Line data Source code
1 : pub(crate) mod split_state;
2 : use std::collections::HashMap;
3 : use std::str::FromStr;
4 : use std::time::Duration;
5 : use std::time::Instant;
6 :
7 : use self::split_state::SplitState;
8 : use diesel::pg::PgConnection;
9 : use diesel::prelude::*;
10 : use diesel::Connection;
11 : use pageserver_api::controller_api::MetadataHealthRecord;
12 : use pageserver_api::controller_api::ShardSchedulingPolicy;
13 : use pageserver_api::controller_api::{NodeSchedulingPolicy, PlacementPolicy};
14 : use pageserver_api::models::TenantConfig;
15 : use pageserver_api::shard::ShardConfigError;
16 : use pageserver_api::shard::ShardIdentity;
17 : use pageserver_api::shard::ShardStripeSize;
18 : use pageserver_api::shard::{ShardCount, ShardNumber, TenantShardId};
19 : use serde::{Deserialize, Serialize};
20 : use utils::generation::Generation;
21 : use utils::id::{NodeId, TenantId};
22 :
23 : use crate::metrics::{
24 : DatabaseQueryErrorLabelGroup, DatabaseQueryLatencyLabelGroup, METRICS_REGISTRY,
25 : };
26 : use crate::node::Node;
27 :
28 : use diesel_migrations::{embed_migrations, EmbeddedMigrations};
29 : const MIGRATIONS: EmbeddedMigrations = embed_migrations!("./migrations");
30 :
31 : /// ## What do we store?
32 : ///
33 : /// The storage controller service does not store most of its state durably.
34 : ///
35 : /// The essential things to store durably are:
36 : /// - generation numbers, as these must always advance monotonically to ensure data safety.
37 : /// - Tenant's PlacementPolicy and TenantConfig, as the source of truth for these is something external.
38 : /// - Node's scheduling policies, as the source of truth for these is something external.
39 : ///
40 : /// Other things we store durably as an implementation detail:
41 : /// - Node's host/port: this could be avoided it we made nodes emit a self-registering heartbeat,
42 : /// but it is operationally simpler to make this service the authority for which nodes
43 : /// it talks to.
44 : ///
45 : /// ## Performance/efficiency
46 : ///
47 : /// The storage controller service does not go via the database for most things: there are
48 : /// a couple of places where we must, and where efficiency matters:
49 : /// - Incrementing generation numbers: the Reconciler has to wait for this to complete
50 : /// before it can attach a tenant, so this acts as a bound on how fast things like
51 : /// failover can happen.
52 : /// - Pageserver re-attach: we will increment many shards' generations when this happens,
53 : /// so it is important to avoid e.g. issuing O(N) queries.
54 : ///
55 : /// Database calls relating to nodes have low performance requirements, as they are very rarely
56 : /// updated, and reads of nodes are always from memory, not the database. We only require that
57 : /// we can UPDATE a node's scheduling mode reasonably quickly to mark a bad node offline.
58 : pub struct Persistence {
59 : connection_pool: diesel::r2d2::Pool<diesel::r2d2::ConnectionManager<PgConnection>>,
60 : }
61 :
62 : /// Legacy format, for use in JSON compat objects in test environment
63 0 : #[derive(Serialize, Deserialize)]
64 : struct JsonPersistence {
65 : tenants: HashMap<TenantShardId, TenantShardPersistence>,
66 : }
67 :
68 0 : #[derive(thiserror::Error, Debug)]
69 : pub(crate) enum DatabaseError {
70 : #[error(transparent)]
71 : Query(#[from] diesel::result::Error),
72 : #[error(transparent)]
73 : Connection(#[from] diesel::result::ConnectionError),
74 : #[error(transparent)]
75 : ConnectionPool(#[from] r2d2::Error),
76 : #[error("Logical error: {0}")]
77 : Logical(String),
78 : #[error("Migration error: {0}")]
79 : Migration(String),
80 : }
81 :
82 : #[derive(measured::FixedCardinalityLabel, Copy, Clone)]
83 : pub(crate) enum DatabaseOperation {
84 : InsertNode,
85 : UpdateNode,
86 : DeleteNode,
87 : ListNodes,
88 : BeginShardSplit,
89 : CompleteShardSplit,
90 : AbortShardSplit,
91 : Detach,
92 : ReAttach,
93 : IncrementGeneration,
94 : ListTenantShards,
95 : InsertTenantShards,
96 : UpdateTenantShard,
97 : DeleteTenant,
98 : UpdateTenantConfig,
99 : UpdateMetadataHealth,
100 : ListMetadataHealth,
101 : ListMetadataHealthUnhealthy,
102 : ListMetadataHealthOutdated,
103 : GetLeader,
104 : UpdateLeader,
105 : }
106 :
107 : #[must_use]
108 : pub(crate) enum AbortShardSplitStatus {
109 : /// We aborted the split in the database by reverting to the parent shards
110 : Aborted,
111 : /// The split had already been persisted.
112 : Complete,
113 : }
114 :
115 : pub(crate) type DatabaseResult<T> = Result<T, DatabaseError>;
116 :
117 : /// Some methods can operate on either a whole tenant or a single shard
118 : pub(crate) enum TenantFilter {
119 : Tenant(TenantId),
120 : Shard(TenantShardId),
121 : }
122 :
123 : impl Persistence {
124 : // The default postgres connection limit is 100. We use up to 99, to leave one free for a human admin under
125 : // normal circumstances. This assumes we have exclusive use of the database cluster to which we connect.
126 : pub const MAX_CONNECTIONS: u32 = 99;
127 :
128 : // We don't want to keep a lot of connections alive: close them down promptly if they aren't being used.
