile_shard(
4753 0 : &self,
4754 0 : shard: &mut TenantShard,
4755 0 : nodes: &Arc<HashMap<NodeId, Node>>,
4756 0 : ) -> Option<ReconcilerWaiter> {
4757 0 : let reconcile_needed = shard.get_reconcile_needed(nodes);
4758 0 :
4759 0 : match reconcile_needed {
4760 0 : ReconcileNeeded::No => return None,
4761 0 : ReconcileNeeded::WaitExisting(waiter) => return Some(waiter),
4762 0 : ReconcileNeeded::Yes => {
4763 0 : // Fall through to try and acquire units for spawning reconciler
4764 0 : }
4765 : };
4766 :
4767 0 : let units = match self.reconciler_concurrency.clone().try_acquire_owned() {
4768 0 : Ok(u) => ReconcileUnits::new(u),
4769 : Err(_) => {
4770 0 : tracing::info!(tenant_id=%shard.tenant_shard_id.tenant_id, shard_id=%shard.tenant_shard_id.shard_slug(),
4771 0 : "Concurrency limited: enqueued for reconcile later");
4772 0 : if !shard.delayed_reconcile {
4773 0 : match self.delayed_reconcile_tx.try_send(shard.tenant_shard_id) {
4774 0 : Err(TrySendError::Closed(_)) => {
4775 0 : // Weird mid-shutdown case?
4776 0 : }
4777 : Err(TrySendError::Full(_)) => {
4778 : // It is safe to skip sending our ID in the channel: we will eventually get retried by the background reconcile task.
4779 0 : tracing::warn!(
4780 0 : "Many shards are waiting to reconcile: delayed_reconcile queue is full"
4781 : );
4782 : }
4783 0 : Ok(()) => {
4784 0 : shard.delayed_reconcile = true;
4785 0 : }
4786 : }
4787 0 : }
4788 :
4789 : // We won't spawn a reconciler, but we will construct a waiter that waits for the shard's sequence
4790 : // number to advance. When this function is eventually called again and succeeds in getting units,
4791 : // it will spawn a reconciler that makes this waiter complete.
4792 0 : return Some(shard.future_reconcile_waiter());
4793 : }
4794 : };
4795 :
4796 0 : let Ok(gate_guard) = self.gate.enter() else {
4797 : // Gate closed: we're shutting down, drop out.
4798 0 : return None;
4799 : };
4800 :
4801 0 : shard.spawn_reconciler(
4802 0 : &self.result_tx,
4803 0 : nodes,
4804 0 : &self.compute_hook,
4805 0 : &self.config,
4806 0 : &self.persistence,
4807 0 : units,
4808 0 : gate_guard,
4809 0 : &self.cancel,
4810 0 : )
4811 0 : }
4812 :
4813 : /// Check all tenants for pending reconciliation work, and reconcile those in need.
4814 : /// Additionally, reschedule tenants that require it.
4815 : ///
4816 : /// Returns how many reconciliation tasks were started, or `1` if no reconciles were
4817 : /// spawned but some _would_ have been spawned if `reconciler_concurrency` units where
4818 : /// available. A return value of 0 indicates that everything is fully reconciled already.
4819 0 : fn reconcile_all(&self) -> usize {
4820 0 : let mut locked = self.inner.write().unwrap();
4821 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
4822 0 : let pageservers = nodes.clone();
4823 0 :
4824 0 : let mut schedule_context = ScheduleContext::default();
4825 0 :
4826 0 : let mut reconciles_spawned = 0;
4827 0 : for (tenant_shard_id, shard) in tenants.iter_mut() {
4828 0 : if tenant_shard_id.is_shard_zero() {
4829 0 : schedule_context = ScheduleContext::default();
4830 0 : }
4831 :
4832 : // Skip checking if this shard is already enqueued for reconciliation
4833 0 : if shard.delayed_reconcile && self.reconciler_concurrency.available_permits() == 0 {
4834 : // If there is something delayed, then return a nonzero count so that
4835 : // callers like reconcile_all_now do not incorrectly get the impression
4836 : // that the system is in a quiescent state.
