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