129 : const IDLE_CONNECTION_TIMEOUT: Duration = Duration::from_secs(10);
130 : const MAX_CONNECTION_LIFETIME: Duration = Duration::from_secs(60);
131 :
132 0 : pub fn new(database_url: String) -> Self {
133 0 : let manager = diesel::r2d2::ConnectionManager::<PgConnection>::new(database_url);
134 0 :
135 0 : // We will use a connection pool: this is primarily to _limit_ our connection count, rather than to optimize time
136 0 : // to execute queries (database queries are not generally on latency-sensitive paths).
137 0 : let connection_pool = diesel::r2d2::Pool::builder()
138 0 : .max_size(Self::MAX_CONNECTIONS)
139 0 : .max_lifetime(Some(Self::MAX_CONNECTION_LIFETIME))
140 0 : .idle_timeout(Some(Self::IDLE_CONNECTION_TIMEOUT))
141 0 : // Always keep at least one connection ready to go
142 0 : .min_idle(Some(1))
143 0 : .test_on_check_out(true)
144 0 : .build(manager)
145 0 : .expect("Could not build connection pool");
146 0 :
147 0 : Self { connection_pool }
148 0 : }
149 :
150 : /// A helper for use during startup, where we would like to tolerate concurrent restarts of the
151 : /// database and the storage controller, therefore the database might not be available right away
152 0 : pub async fn await_connection(
153 0 : database_url: &str,
154 0 : timeout: Duration,
155 0 : ) -> Result<(), diesel::ConnectionError> {
156 0 : let started_at = Instant::now();
157 0 : loop {
158 0 : match PgConnection::establish(database_url) {
159 : Ok(_) => {
160 0 : tracing::info!("Connected to database.");
161 0 : return Ok(());
162 : }
163 0 : Err(e) => {
164 0 : if started_at.elapsed() > timeout {
165 0 : return Err(e);
166 : } else {
167 0 : tracing::info!("Database not yet available, waiting... ({e})");
168 0 : tokio::time::sleep(Duration::from_millis(100)).await;
169 : }
170 : }
171 : }
172 : }
173 0 : }
174 :
175 : /// Execute the diesel migrations that are built into this binary
176 0 : pub(crate) async fn migration_run(&self) -> DatabaseResult<()> {
177 0 : use diesel_migrations::{HarnessWithOutput, MigrationHarness};
178 0 :
179 0 : self.with_conn(move |conn| -> DatabaseResult<()> {
180 0 : HarnessWithOutput::write_to_stdout(conn)
181 0 : .run_pending_migrations(MIGRATIONS)
182 0 : .map(|_| ())
183 0 : .map_err(|e| DatabaseError::Migration(e.to_string()))
184 0 : })
185 0 : .await
186 0 : }
187 :
188 : /// Wraps `with_conn` in order to collect latency and error metrics
189 0 : async fn with_measured_conn<F, R>(&self, op: DatabaseOperation, func: F) -> DatabaseResult<R>
190 0 : where
191 0 : F: Fn(&mut PgConnection) -> DatabaseResult<R> + Send + 'static,
192 0 : R: Send + 'static,
193 0 : {
194 0 : let latency = &METRICS_REGISTRY
195 0 : .metrics_group
196 0 : .storage_controller_database_query_latency;
197 0 : let _timer = latency.start_timer(DatabaseQueryLatencyLabelGroup { operation: op });
198 :
199 0 : let res = self.with_conn(func).await;
200 :
201 0 : if let Err(err) = &res {
202 0 : let error_counter = &METRICS_REGISTRY
203 0 : .metrics_group
204 0 : .storage_controller_database_query_error;
205 0 : error_counter.inc(DatabaseQueryErrorLabelGroup {
206 0 : error_type: err.error_label(),
207 0 : operation: op,
208 0 : })
209 0 : }
210 :
211 0 : res
212 0 : }
213 :
214 : /// Call the provided function in a tokio blocking thread, with a Diesel database connection.
215 0 : async fn with_conn<F, R>(&self, func: F) -> DatabaseResult<R>
216 0 : where
217 0 : F: Fn(&mut PgConnection) -> DatabaseResult<R> + Send + 'static,
218 0 : R: Send + 'static,
219 0 : {
220 : // A generous allowance for how many times we may retry serializable transactions
221 : // before giving up. This is not expected to be hit: it is a defensive measure in case we
222 : // somehow engineer a situation where duelling transactions might otherwise live-lock.