4837 0 : reconciles_spawned = std::cmp::max(1, reconciles_spawned);
4838 0 : continue;
4839 0 : }
4840 0 :
4841 0 : // Eventual consistency: if an earlier reconcile job failed, and the shard is still
4842 0 : // dirty, spawn another rone
4843 0 : if self.maybe_reconcile_shard(shard, &pageservers).is_some() {
4844 0 : reconciles_spawned += 1;
4845 0 : }
4846 :
4847 0 : schedule_context.avoid(&shard.intent.all_pageservers());
4848 : }
4849 :
4850 0 : reconciles_spawned
4851 0 : }
4852 :
4853 : /// `optimize` in this context means identifying shards which have valid scheduled locations, but
4854 : /// could be scheduled somewhere better:
4855 : /// - Cutting over to a secondary if the node with the secondary is more lightly loaded
4856 : /// * e.g. after a node fails then recovers, to move some work back to it
4857 : /// - Cutting over to a secondary if it improves the spread of shard attachments within a tenant
4858 : /// * e.g. after a shard split, the initial attached locations will all be on the node where
4859 : /// we did the split, but are probably better placed elsewhere.
4860 : /// - Creating new secondary locations if it improves the spreading of a sharded tenant
4861 : /// * e.g. after a shard split, some locations will be on the same node (where the split
4862 : /// happened), and will probably be better placed elsewhere.
4863 : ///
4864 : /// To put it more briefly: whereas the scheduler respects soft constraints in a ScheduleContext at
4865 : /// the time of scheduling, this function looks for cases where a better-scoring location is available
4866 : /// according to those same soft constraints.
4867 0 : async fn optimize_all(&self) -> usize {
4868 0 : // Limit on how many shards' optmizations each call to this function will execute. Combined
4869 0 : // with the frequency of background calls, this acts as an implicit rate limit that runs a small
4870 0 : // trickle of optimizations in the background, rather than executing a large number in parallel
4871 0 : // when a change occurs.
4872 0 : const MAX_OPTIMIZATIONS_EXEC_PER_PASS: usize = 2;
4873 0 :
4874 0 : // Synchronous prepare: scan shards for possible scheduling optimizations
4875 0 : let candidate_work = self.optimize_all_plan();
4876 0 : let candidate_work_len = candidate_work.len();
4877 :
4878 : // Asynchronous validate: I/O to pageservers to make sure shards are in a good state to apply validation
4879 0 : let validated_work = self.optimize_all_validate(candidate_work).await;
4880 :
4881 0 : let was_work_filtered = validated_work.len() != candidate_work_len;
4882 0 :
4883 0 : // Synchronous apply: update the shards' intent states according to validated optimisations
4884 0 : let mut reconciles_spawned = 0;
4885 0 : let mut optimizations_applied = 0;
4886 0 : let mut locked = self.inner.write().unwrap();
4887 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
4888 0 : for (tenant_shard_id, optimization) in validated_work {
4889 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
4890 : // Shard was dropped between planning and execution;
4891 0 : continue;
4892 : };
4893 0 : if shard.apply_optimization(scheduler, optimization) {
4894 0 : optimizations_applied += 1;
4895 0 : if self.maybe_reconcile_shard(shard, nodes).is_some() {
4896 0 : reconciles_spawned += 1;
4897 0 : }
4898 0 : }
4899 :
4900 0 : if optimizations_applied >= MAX_OPTIMIZATIONS_EXEC_PER_PASS {
4901 0 : break;
4902 0 : }
4903 : }
4904 :
4905 0 : if was_work_filtered {
4906 0 : // If we filtered any work out during validation, ensure we return a nonzero value to indicate
4907 0 : // to callers that the system is not in a truly quiet state, it's going to do some work as soon
4908 0 : // as these validations start passing.
4909 0 : reconciles_spawned = std::cmp::max(reconciles_spawned, 1);
4910 0 : }
4911 :
4912 0 : reconciles_spawned
4913 0 : }
4914 :
4915 0 : fn optimize_all_plan(&self) -> Vec<(TenantShardId, ScheduleOptimization)> {
4916 0 : let mut schedule_context = ScheduleContext::default();
4917 0 :
4918 0 : let mut tenant_shards: Vec<&TenantShard> = Vec::new();
4919 0 :
4920 0 : // How many candidate optimizations we will generate, before evaluating them for readniess: setting
4921 0 : // this higher than the execution limit gives us a chance to execute some work even if the first
4922 0 : // few optimizations we find are not ready.
