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 itertools::Itertools;
12 : use pageserver_api::controller_api::AvailabilityZone;
13 : use pageserver_api::controller_api::MetadataHealthRecord;
14 : use pageserver_api::controller_api::SafekeeperDescribeResponse;
15 : use pageserver_api::controller_api::ShardSchedulingPolicy;
16 : use pageserver_api::controller_api::SkSchedulingPolicy;
17 : use pageserver_api::controller_api::{NodeSchedulingPolicy, PlacementPolicy};
18 : use pageserver_api::models::TenantConfig;
19 : use pageserver_api::shard::ShardConfigError;
20 : use pageserver_api::shard::ShardIdentity;
21 : use pageserver_api::shard::ShardStripeSize;
22 : use pageserver_api::shard::{ShardCount, ShardNumber, TenantShardId};
23 : use serde::{Deserialize, Serialize};
24 : use utils::generation::Generation;
25 : use utils::id::{NodeId, TenantId};
26 :
27 : use crate::metrics::{
28 : DatabaseQueryErrorLabelGroup, DatabaseQueryLatencyLabelGroup, METRICS_REGISTRY,
29 : };
30 : use crate::node::Node;
31 :
32 : use diesel_migrations::{embed_migrations, EmbeddedMigrations};
33 : const MIGRATIONS: EmbeddedMigrations = embed_migrations!("./migrations");
34 :
35 : /// ## What do we store?
36 : ///
37 : /// The storage controller service does not store most of its state durably.
38 : ///
39 : /// The essential things to store durably are:
40 : /// - generation numbers, as these must always advance monotonically to ensure data safety.
41 : /// - Tenant's PlacementPolicy and TenantConfig, as the source of truth for these is something external.
42 : /// - Node's scheduling policies, as the source of truth for these is something external.
43 : ///
44 : /// Other things we store durably as an implementation detail:
45 : /// - Node's host/port: this could be avoided it we made nodes emit a self-registering heartbeat,
46 : /// but it is operationally simpler to make this service the authority for which nodes
47 : /// it talks to.
48 : ///
49 : /// ## Performance/efficiency
50 : ///
51 : /// The storage controller service does not go via the database for most things: there are
52 : /// a couple of places where we must, and where efficiency matters:
53 : /// - Incrementing generation numbers: the Reconciler has to wait for this to complete
54 : /// before it can attach a tenant, so this acts as a bound on how fast things like
55 : /// failover can happen.
56 : /// - Pageserver re-attach: we will increment many shards' generations when this happens,
57 : /// so it is important to avoid e.g. issuing O(N) queries.
58 : ///
59 : /// Database calls relating to nodes have low performance requirements, as they are very rarely
60 : /// updated, and reads of nodes are always from memory, not the database. We only require that
61 : /// we can UPDATE a node's scheduling mode reasonably quickly to mark a bad node offline.
62 : pub struct Persistence {
63 : connection_pool: diesel::r2d2::Pool<diesel::r2d2::ConnectionManager<PgConnection>>,
64 : }
65 :
66 : /// Legacy format, for use in JSON compat objects in test environment
67 0 : #[derive(Serialize, Deserialize)]
68 : struct JsonPersistence {
69 : tenants: HashMap<TenantShardId, TenantShardPersistence>,
70 : }
71 :
72 : #[derive(thiserror::Error, Debug)]
73 : pub(crate) enum DatabaseError {
74 : #[error(transparent)]
75 : Query(#[from] diesel::result::Error),
76 : #[error(transparent)]
77 : Connection(#[from] diesel::result::ConnectionError),
78 : #[error(transparent)]
79 : ConnectionPool(#[from] r2d2::Error),
80 : #[error("Logical error: {0}")]
81 : Logical(String),
82 : #[error("Migration error: {0}")]
83 : Migration(String),
84 : }
85 :
86 : #[derive(measured::FixedCardinalityLabel, Copy, Clone)]
87 : pub(crate) enum DatabaseOperation {
88 : InsertNode,
89 : UpdateNode,
90 : DeleteNode,
91 : ListNodes,
92 : BeginShardSplit,
93 : CompleteShardSplit,
94 : AbortShardSplit,
95 : Detach,
96 : ReAttach,
97 : IncrementGeneration,
98 : TenantGenerations,
99 : ShardGenerations,
100 : ListTenantShards,
101 : LoadTenant,
102 : InsertTenantShards,
103 : UpdateTenantShard,
104 : DeleteTenant,
105 : UpdateTenantConfig,
106 : UpdateMetadataHealth,
107 : ListMetadataHealth,
108 : ListMetadataHealthUnhealthy,
109 : ListMetadataHealthOutdated,
110 : ListSafekeepers,
111 : GetLeader,
112 : UpdateLeader,
113 : SetPreferredAzs,
114 : }
115 :
116 : #[must_use]
117 : pub(crate) enum AbortShardSplitStatus {
118 : /// We aborted the split in the database by reverting to the parent shards
119 : Aborted,
120 : /// The split had already been persisted.