223 : const MAX_RETRIES: usize = 128;
224 :
225 0 : let mut conn = self.connection_pool.get()?;
226 0 : tokio::task::spawn_blocking(move || -> DatabaseResult<R> {
227 0 : let mut retry_count = 0;
228 : loop {
229 0 : match conn.build_transaction().serializable().run(|c| func(c)) {
230 0 : Ok(r) => break Ok(r),
231 : Err(
232 0 : err @ DatabaseError::Query(diesel::result::Error::DatabaseError(
233 0 : diesel::result::DatabaseErrorKind::SerializationFailure,
234 0 : _,
235 0 : )),
236 0 : ) => {
237 0 : retry_count += 1;
238 0 : if retry_count > MAX_RETRIES {
239 0 : tracing::error!(
240 0 : "Exceeded max retries on SerializationFailure errors: {err:?}"
241 : );
242 0 : break Err(err);
243 : } else {
244 : // Retry on serialization errors: these are expected, because even though our
245 : // transactions don't fight for the same rows, they will occasionally collide
246 : // on index pages (e.g. increment_generation for unrelated shards can collide)
247 0 : tracing::debug!(
248 0 : "Retrying transaction on serialization failure {err:?}"
249 : );
250 0 : continue;
251 : }
252 : }
253 0 : Err(e) => break Err(e),
254 : }
255 : }
256 0 : })
257 0 : .await
258 0 : .expect("Task panic")
259 0 : }
260 :
261 : /// When a node is first registered, persist it before using it for anything
262 0 : pub(crate) async fn insert_node(&self, node: &Node) -> DatabaseResult<()> {
263 0 : let np = node.to_persistent();
264 0 : self.with_measured_conn(
265 0 : DatabaseOperation::InsertNode,
266 0 : move |conn| -> DatabaseResult<()> {
267 0 : diesel::insert_into(crate::schema::nodes::table)
268 0 : .values(&np)
269 0 : .execute(conn)?;
270 0 : Ok(())
271 0 : },
272 0 : )
273 0 : .await
274 0 : }
275 :
276 : /// At startup, populate the list of nodes which our shards may be placed on
277 0 : pub(crate) async fn list_nodes(&self) -> DatabaseResult<Vec<NodePersistence>> {
278 0 : let nodes: Vec<NodePersistence> = self
279 0 : .with_measured_conn(
280 0 : DatabaseOperation::ListNodes,
281 0 : move |conn| -> DatabaseResult<_> {
282 0 : Ok(crate::schema::nodes::table.load::<NodePersistence>(conn)?)
283 0 : },
284 0 : )
285 0 : .await?;
286 :
287 0 : tracing::info!("list_nodes: loaded {} nodes", nodes.len());
288 :
289 0 : Ok(nodes)
290 0 : }
291 :
292 0 : pub(crate) async fn update_node(
293 0 : &self,
294 0 : input_node_id: NodeId,
295 0 : input_scheduling: NodeSchedulingPolicy,
296 0 : ) -> DatabaseResult<()> {
297 : use crate::schema::nodes::dsl::*;
298 0 : let updated = self
299 0 : .with_measured_conn(DatabaseOperation::UpdateNode, move |conn| {
300 0 : let updated = diesel::update(nodes)
301 0 : .filter(node_id.eq(input_node_id.0 as i64))
302 0 : .set((scheduling_policy.eq(String::from(input_scheduling)),))
303 0 : .execute(conn)?;
304 0 : Ok(updated)
305 0 : })
306 0 : .await?;
307 :
308 0 : if updated != 1 {
309 0 : Err(DatabaseError::Logical(format!(
310 0 : "Node {node_id:?} not found for update",
311 0 : )))
312 : } else {
313 0 : Ok(())
314 : }
315 0 : }
316 :
317 : /// At startup, load the high level state for shards, such as their config + policy. This will
318 : /// be enriched at runtime with state discovered on pageservers.
319 0 : pub(crate) async fn list_tenant_shards(&self) -> DatabaseResult<Vec<TenantShardPersistence>> {
320 0 : self.with_measured_conn(
321 0 : DatabaseOperation::ListTenantShards,
322 0 : move |conn| -> DatabaseResult<_> {
323 0 : Ok(crate::schema::tenant_shards::table.load::<TenantShardPersistence>(conn)?)
324 0 : },
325 0 : )
326 0 : .await
327 0 : }
328 :
329 : /// Tenants must be persisted before we schedule them for the first time. This enables us
330 : /// to correctly retain generation monotonicity, and the externally provided placement policy & config.
331 0 : pub(crate) async fn insert_tenant_shards(
332 0 : &self,
333 0 : shards: Vec<TenantShardPersistence>,
334 0 : ) -> DatabaseResult<()> {
335 0 : use crate::schema::metadata_health;
336 0 : use crate::schema::tenant_shards;
337 0 :
338 0 : let now = chrono::Utc::now();
339 0 :
340 0 : let metadata_health_records = shards
341 0 : .iter()
342 0 : .map(|t| MetadataHealthPersistence {
343 0 : tenant_id: t.tenant_id.clone(),
344 0 : shard_number: t.shard_number,
345 0 : shard_count: t.shard_count,
346 0 : healthy: true,
347 0 : last_scrubbed_at: now,
348 0 : })
349 0 : .collect::<Vec<_>>();
350 0 :
351 0 : self.with_measured_conn(
352 0 : DatabaseOperation::InsertTenantShards,
353 0 : move |conn| -> DatabaseResult<()> {
354 0 : diesel::insert_into(tenant_shards::table)
355 0 : .values(&shards)
356 0 : .execute(conn)?;
357 :
358 0 : diesel::insert_into(metadata_health::table)
359 0 : .values(&metadata_health_records)
360 0 : .execute(conn)?;
361 0 : Ok(())
362 0 : },
363 0 : )
364 0 : .await
365 0 : }
366 :
367 : /// Ordering: call this _after_ deleting the tenant on pageservers, but _before_ dropping state for
368 : /// the tenant from memory on this server.
369 0 : pub(crate) async fn delete_tenant(&self, del_tenant_id: TenantId) -> DatabaseResult<()> {
370 0 : use crate::schema::tenant_shards::dsl::*;
371 0 : self.with_measured_conn(
372 0 : DatabaseOperation::DeleteTenant,
373 0 : move |conn| -> DatabaseResult<()> {
374 0 : // `metadata_health` status (if exists) is also deleted based on the cascade behavior.