4923 0 : const MAX_OPTIMIZATIONS_PLAN_PER_PASS: usize = 8;
4924 0 :
4925 0 : let mut work = Vec::new();
4926 0 :
4927 0 : let mut locked = self.inner.write().unwrap();
4928 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
4929 0 : for (tenant_shard_id, shard) in tenants.iter() {
4930 0 : if tenant_shard_id.is_shard_zero() {
4931 0 : // Reset accumulators on the first shard in a tenant
4932 0 : schedule_context = ScheduleContext::default();
4933 0 : schedule_context.mode = ScheduleMode::Speculative;
4934 0 : tenant_shards.clear();
4935 0 : }
4936 :
4937 0 : if work.len() >= MAX_OPTIMIZATIONS_PLAN_PER_PASS {
4938 0 : break;
4939 0 : }
4940 0 :
4941 0 : match shard.get_scheduling_policy() {
4942 0 : ShardSchedulingPolicy::Active => {
4943 0 : // Ok to do optimization
4944 0 : }
4945 : ShardSchedulingPolicy::Essential
4946 : | ShardSchedulingPolicy::Pause
4947 : | ShardSchedulingPolicy::Stop => {
4948 : // Policy prevents optimizing this shard.
4949 0 : continue;
4950 : }
4951 : }
4952 :
4953 : // Accumulate the schedule context for all the shards in a tenant: we must have
4954 : // the total view of all shards before we can try to optimize any of them.
4955 0 : schedule_context.avoid(&shard.intent.all_pageservers());
4956 0 : if let Some(attached) = shard.intent.get_attached() {
4957 0 : schedule_context.push_attached(*attached);
4958 0 : }
4959 0 : tenant_shards.push(shard);
4960 0 :
4961 0 : // Once we have seen the last shard in the tenant, proceed to search across all shards
4962 0 : // in the tenant for optimizations
4963 0 : if shard.shard.number.0 == shard.shard.count.count() - 1 {
4964 0 : if tenant_shards.iter().any(|s| s.reconciler.is_some()) {
4965 : // Do not start any optimizations while another change to the tenant is ongoing: this
4966 : // is not necessary for correctness, but simplifies operations and implicitly throttles
4967 : // optimization changes to happen in a "trickle" over time.
4968 0 : continue;
4969 0 : }
4970 0 :
4971 0 : if tenant_shards.iter().any(|s| {
4972 0 : !matches!(s.splitting, SplitState::Idle)
4973 0 : || matches!(s.policy, PlacementPolicy::Detached)
4974 0 : }) {
4975 : // Never attempt to optimize a tenant that is currently being split, or
4976 : // a tenant that is meant to be detached
4977 0 : continue;
4978 0 : }
4979 :
4980 : // TODO: optimization calculations are relatively expensive: create some fast-path for
4981 : // the common idle case (avoiding the search on tenants that we have recently checked)
4982 :
4983 0 : for shard in &tenant_shards {
4984 0 : if let Some(optimization) =
4985 : // If idle, maybe ptimize attachments: if a shard has a secondary location that is preferable to
4986 : // its primary location based on soft constraints, cut it over.
4987 0 : shard.optimize_attachment(nodes, &schedule_context)
4988 : {
4989 0 : work.push((shard.tenant_shard_id, optimization));
4990 0 : break;
4991 0 : } else if let Some(optimization) =
4992 : // If idle, maybe optimize secondary locations: if a shard has a secondary location that would be
4993 : // better placed on another node, based on ScheduleContext, then adjust it. This
4994 : // covers cases like after a shard split, where we might have too many shards
4995 : // in the same tenant with secondary locations on the node where they originally split.
4996 0 : shard.optimize_secondary(scheduler, &schedule_context)
4997 : {
4998 0 : work.push((shard.tenant_shard_id, optimization));
4999 0 : break;
5000 0 : }
5001 :
5002 : // TODO: extend this mechanism to prefer attaching on nodes with fewer attached
5003 : // tenants (i.e. extend schedule state to distinguish attached from secondary counts),
5004 : // for the total number of attachments on a node (not just within a tenant.)
5005 : }
5006 0 : }
5007 : }
5008 :
5009 0 : work
5010 0 : }
5011 :
5012 0 : async fn optimize_all_validate(
5013 0 : &self,
5014 0 : candidate_work: Vec<(TenantShardId, ScheduleOptimization)>,
5015 0 : ) -> Vec<(TenantShardId, ScheduleOptimization)> {
5016 0 : // Take a clone of the node map to use outside the lock in async validation phase
5017 0 : let validation_nodes = { self.inner.read().unwrap().nodes.clone() };
5018 0 :
5019 0 : let mut want_secondary_status = Vec::new();
5020 0 :
5021 0 : // Validate our plans: this is an async phase where we may do I/O to pageservers to
5022 0 : // check that the state of locations is acceptable to run the optimization, such as
5023 0 : // checking that a secondary location is sufficiently warmed-up to cleanly cut over
5024 0 : // in a live migration.