375 0 : diesel::delete(tenant_shards)
376 0 : .filter(tenant_id.eq(del_tenant_id.to_string()))
377 0 : .execute(conn)?;
378 0 : Ok(())
379 0 : },
380 0 : )
381 0 : .await
382 0 : }
383 :
384 0 : pub(crate) async fn delete_node(&self, del_node_id: NodeId) -> DatabaseResult<()> {
385 0 : use crate::schema::nodes::dsl::*;
386 0 : self.with_measured_conn(
387 0 : DatabaseOperation::DeleteNode,
388 0 : move |conn| -> DatabaseResult<()> {
389 0 : diesel::delete(nodes)
390 0 : .filter(node_id.eq(del_node_id.0 as i64))
391 0 : .execute(conn)?;
392 :
393 0 : Ok(())
394 0 : },
395 0 : )
396 0 : .await
397 0 : }
398 :
399 : /// When a tenant invokes the /re-attach API, this function is responsible for doing an efficient
400 : /// batched increment of the generations of all tenants whose generation_pageserver is equal to
401 : /// the node that called /re-attach.
402 0 : #[tracing::instrument(skip_all, fields(node_id))]
403 : pub(crate) async fn re_attach(
404 : &self,
405 : input_node_id: NodeId,
406 : ) -> DatabaseResult<HashMap<TenantShardId, Generation>> {
407 : use crate::schema::nodes::dsl::scheduling_policy;
408 : use crate::schema::nodes::dsl::*;
409 : use crate::schema::tenant_shards::dsl::*;
410 : let updated = self
411 0 : .with_measured_conn(DatabaseOperation::ReAttach, move |conn| {
412 0 : let rows_updated = diesel::update(tenant_shards)
413 0 : .filter(generation_pageserver.eq(input_node_id.0 as i64))
414 0 : .set(generation.eq(generation + 1))
415 0 : .execute(conn)?;
416 :
417 0 : tracing::info!("Incremented {} tenants' generations", rows_updated);
418 :
419 : // TODO: UPDATE+SELECT in one query
420 :
421 0 : let updated = tenant_shards
422 0 : .filter(generation_pageserver.eq(input_node_id.0 as i64))
423 0 : .select(TenantShardPersistence::as_select())
424 0 : .load(conn)?;
425 :
426 : // If the node went through a drain and restart phase before re-attaching,
427 : // then reset it's node scheduling policy to active.
428 0 : diesel::update(nodes)
429 0 : .filter(node_id.eq(input_node_id.0 as i64))
430 0 : .filter(
431 0 : scheduling_policy
432 0 : .eq(String::from(NodeSchedulingPolicy::PauseForRestart))
433 0 : .or(scheduling_policy.eq(String::from(NodeSchedulingPolicy::Draining)))
434 0 : .or(scheduling_policy.eq(String::from(NodeSchedulingPolicy::Filling))),
435 0 : )
436 0 : .set(scheduling_policy.eq(String::from(NodeSchedulingPolicy::Active)))
437 0 : .execute(conn)?;
438 :
439 0 : Ok(updated)
440 0 : })
441 : .await?;
442 :
443 : let mut result = HashMap::new();
444 : for tsp in updated {
445 : let tenant_shard_id = TenantShardId {
446 : tenant_id: TenantId::from_str(tsp.tenant_id.as_str())
447 0 : .map_err(|e| DatabaseError::Logical(format!("Malformed tenant id: {e}")))?,
448 : shard_number: ShardNumber(tsp.shard_number as u8),
449 : shard_count: ShardCount::new(tsp.shard_count as u8),
450 : };
451 :
452 : let Some(g) = tsp.generation else {
453 : // If the generation_pageserver column was non-NULL, then the generation column should also be non-NULL:
454 : // we only set generation_pageserver when setting generation.
455 : return Err(DatabaseError::Logical(
456 : "Generation should always be set after incrementing".to_string(),
457 : ));
458 : };
459 : result.insert(tenant_shard_id, Generation::new(g as u32));
460 : }
461 :
462 : Ok(result)
463 : }
464 :
465 : /// Reconciler calls this immediately before attaching to a new pageserver, to acquire a unique, monotonically
466 : /// advancing generation number. We also store the NodeId for which the generation was issued, so that in
467 : /// [`Self::re_attach`] we can do a bulk UPDATE on the generations for that node.
468 0 : pub(crate) async fn increment_generation(
469 0 : &self,
470 0 : tenant_shard_id: TenantShardId,
471 0 : node_id: NodeId,
472 0 : ) -> anyhow::Result<Generation> {
473 : use crate::schema::tenant_shards::dsl::*;
474 0 : let updated = self
475 0 : .with_measured_conn(DatabaseOperation::IncrementGeneration, move |conn| {
476 0 : let updated = diesel::update(tenant_shards)
477 0 : .filter(tenant_id.eq(tenant_shard_id.tenant_id.to_string()))
478 0 : .filter(shard_number.eq(tenant_shard_id.shard_number.0 as i32))
479 0 : .filter(shard_count.eq(tenant_shard_id.shard_count.literal() as i32))
480 0 : .set((
481 0 : generation.eq(generation + 1),
482 0 : generation_pageserver.eq(node_id.0 as i64),
483 0 : ))
484 0 : // TODO: only returning() the generation column
485 0 : .returning(TenantShardPersistence::as_returning())
486 0 : .get_result(conn)?;
487 :
488 0 : Ok(updated)
489 0 : })
490 0 : .await?;
491 :
492 : // Generation is always non-null in the rseult: if the generation column had been NULL, then we
493 : // should have experienced an SQL Confilict error while executing a query that tries to increment it.
494 0 : debug_assert!(updated.generation.is_some());
495 0 : let Some(g) = updated.generation else {
496 0 : return Err(DatabaseError::Logical(
497 0 : "Generation should always be set after incrementing".to_string(),
498 0 : )
499 0 : .into());
500 : };
501 :
502 0 : Ok(Generation::new(g as u32))
503 0 : }
504 :
505 : #[allow(non_local_definitions)]
506 : /// For use when updating a persistent property of a tenant, such as its config or placement_policy.
507 : ///
508 : /// Do not use this for settting generation, unless in the special onboarding code path (/location_config)
509 : /// API: use [`Self::increment_generation`] instead. Setting the generation via this route is a one-time thing
510 : /// that we only do the first time a tenant is set to an attached policy via /location_config.
511 0 : pub(crate) async fn update_tenant_shard(
512 0 : &self,
513 0 : tenant: TenantFilter,
514 0 : input_placement_policy: Option<PlacementPolicy>,
515 0 : input_config: Option<TenantConfig>,
516 0 : input_generation: Option<Generation>,
517 0 : input_scheduling_policy: Option<ShardSchedulingPolicy>,
518 0 : ) -> DatabaseResult<()> {
519 0 : use crate::schema::tenant_shards::dsl::*;
520 0 :
521 0 : self.with_measured_conn(DatabaseOperation::UpdateTenantShard, move |conn| {
522 0 : let query = match tenant {
523 0 : TenantFilter::Shard(tenant_shard_id) => diesel::update(tenant_shards)
524 0 : .filter(tenant_id.eq(tenant_shard_id.tenant_id.to_string()))
525 0 : .filter(shard_number.eq(tenant_shard_id.shard_number.0 as i32))
526 0 : .filter(shard_count.eq(tenant_shard_id.shard_count.literal() as i32))
527 0 : .into_boxed(),
528 0 : TenantFilter::Tenant(input_tenant_id) => diesel::update(tenant_shards)
529 0 : .filter(tenant_id.eq(input_tenant_id.to_string()))
530 0 : .into_boxed(),
531 : };
532 :
533 0 : #[derive(AsChangeset)]
534 : #[diesel(table_name = crate::schema::tenant_shards)]
535 : struct ShardUpdate {
536 : generation: Option<i32>,
537 : placement_policy: Option<String>,
538 : config: Option<String>,
539 : scheduling_policy: Option<String>,
540 : }
541 :
542 0 : let update = ShardUpdate {
543 0 : generation: input_generation.map(|g| g.into().unwrap() as i32),
544 0 : placement_policy: input_placement_policy
545 0 : .as_ref()
546 0 : .map(|p| serde_json::to_string(&p).unwrap()),
547 0 : config: input_config
548 0 : .as_ref()
549 0 : .map(|c| serde_json::to_string(&c).unwrap()),
550 0 : scheduling_policy: input_scheduling_policy
551 0 : .map(|p| serde_json::to_string(&p).unwrap()),
552 0 : };
553 0 :
554 0 : query.set(update).execute(conn)?;
555 :
556 0 : Ok(())
557 0 : })
558 0 : .await?;
559 :
560 0 : Ok(())
561 0 : }
562 :
563 0 : pub(crate) async fn detach(&self, tenant_shard_id: TenantShardId) -> anyhow::Result<()> {
564 0 : use crate::schema::tenant_shards::dsl::*;
565 0 : self.with_measured_conn(DatabaseOperation::Detach, move |conn| {
566 0 : let updated = diesel::update(tenant_shards)
567 0 : .filter(tenant_id.eq(tenant_shard_id.tenant_id.to_string()))
568 0 : .filter(shard_number.eq(tenant_shard_id.shard_number.0 as i32))
569 0 : .filter(shard_count.eq(tenant_shard_id.shard_count.literal() as i32))
570 0 : .set((
571 0 : generation_pageserver.eq(Option::<i64>::None),
572 0 : placement_policy.eq(serde_json::to_string(&PlacementPolicy::Detached).unwrap()),
573 0 : ))
574 0 : .execute(conn)?;
575 :
576 0 : Ok(updated)
577 0 : })
578 0 : .await?;
579 :
580 0 : Ok(())
581 0 : }
582 :
583 : // When we start shard splitting, we must durably mark the tenant so that
584 : // on restart, we know that we must go through recovery.
585 : //
586 : // We create the child shards here, so that they will be available for increment_generation calls
587 : // if some pageserver holding a child shard needs to restart before the overall tenant split is complete.
588 0 : pub(crate) async fn begin_shard_split(
589 0 : &self,
590 0 : old_shard_count: ShardCount,
591 0 : split_tenant_id: TenantId,
592 0 : parent_to_children: Vec<(TenantShardId, Vec<TenantShardPersistence>)>,
593 0 : ) -> DatabaseResult<()> {
594 0 : use crate::schema::tenant_shards::dsl::*;
595 0 : self.with_measured_conn(DatabaseOperation::BeginShardSplit, move |conn| -> DatabaseResult<()> {
596 : // Mark parent shards as splitting
597 :
598 0 : let updated = diesel::update(tenant_shards)
599 0 : .filter(tenant_id.eq(split_tenant_id.to_string()))
600 0 : .filter(shard_count.eq(old_shard_count.literal() as i32))
601 0 : .set((splitting.eq(1),))
602 0 : .execute(conn)?;
603 0 : if u8::try_from(updated)
604 0 : .map_err(|_| DatabaseError::Logical(
605 0 : format!("Overflow existing shard count {} while splitting", updated))
606 0 : )? != old_shard_count.count() {
607 : // Perhaps a deletion or another split raced with this attempt to split, mutating
608 : // the parent shards that we intend to split. In this case the split request should fail.