5025 0 : let mut validated_work = Vec::new();
5026 0 : for (tenant_shard_id, optimization) in candidate_work {
5027 0 : match optimization.action {
5028 : ScheduleOptimizationAction::MigrateAttachment(MigrateAttachment {
5029 : old_attached_node_id: _,
5030 0 : new_attached_node_id,
5031 0 : }) => {
5032 0 : match validation_nodes.get(&new_attached_node_id) {
5033 0 : None => {
5034 0 : // Node was dropped between planning and validation
5035 0 : }
5036 0 : Some(node) => {
5037 0 : if !node.is_available() {
5038 0 : tracing::info!("Skipping optimization migration of {tenant_shard_id} to {new_attached_node_id} because node unavailable");
5039 0 : } else {
5040 0 : // Accumulate optimizations that require fetching secondary status, so that we can execute these
5041 0 : // remote API requests concurrently.
5042 0 : want_secondary_status.push((
5043 0 : tenant_shard_id,
5044 0 : node.clone(),
5045 0 : optimization,
5046 0 : ));
5047 0 : }
5048 : }
5049 : }
5050 : }
5051 : ScheduleOptimizationAction::ReplaceSecondary(_) => {
5052 : // No extra checks needed to replace a secondary: this does not interrupt client access
5053 0 : validated_work.push((tenant_shard_id, optimization))
5054 : }
5055 : };
5056 : }
5057 :
5058 : // Call into pageserver API to find out if the destination secondary location is warm enough for a reasonably smooth migration: we
5059 : // do this so that we avoid spawning a Reconciler that would have to wait minutes/hours for a destination to warm up: that reconciler
5060 : // would hold a precious reconcile semaphore unit the whole time it was waiting for the destination to warm up.
5061 0 : let results = self
5062 0 : .tenant_for_shards_api(
5063 0 : want_secondary_status
5064 0 : .iter()
5065 0 : .map(|i| (i.0, i.1.clone()))
5066 0 : .collect(),
5067 0 : |tenant_shard_id, client| async move {
5068 0 : client.tenant_secondary_status(tenant_shard_id).await
5069 0 : },
5070 0 : 1,
5071 0 : 1,
5072 0 : SHORT_RECONCILE_TIMEOUT,
5073 0 : &self.cancel,
5074 0 : )
5075 0 : .await;
5076 :
5077 0 : for ((tenant_shard_id, node, optimization), secondary_status) in
5078 0 : want_secondary_status.into_iter().zip(results.into_iter())
5079 : {
5080 0 : match secondary_status {
5081 0 : Err(e) => {
5082 0 : tracing::info!("Skipping migration of {tenant_shard_id} to {node}, error querying secondary: {e}");
5083 : }
5084 0 : Ok(progress) => {
5085 0 : // We require secondary locations to have less than 10GiB of downloads pending before we will use
5086 0 : // them in an optimization
5087 0 : const DOWNLOAD_FRESHNESS_THRESHOLD: u64 = 10 * 1024 * 1024 * 1024;
5088 0 :
5089 0 : if progress.heatmap_mtime.is_none()
5090 0 : || progress.bytes_total < DOWNLOAD_FRESHNESS_THRESHOLD
5091 0 : && progress.bytes_downloaded != progress.bytes_total
5092 0 : || progress.bytes_total - progress.bytes_downloaded
5093 0 : > DOWNLOAD_FRESHNESS_THRESHOLD
5094 : {
5095 0 : tracing::info!("Skipping migration of {tenant_shard_id} to {node} because secondary isn't ready: {progress:?}");
5096 : } else {
5097 : // Location looks ready: proceed
5098 0 : tracing::info!(
5099 0 : "{tenant_shard_id} secondary on {node} is warm enough for migration: {progress:?}"
5100 : );
5101 0 : validated_work.push((tenant_shard_id, optimization))
5102 : }
5103 : }
5104 : }
5105 : }
5106 :
5107 0 : validated_work
5108 0 : }
5109 :
5110 : /// Look for shards which are oversized and in need of splitting
5111 0 : async fn autosplit_tenants(self: &Arc<Self>) {
5112 0 : let Some(split_threshold) = self.config.split_threshold else {
5113 : // Auto-splitting is disabled
5114 0 : return;
5115 : };
5116 :
5117 0 : let nodes = self.inner.read().unwrap().nodes.clone();
5118 0 :
5119 0 : const SPLIT_TO_MAX: ShardCount = ShardCount::new(8);
5120 0 :
5121 0 : let mut top_n = Vec::new();
5122 0 :
5123 0 : // Call into each node to look for big tenants
5124 0 : let top_n_request = TopTenantShardsRequest {
5125 0 : // We currently split based on logical size, for simplicity: logical size is a signal of
5126 0 : // the user's intent to run a large database, whereas physical/resident size can be symptoms
5127 0 : // of compaction issues. Eventually we should switch to using resident size to bound the
5128 0 : // disk space impact of one shard.