609 0 : return Err(DatabaseError::Logical(
610 0 : format!("Unexpected existing shard count {updated} when preparing tenant for split (expected {})", old_shard_count.count())
611 0 : ));
612 0 : }
613 0 :
614 0 : // FIXME: spurious clone to sidestep closure move rules
615 0 : let parent_to_children = parent_to_children.clone();
616 :
617 : // Insert child shards
618 0 : for (parent_shard_id, children) in parent_to_children {
619 0 : let mut parent = crate::schema::tenant_shards::table
620 0 : .filter(tenant_id.eq(parent_shard_id.tenant_id.to_string()))
621 0 : .filter(shard_number.eq(parent_shard_id.shard_number.0 as i32))
622 0 : .filter(shard_count.eq(parent_shard_id.shard_count.literal() as i32))
623 0 : .load::<TenantShardPersistence>(conn)?;
624 0 : let parent = if parent.len() != 1 {
625 0 : return Err(DatabaseError::Logical(format!(
626 0 : "Parent shard {parent_shard_id} not found"
627 0 : )));
628 : } else {
629 0 : parent.pop().unwrap()
630 : };
631 0 : for mut shard in children {
632 : // Carry the parent's generation into the child
633 0 : shard.generation = parent.generation;
634 0 :
635 0 : debug_assert!(shard.splitting == SplitState::Splitting);
636 0 : diesel::insert_into(tenant_shards)
637 0 : .values(shard)
638 0 : .execute(conn)?;
639 : }
640 : }
641 :
642 0 : Ok(())
643 0 : })
644 0 : .await
645 0 : }
646 :
647 : // When we finish shard splitting, we must atomically clean up the old shards
648 : // and insert the new shards, and clear the splitting marker.
649 0 : pub(crate) async fn complete_shard_split(
650 0 : &self,
651 0 : split_tenant_id: TenantId,
652 0 : old_shard_count: ShardCount,
653 0 : ) -> DatabaseResult<()> {
654 0 : use crate::schema::tenant_shards::dsl::*;
655 0 : self.with_measured_conn(
656 0 : DatabaseOperation::CompleteShardSplit,
657 0 : move |conn| -> DatabaseResult<()> {
658 0 : // Drop parent shards
659 0 : diesel::delete(tenant_shards)
660 0 : .filter(tenant_id.eq(split_tenant_id.to_string()))
661 0 : .filter(shard_count.eq(old_shard_count.literal() as i32))
662 0 : .execute(conn)?;
663 :
664 : // Clear sharding flag
665 0 : let updated = diesel::update(tenant_shards)
666 0 : .filter(tenant_id.eq(split_tenant_id.to_string()))
667 0 : .set((splitting.eq(0),))
668 0 : .execute(conn)?;
669 0 : debug_assert!(updated > 0);
670 :
671 0 : Ok(())
672 0 : },
673 0 : )
674 0 : .await
675 0 : }
676 :
677 : /// Used when the remote part of a shard split failed: we will revert the database state to have only
678 : /// the parent shards, with SplitState::Idle.
679 0 : pub(crate) async fn abort_shard_split(
680 0 : &self,
681 0 : split_tenant_id: TenantId,
682 0 : new_shard_count: ShardCount,
683 0 : ) -> DatabaseResult<AbortShardSplitStatus> {
684 0 : use crate::schema::tenant_shards::dsl::*;
685 0 : self.with_measured_conn(
686 0 : DatabaseOperation::AbortShardSplit,
687 0 : move |conn| -> DatabaseResult<AbortShardSplitStatus> {
688 : // Clear the splitting state on parent shards
689 0 : let updated = diesel::update(tenant_shards)
690 0 : .filter(tenant_id.eq(split_tenant_id.to_string()))
691 0 : .filter(shard_count.ne(new_shard_count.literal() as i32))
692 0 : .set((splitting.eq(0),))
693 0 : .execute(conn)?;
694 :
695 : // Parent shards are already gone: we cannot abort.
696 0 : if updated == 0 {
697 0 : return Ok(AbortShardSplitStatus::Complete);
698 0 : }
699 0 :
700 0 : // Sanity check: if parent shards were present, their cardinality should
701 0 : // be less than the number of child shards.
702 0 : if updated >= new_shard_count.count() as usize {
703 0 : return Err(DatabaseError::Logical(format!(
704 0 : "Unexpected parent shard count {updated} while aborting split to \
705 0 : count {new_shard_count:?} on tenant {split_tenant_id}"
706 0 : )));
707 0 : }
708 0 :
709 0 : // Erase child shards
710 0 : diesel::delete(tenant_shards)
711 0 : .filter(tenant_id.eq(split_tenant_id.to_string()))
712 0 : .filter(shard_count.eq(new_shard_count.literal() as i32))
713 0 : .execute(conn)?;
714 :
715 0 : Ok(AbortShardSplitStatus::Aborted)
716 0 : },
717 0 : )
718 0 : .await
719 0 : }
720 :
721 : /// Stores all the latest metadata health updates durably. Updates existing entry on conflict.
722 : ///
723 : /// **Correctness:** `metadata_health_updates` should all belong the tenant shards managed by the storage controller.
724 : #[allow(dead_code)]
725 0 : pub(crate) async fn update_metadata_health_records(
726 0 : &self,
727 0 : healthy_records: Vec<MetadataHealthPersistence>,
728 0 : unhealthy_records: Vec<MetadataHealthPersistence>,
729 0 : now: chrono::DateTime<chrono::Utc>,
730 0 : ) -> DatabaseResult<()> {
731 0 : use crate::schema::metadata_health::dsl::*;
732 0 :
733 0 : self.with_measured_conn(
734 0 : DatabaseOperation::UpdateMetadataHealth,
735 0 : move |conn| -> DatabaseResult<_> {
736 0 : diesel::insert_into(metadata_health)
737 0 : .values(&healthy_records)
738 0 : .on_conflict((tenant_id, shard_number, shard_count))
739 0 : .do_update()
740 0 : .set((healthy.eq(true), last_scrubbed_at.eq(now)))
741 0 : .execute(conn)?;
742 :
743 0 : diesel::insert_into(metadata_health)
744 0 : .values(&unhealthy_records)
745 0 : .on_conflict((tenant_id, shard_number, shard_count))
746 0 : .do_update()
747 0 : .set((healthy.eq(false), last_scrubbed_at.eq(now)))
748 0 : .execute(conn)?;
749 0 : Ok(())
750 0 : },
751 0 : )
752 0 : .await
753 0 : }
754 :
755 : /// Lists all the metadata health records.