5129 0 : order_by: models::TenantSorting::MaxLogicalSize,
5130 0 : limit: 10,
5131 0 : where_shards_lt: Some(SPLIT_TO_MAX),
5132 0 : where_gt: Some(split_threshold),
5133 0 : };
5134 0 : for node in nodes.values() {
5135 0 : let request_ref = &top_n_request;
5136 0 : match node
5137 0 : .with_client_retries(
5138 0 : |client| async move {
5139 0 : let request = request_ref.clone();
5140 0 : client.top_tenant_shards(request.clone()).await
5141 0 : },
5142 0 : &self.config.jwt_token,
5143 0 : 3,
5144 0 : 3,
5145 0 : Duration::from_secs(5),
5146 0 : &self.cancel,
5147 0 : )
5148 0 : .await
5149 : {
5150 0 : Some(Ok(node_top_n)) => {
5151 0 : top_n.extend(node_top_n.shards.into_iter());
5152 0 : }
5153 : Some(Err(mgmt_api::Error::Cancelled)) => {
5154 0 : continue;
5155 : }
5156 0 : Some(Err(e)) => {
5157 0 : tracing::warn!("Failed to fetch top N tenants from {node}: {e}");
5158 0 : continue;
5159 : }
5160 : None => {
5161 : // Node is shutting down
5162 0 : continue;
5163 : }
5164 : };
5165 : }
5166 :
5167 : // Pick the biggest tenant to split first
5168 0 : top_n.sort_by_key(|i| i.resident_size);
5169 0 : let Some(split_candidate) = top_n.into_iter().next() else {
5170 0 : tracing::debug!("No split-elegible shards found");
5171 0 : return;
5172 : };
5173 :
5174 : // We spawn a task to run this, so it's exactly like some external API client requesting it. We don't
5175 : // want to block the background reconcile loop on this.
5176 0 : tracing::info!("Auto-splitting tenant for size threshold {split_threshold}: current size {split_candidate:?}");
5177 :
5178 0 : let this = self.clone();
5179 0 : tokio::spawn(
5180 0 : async move {
5181 0 : match this
5182 0 : .tenant_shard_split(
5183 0 : split_candidate.id.tenant_id,
5184 0 : TenantShardSplitRequest {
5185 0 : // Always split to the max number of shards: this avoids stepping through
5186 0 : // intervening shard counts and encountering the overrhead of a split+cleanup
5187 0 : // each time as a tenant grows, and is not too expensive because our max shard
5188 0 : // count is relatively low anyway.
5189 0 : // This policy will be adjusted in future once we support higher shard count.
5190 0 : new_shard_count: SPLIT_TO_MAX.literal(),
5191 0 : new_stripe_size: Some(ShardParameters::DEFAULT_STRIPE_SIZE),
5192 0 : },
5193 0 : )
5194 0 : .await
5195 : {
5196 : Ok(_) => {
5197 0 : tracing::info!("Successful auto-split");
5198 : }
5199 0 : Err(e) => {
5200 0 : tracing::error!("Auto-split failed: {e}");
5201 : }
5202 : }
5203 0 : }
5204 0 : .instrument(tracing::info_span!("auto_split", tenant_id=%split_candidate.id.tenant_id)),
5205 : );
5206 0 : }
5207 :
5208 : /// Useful for tests: run whatever work a background [`Self::reconcile_all`] would have done, but
5209 : /// also wait for any generated Reconcilers to complete. Calling this until it returns zero should
5210 : /// put the system into a quiescent state where future background reconciliations won't do anything.
5211 0 : pub(crate) async fn reconcile_all_now(&self) -> Result<usize, ReconcileWaitError> {
5212 0 : let reconciles_spawned = self.reconcile_all();
5213 0 : let reconciles_spawned = if reconciles_spawned == 0 {
5214 : // Only optimize when we are otherwise idle
5215 0 : self.optimize_all().await
5216 : } else {
5217 0 : reconciles_spawned
5218 : };
5219 :
5220 0 : let waiters = {
5221 0 : let mut waiters = Vec::new();
5222 0 : let locked = self.inner.read().unwrap();
5223 0 : for (_tenant_shard_id, shard) in locked.tenants.iter() {
5224 0 : if let Some(waiter) = shard.get_waiter() {
5225 0 : waiters.push(waiter);
5226 0 : }
5227 : }
5228 0 : waiters
5229 0 : };
5230 0 :
5231 0 : let waiter_count = waiters.len();
5232 0 : match self.await_waiters(waiters, RECONCILE_TIMEOUT).await {
5233 0 : Ok(()) => {}
5234 0 : Err(ReconcileWaitError::Failed(_, reconcile_error))
5235 0 : if matches!(*reconcile_error, ReconcileError::Cancel) =>
5236 0 : {
5237 0 : // Ignore reconciler cancel errors: this reconciler might have shut down
5238 0 : // because some other change superceded it. We will return a nonzero number,
5239 0 : // so the caller knows they might have to call again to quiesce the system.