756 : #[allow(dead_code)]
757 0 : pub(crate) async fn list_metadata_health_records(
758 0 : &self,
759 0 : ) -> DatabaseResult<Vec<MetadataHealthPersistence>> {
760 0 : self.with_measured_conn(
761 0 : DatabaseOperation::ListMetadataHealth,
762 0 : move |conn| -> DatabaseResult<_> {
763 0 : Ok(
764 0 : crate::schema::metadata_health::table
765 0 : .load::<MetadataHealthPersistence>(conn)?,
766 : )
767 0 : },
768 0 : )
769 0 : .await
770 0 : }
771 :
772 : /// Lists all the metadata health records that is unhealthy.
773 : #[allow(dead_code)]
774 0 : pub(crate) async fn list_unhealthy_metadata_health_records(
775 0 : &self,
776 0 : ) -> DatabaseResult<Vec<MetadataHealthPersistence>> {
777 0 : use crate::schema::metadata_health::dsl::*;
778 0 : self.with_measured_conn(
779 0 : DatabaseOperation::ListMetadataHealthUnhealthy,
780 0 : move |conn| -> DatabaseResult<_> {
781 0 : Ok(crate::schema::metadata_health::table
782 0 : .filter(healthy.eq(false))
783 0 : .load::<MetadataHealthPersistence>(conn)?)
784 0 : },
785 0 : )
786 0 : .await
787 0 : }
788 :
789 : /// Lists all the metadata health records that have not been updated since an `earlier` time.
790 : #[allow(dead_code)]
791 0 : pub(crate) async fn list_outdated_metadata_health_records(
792 0 : &self,
793 0 : earlier: chrono::DateTime<chrono::Utc>,
794 0 : ) -> DatabaseResult<Vec<MetadataHealthPersistence>> {
795 0 : use crate::schema::metadata_health::dsl::*;
796 0 :
797 0 : self.with_measured_conn(
798 0 : DatabaseOperation::ListMetadataHealthOutdated,
799 0 : move |conn| -> DatabaseResult<_> {
800 0 : let query = metadata_health.filter(last_scrubbed_at.lt(earlier));
801 0 : let res = query.load::<MetadataHealthPersistence>(conn)?;
802 :
803 0 : Ok(res)
804 0 : },
805 0 : )
806 0 : .await
807 0 : }
808 :
809 : /// Get the current entry from the `leader` table if one exists.
810 : /// It is an error for the table to contain more than one entry.
811 0 : pub(crate) async fn get_leader(&self) -> DatabaseResult<Option<ControllerPersistence>> {
812 0 : let mut leader: Vec<ControllerPersistence> = self
813 0 : .with_measured_conn(
814 0 : DatabaseOperation::GetLeader,
815 0 : move |conn| -> DatabaseResult<_> {
816 0 : Ok(crate::schema::controllers::table.load::<ControllerPersistence>(conn)?)
817 0 : },
818 0 : )
819 0 : .await?;
820 :
821 0 : if leader.len() > 1 {
822 0 : return Err(DatabaseError::Logical(format!(
823 0 : "More than one entry present in the leader table: {leader:?}"
824 0 : )));
825 0 : }
826 0 :
827 0 : Ok(leader.pop())
828 0 : }
829 :
830 : /// Update the new leader with compare-exchange semantics. If `prev` does not
831 : /// match the current leader entry, then the update is treated as a failure.
832 : /// When `prev` is not specified, the update is forced.
833 0 : pub(crate) async fn update_leader(
834 0 : &self,
835 0 : prev: Option<ControllerPersistence>,
836 0 : new: ControllerPersistence,
837 0 : ) -> DatabaseResult<()> {
838 : use crate::schema::controllers::dsl::*;
839 :
840 0 : let updated = self
841 0 : .with_measured_conn(
842 0 : DatabaseOperation::UpdateLeader,
843 0 : move |conn| -> DatabaseResult<usize> {
844 0 : let updated = match &prev {
845 0 : Some(prev) => diesel::update(controllers)
846 0 : .filter(address.eq(prev.address.clone()))
847 0 : .filter(started_at.eq(prev.started_at))
848 0 : .set((
849 0 : address.eq(new.address.clone()),
850 0 : started_at.eq(new.started_at),
851 0 : ))
852 0 : .execute(conn)?,
853 0 : None => diesel::insert_into(controllers)
854 0 : .values(new.clone())
855 0 : .execute(conn)?,
856 : };
857 :
858 0 : Ok(updated)
859 0 : },
860 0 : )
861 0 : .await?;
862 :
863 0 : if updated == 0 {
864 0 : return Err(DatabaseError::Logical(
865 0 : "Leader table update failed".to_string(),
866 0 : ));
867 0 : }
868 0 :
869 0 : Ok(())
870 0 : }
871 : }
872 :
873 : /// Parts of [`crate::tenant_shard::TenantShard`] that are stored durably
874 0 : #[derive(Queryable, Selectable, Insertable, Serialize, Deserialize, Clone, Eq, PartialEq)]
875 : #[diesel(table_name = crate::schema::tenant_shards)]
876 : pub(crate) struct TenantShardPersistence {
877 : #[serde(default)]
878 : pub(crate) tenant_id: String,
879 : #[serde(default)]
880 : pub(crate) shard_number: i32,
881 : #[serde(default)]
882 : pub(crate) shard_count: i32,
883 : #[serde(default)]
884 : pub(crate) shard_stripe_size: i32,
885 :
886 : // Latest generation number: next time we attach, increment this
887 : // and use the incremented number when attaching.