5240 0 : }
5241 0 : Err(e) => {
5242 0 : return Err(e);
5243 : }
5244 : };
5245 :
5246 0 : tracing::info!(
5247 0 : "{} reconciles in reconcile_all, {} waiters",
5248 : reconciles_spawned,
5249 : waiter_count
5250 : );
5251 :
5252 0 : Ok(std::cmp::max(waiter_count, reconciles_spawned))
5253 0 : }
5254 :
5255 0 : pub async fn shutdown(&self) {
5256 0 : // Note that this already stops processing any results from reconciles: so
5257 0 : // we do not expect that our [`TenantShard`] objects will reach a neat
5258 0 : // final state.
5259 0 : self.cancel.cancel();
5260 0 :
5261 0 : // The cancellation tokens in [`crate::reconciler::Reconciler`] are children
5262 0 : // of our cancellation token, so we do not need to explicitly cancel each of
5263 0 : // them.
5264 0 :
5265 0 : // Background tasks and reconcilers hold gate guards: this waits for them all
5266 0 : // to complete.
5267 0 : self.gate.close().await;
5268 0 : }
5269 :
5270 : /// Drain a node by moving the shards attached to it as primaries.
5271 : /// This is a long running operation and it should run as a separate Tokio task.
5272 0 : pub(crate) async fn drain_node(
5273 0 : &self,
5274 0 : node_id: NodeId,
5275 0 : cancel: CancellationToken,
5276 0 : ) -> Result<(), OperationError> {
5277 0 : let mut last_inspected_shard: Option<TenantShardId> = None;
5278 0 : let mut inspected_all_shards = false;
5279 0 : let mut waiters = Vec::new();
5280 0 : let mut schedule_context = ScheduleContext::default();
5281 :
5282 0 : while !inspected_all_shards {
5283 0 : if cancel.is_cancelled() {
5284 0 : return Err(OperationError::Cancelled);
5285 0 : }
5286 0 :
5287 0 : {
5288 0 : let mut locked = self.inner.write().unwrap();
5289 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
5290 :
5291 0 : let node = nodes.get(&node_id).ok_or(OperationError::NodeStateChanged(
5292 0 : format!("node {node_id} was removed").into(),
5293 0 : ))?;
5294 :
5295 0 : let current_policy = node.get_scheduling();
5296 0 : if !matches!(current_policy, NodeSchedulingPolicy::Draining) {
5297 : // TODO(vlad): maybe cancel pending reconciles before erroring out. need to think
5298 : // about it
5299 0 : return Err(OperationError::NodeStateChanged(
5300 0 : format!("node {node_id} changed state to {current_policy:?}").into(),
5301 0 : ));
5302 0 : }
5303 0 :
5304 0 : let mut cursor = tenants.iter_mut().skip_while({
5305 0 : let skip_past = last_inspected_shard;
5306 0 : move |(tid, _)| match skip_past {
5307 0 : Some(last) => **tid != last,
5308 0 : None => false,
5309 0 : }
5310 0 : });
5311 :
5312 0 : while waiters.len() < MAX_RECONCILES_PER_OPERATION {
5313 0 : let (tid, tenant_shard) = match cursor.next() {
5314 0 : Some(some) => some,
5315 : None => {
5316 0 : inspected_all_shards = true;
5317 0 : break;
5318 : }
5319 : };
5320 :
5321 0 : if tenant_shard.intent.demote_attached(scheduler, node_id) {
5322 0 : match tenant_shard.schedule(scheduler, &mut schedule_context) {
5323 0 : Err(e) => {
5324 0 : tracing::warn!(
5325 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
5326 0 : "Scheduling error when draining pageserver {} : {e}", node_id
5327 : );
5328 : }
5329 : Ok(()) => {
5330 0 : let scheduled_to = tenant_shard.intent.get_attached();
5331 0 : tracing::info!(
5332 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
5333 0 : "Rescheduled shard while draining node {}: {} -> {:?}",
5334 : node_id,
5335 : node_id,
5336 : scheduled_to
5337 : );
5338 :
5339 0 : let waiter = self.maybe_reconcile_shard(tenant_shard, nodes);
5340 0 : if let Some(some) = waiter {
5341 0 : waiters.push(some);
5342 0 : }
5343 : }
5344 : }
5345 0 : }
5346 :
5347 0 : last_inspected_shard = Some(*tid);
5348 : }
5349 : }
5350 :
5351 0 : waiters = self
5352 0 : .await_waiters_remainder(waiters, SHORT_RECONCILE_TIMEOUT)
5353 0 : .await;
5354 : }
5355 :
5356 0 : while !waiters.is_empty() {
5357 0 : tracing::info!("Awaiting {} pending drain reconciliations", waiters.len());
5358 :
5359 0 : waiters = self
5360 0 : .await_waiters_remainder(waiters, SHORT_RECONCILE_TIMEOUT)
5361 0 : .await;
5362 : }
5363 :
5364 : // At this point we have done the best we could to drain shards from this node.