888 : //
889 : // Generation is only None when first onboarding a tenant, where it may
890 : // be in PlacementPolicy::Secondary and therefore have no valid generation state.
891 : pub(crate) generation: Option<i32>,
892 :
893 : // Currently attached pageserver
894 : #[serde(rename = "pageserver")]
895 : pub(crate) generation_pageserver: Option<i64>,
896 :
897 : #[serde(default)]
898 : pub(crate) placement_policy: String,
899 : #[serde(default)]
900 : pub(crate) splitting: SplitState,
901 : #[serde(default)]
902 : pub(crate) config: String,
903 : #[serde(default)]
904 : pub(crate) scheduling_policy: String,
905 : }
906 :
907 : impl TenantShardPersistence {
908 0 : pub(crate) fn get_shard_identity(&self) -> Result<ShardIdentity, ShardConfigError> {
909 0 : if self.shard_count == 0 {
910 0 : Ok(ShardIdentity::unsharded())
911 : } else {
912 0 : Ok(ShardIdentity::new(
913 0 : ShardNumber(self.shard_number as u8),
914 0 : ShardCount::new(self.shard_count as u8),
915 0 : ShardStripeSize(self.shard_stripe_size as u32),
916 0 : )?)
917 : }
918 0 : }
919 :
920 0 : pub(crate) fn get_tenant_shard_id(&self) -> Result<TenantShardId, hex::FromHexError> {
921 0 : Ok(TenantShardId {
922 0 : tenant_id: TenantId::from_str(self.tenant_id.as_str())?,
923 0 : shard_number: ShardNumber(self.shard_number as u8),
924 0 : shard_count: ShardCount::new(self.shard_count as u8),
925 : })
926 0 : }
927 : }
928 :
929 : /// Parts of [`crate::node::Node`] that are stored durably
930 0 : #[derive(Serialize, Deserialize, Queryable, Selectable, Insertable, Eq, PartialEq)]
931 : #[diesel(table_name = crate::schema::nodes)]
932 : pub(crate) struct NodePersistence {
933 : pub(crate) node_id: i64,
934 : pub(crate) scheduling_policy: String,
935 : pub(crate) listen_http_addr: String,
936 : pub(crate) listen_http_port: i32,
937 : pub(crate) listen_pg_addr: String,
938 : pub(crate) listen_pg_port: i32,
939 : }
940 :
941 : /// Tenant metadata health status that are stored durably.
942 0 : #[derive(Queryable, Selectable, Insertable, Serialize, Deserialize, Clone, Eq, PartialEq)]
943 : #[diesel(table_name = crate::schema::metadata_health)]
944 : pub(crate) struct MetadataHealthPersistence {
945 : #[serde(default)]
946 : pub(crate) tenant_id: String,
947 : #[serde(default)]
948 : pub(crate) shard_number: i32,
949 : #[serde(default)]
950 : pub(crate) shard_count: i32,
951 :
952 : pub(crate) healthy: bool,
953 : pub(crate) last_scrubbed_at: chrono::DateTime<chrono::Utc>,
954 : }
955 :
956 : impl MetadataHealthPersistence {
957 0 : pub fn new(
958 0 : tenant_shard_id: TenantShardId,
959 0 : healthy: bool,
960 0 : last_scrubbed_at: chrono::DateTime<chrono::Utc>,
961 0 : ) -> Self {
962 0 : let tenant_id = tenant_shard_id.tenant_id.to_string();
963 0 : let shard_number = tenant_shard_id.shard_number.0 as i32;
964 0 : let shard_count = tenant_shard_id.shard_count.literal() as i32;
965 0 :
966 0 : MetadataHealthPersistence {
967 0 : tenant_id,
968 0 : shard_number,
969 0 : shard_count,
970 0 : healthy,
971 0 : last_scrubbed_at,
972 0 : }
973 0 : }
974 :
975 : #[allow(dead_code)]
976 0 : pub(crate) fn get_tenant_shard_id(&self) -> Result<TenantShardId, hex::FromHexError> {
977 0 : Ok(TenantShardId {
978 0 : tenant_id: TenantId::from_str(self.tenant_id.as_str())?,
979 0 : shard_number: ShardNumber(self.shard_number as u8),
980 0 : shard_count: ShardCount::new(self.shard_count as u8),
981 : })
982 0 : }
983 : }
984 :
985 : impl From<MetadataHealthPersistence> for MetadataHealthRecord {
986 0 : fn from(value: MetadataHealthPersistence) -> Self {
987 0 : MetadataHealthRecord {
988 0 : tenant_shard_id: value
989 0 : .get_tenant_shard_id()
990 0 : .expect("stored tenant id should be valid"),
991 0 : healthy: value.healthy,
992 0 : last_scrubbed_at: value.last_scrubbed_at,
993 0 : }
994 0 : }
995 : }
996 :
997 : #[derive(
998 0 : Serialize, Deserialize, Queryable, Selectable, Insertable, Eq, PartialEq, Debug, Clone,
999 : )]
1000 : #[diesel(table_name = crate::schema::controllers)]
1001 : pub(crate) struct ControllerPersistence {
1002 : pub(crate) address: String,
1003 : pub(crate) started_at: chrono::DateTime<chrono::Utc>,
1004 : }
|