5365 : // Set the node scheduling policy to `[NodeSchedulingPolicy::PauseForRestart]`
5366 : // to complete the drain.
5367 0 : if let Err(err) = self
5368 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::PauseForRestart))
5369 0 : .await
5370 : {
5371 : // This is not fatal. Anything that is polling the node scheduling policy to detect
5372 : // the end of the drain operations will hang, but all such places should enforce an
5373 : // overall timeout. The scheduling policy will be updated upon node re-attach and/or
5374 : // by the counterpart fill operation.
5375 0 : return Err(OperationError::FinalizeError(
5376 0 : format!(
5377 0 : "Failed to finalise drain of {node_id} by setting scheduling policy to PauseForRestart: {err}"
5378 0 : )
5379 0 : .into(),
5380 0 : ));
5381 0 : }
5382 0 :
5383 0 : Ok(())
5384 0 : }
5385 :
5386 : /// Create a node fill plan (pick secondaries to promote) that meets the following requirements:
5387 : /// 1. The node should be filled until it reaches the expected cluster average of
5388 : /// attached shards. If there are not enough secondaries on the node, the plan stops early.
5389 : /// 2. Select tenant shards to promote such that the number of attached shards is balanced
5390 : /// throughout the cluster. We achieve this by picking tenant shards from each node,
5391 : /// starting from the ones with the largest number of attached shards, until the node
5392 : /// reaches the expected cluster average.
5393 0 : fn fill_node_plan(&self, node_id: NodeId) -> Vec<TenantShardId> {
5394 0 : let mut locked = self.inner.write().unwrap();
5395 0 : let fill_requirement = locked.scheduler.compute_fill_requirement(node_id);
5396 0 :
5397 0 : let mut tids_by_node = locked
5398 0 : .tenants
5399 0 : .iter_mut()
5400 0 : .filter_map(|(tid, tenant_shard)| {
5401 0 : if tenant_shard.intent.get_secondary().contains(&node_id) {
5402 0 : if let Some(primary) = tenant_shard.intent.get_attached() {
5403 0 : return Some((*primary, *tid));
5404 0 : }
5405 0 : }
5406 :
5407 0 : None
5408 0 : })
5409 0 : .into_group_map();
5410 0 :
5411 0 : let expected_attached = locked.scheduler.expected_attached_shard_count();
5412 0 : let nodes_by_load = locked.scheduler.nodes_by_attached_shard_count();
5413 0 :
5414 0 : let mut plan = Vec::new();
5415 0 : for (node_id, attached) in nodes_by_load {
5416 0 : if plan.len() >= fill_requirement
5417 0 : || tids_by_node.is_empty()
5418 0 : || attached <= expected_attached
5419 : {
5420 0 : break;
5421 0 : }
5422 0 :
5423 0 : let can_take = attached - expected_attached;
5424 0 : let mut remove_node = false;
5425 0 : for _ in 0..can_take {
5426 0 : match tids_by_node.get_mut(&node_id) {
5427 0 : Some(tids) => match tids.pop() {
5428 0 : Some(tid) => {
5429 0 : plan.push(tid);
5430 0 : }
5431 : None => {
5432 0 : remove_node = true;
5433 0 : break;
5434 : }
5435 : },
5436 : None => {
5437 0 : break;
5438 : }
5439 : }
5440 : }
5441 :
5442 0 : if remove_node {
5443 0 : tids_by_node.remove(&node_id);
5444 0 : }
5445 : }
5446 :
5447 0 : plan
5448 0 : }
5449 :
5450 : /// Fill a node by promoting its secondaries until the cluster is balanced
5451 : /// with regards to attached shard counts. Note that this operation only
5452 : /// makes sense as a counterpart to the drain implemented in [`Service::drain_node`].
5453 : /// This is a long running operation and it should run as a separate Tokio task.
5454 0 : pub(crate) async fn fill_node(
5455 0 : &self,
5456 0 : node_id: NodeId,
5457 0 : cancel: CancellationToken,
5458 0 : ) -> Result<(), OperationError> {
5459 0 : // TODO(vlad): Currently this operates on the assumption that all
5460 0 : // secondaries are warm. This is not always true (e.g. we just migrated the
5461 0 : // tenant). Take that into consideration by checking the secondary status.
5462 0 : let mut tids_to_promote = self.fill_node_plan(node_id);
5463 0 :
5464 0 : let mut waiters = Vec::new();
5465 0 : let mut schedule_context = ScheduleContext::default();
5466 :
5467 : // Execute the plan we've composed above. Before aplying each move from the plan,
5468 : // we validate to ensure that it has not gone stale in the meantime.
5469 0 : while !tids_to_promote.is_empty() {
5470 0 : if cancel.is_cancelled() {
5471 0 : return Err(OperationError::Cancelled);
5472 0 : }
5473 0 :
5474 0 : {
5475 0 : let mut locked = self.inner.write().unwrap();
5476 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
5477 :
5478 0 : let node = nodes.get(&node_id).ok_or(OperationError::NodeStateChanged(
5479 0 : format!("node {node_id} was removed").into(),
5480 0 : ))?;
5481 :
5482 0 : let current_policy = node.get_scheduling();
5483 0 : if !matches!(current_policy, NodeSchedulingPolicy::Filling) {
5484 : // TODO(vlad): maybe cancel pending reconciles before erroring out. need to think
5485 : // about it
5486 0 : return Err(OperationError::NodeStateChanged(
5487 0 : format!("node {node_id} changed state to {current_policy:?}").into(),
5488 0 : ));
5489 0 : }
5490 :
5491 0 : while waiters.len() < MAX_RECONCILES_PER_OPERATION {
5492 0 : if let Some(tid) = tids_to_promote.pop() {
5493 0 : if let Some(tenant_shard) = tenants.get_mut(&tid) {
5494 : // If the node being filled is not a secondary anymore,
5495 : // skip the promotion.
5496 0 : if !tenant_shard.intent.get_secondary().contains(&node_id) {
5497 0 : continue;
5498 0 : }
5499 0 :
5500 0 : let previously_attached_to = *tenant_shard.intent.get_attached();
5501 0 :
5502 0 : tenant_shard.intent.promote_attached(scheduler, node_id);
5503 0 : match tenant_shard.schedule(scheduler, &mut schedule_context) {
5504 0 : Err(e) => {
5505 0 : tracing::warn!(
5506 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
5507 0 : "Scheduling error when filling pageserver {} : {e}", node_id
5508 : );
5509 : }
5510 : Ok(()) => {
5511 0 : tracing::info!(
5512 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
5513 0 : "Rescheduled shard while filling node {}: {:?} -> {}",
5514 : node_id,
5515 : previously_attached_to,
5516 : node_id
5517 : );
5518 :
5519 0 : if let Some(waiter) =
5520 0 : self.maybe_reconcile_shard(tenant_shard, nodes)
5521 0 : {
5522 0 : waiters.push(waiter);
5523 0 : }
5524 : }
5525 : }
5526 0 : }
5527 : } else {
5528 0 : break;
5529 : }
5530 : }
5531 : }
5532 :
5533 0 : waiters = self
5534 0 : .await_waiters_remainder(waiters, SHORT_RECONCILE_TIMEOUT)
5535 0 : .await;
5536 : }
5537 :
5538 0 : while !waiters.is_empty() {
5539 0 : tracing::info!("Awaiting {} pending fill reconciliations", waiters.len());
5540 :
5541 0 : waiters = self
5542 0 : .await_waiters_remainder(waiters, SHORT_RECONCILE_TIMEOUT)
5543 0 : .await;
5544 : }
5545 :
5546 0 : if let Err(err) = self
5547 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
5548 0 : .await
5549 : {
5550 : // This isn't a huge issue since the filling process starts upon request. However, it
5551 : // will prevent the next drain from starting. The only case in which this can fail
5552 : // is database unavailability. Such a case will require manual intervention.
5553 0 : return Err(OperationError::FinalizeError(
5554 0 : format!("Failed to finalise fill of {node_id} by setting scheduling policy to Active: {err}")
5555 0 : .into(),
5556 0 : ));
5557 0 : }
5558 0 :
5559 0 : Ok(())
5560 0 : }
5561 : }
|