121 : Complete,
122 : }
123 :
124 : pub(crate) type DatabaseResult<T> = Result<T, DatabaseError>;
125 :
126 : /// Some methods can operate on either a whole tenant or a single shard
127 : pub(crate) enum TenantFilter {
128 : Tenant(TenantId),
129 : Shard(TenantShardId),
130 : }
131 :
132 : /// Represents the results of looking up generation+pageserver for the shards of a tenant
133 : pub(crate) struct ShardGenerationState {
134 : pub(crate) tenant_shard_id: TenantShardId,
135 : pub(crate) generation: Option<Generation>,
136 : pub(crate) generation_pageserver: Option<NodeId>,
137 : }
138 :
139 : impl Persistence {
140 : // The default postgres connection limit is 100. We use up to 99, to leave one free for a human admin under
141 : // normal circumstances. This assumes we have exclusive use of the database cluster to which we connect.
142 : pub const MAX_CONNECTIONS: u32 = 99;
143 :
144 : // We don't want to keep a lot of connections alive: close them down promptly if they aren't being used.
145 : const IDLE_CONNECTION_TIMEOUT: Duration = Duration::from_secs(10);
146 : const MAX_CONNECTION_LIFETIME: Duration = Duration::from_secs(60);
147 :
148 0 : pub fn new(database_url: String) -> Self {
149 0 : let manager = diesel::r2d2::ConnectionManager::<PgConnection>::new(database_url);
150 0 :
151 0 : // We will use a connection pool: this is primarily to _limit_ our connection count, rather than to optimize time
152 0 : // to execute queries (database queries are not generally on latency-sensitive paths).
153 0 : let connection_pool = diesel::r2d2::Pool::builder()
154 0 : .max_size(Self::MAX_CONNECTIONS)
155 0 : .max_lifetime(Some(Self::MAX_CONNECTION_LIFETIME))
156 0 : .idle_timeout(Some(Self::IDLE_CONNECTION_TIMEOUT))
157 0 : // Always keep at least one connection ready to go
158 0 : .min_idle(Some(1))
159 0 : .test_on_check_out(true)
160 0 : .build(manager)
161 0 : .expect("Could not build connection pool");
162 0 :
163 0 : Self { connection_pool }
164 0 : }
165 :
166 : /// A helper for use during startup, where we would like to tolerate concurrent restarts of the
167 : /// database and the storage controller, therefore the database might not be available right away
168 0 : pub async fn await_connection(
169 0 : database_url: &str,
170 0 : timeout: Duration,
171 0 : ) -> Result<(), diesel::ConnectionError> {
172 0 : let started_at = Instant::now();
173 : loop {
174 0 : match PgConnection::establish(database_url) {
175 : Ok(_) => {
176 0 : tracing::info!("Connected to database.");
177 0 : return Ok(());
178 : }
179 0 : Err(e) => {
180 0 : if started_at.elapsed() > timeout {
181 0 : return Err(e);
182 : } else {
183 0 : tracing::info!("Database not yet available, waiting... ({e})");
184 0 : tokio::time::sleep(Duration::from_millis(100)).await;
185 : }
186 : }
187 : }
188 : }
189 0 : }
190 :
191 : /// Execute the diesel migrations that are built into this binary
192 0 : pub(crate) async fn migration_run(&self) -> DatabaseResult<()> {
193 : use diesel_migrations::{HarnessWithOutput, MigrationHarness};
194 :
195 0 : self.with_conn(move |conn| -> DatabaseResult<()> {
196 0 : HarnessWithOutput::write_to_stdout(conn)
197 0 : .run_pending_migrations(MIGRATIONS)
198 0 : .map(|_| ())
199 0 : .map_err(|e| DatabaseError::Migration(e.to_string()))
200 0 : })
201 0 : .await
202 0 : }
203 :
204 : /// Wraps `with_conn` in order to collect latency and error metrics
205 0 : async fn with_measured_conn<F, R>(&self, op: DatabaseOperation, func: F) -> DatabaseResult<R>
206 0 : where
207 0 : F: Fn(&mut PgConnection) -> DatabaseResult<R> + Send + 'static,
208 0 : R: Send + 'static,
209 0 : {
210 0 : let latency = &METRICS_REGISTRY
211 0 : .metrics_group
212 0 : .storage_controller_database_query_latency;
213 0 : let _timer = latency.start_timer(DatabaseQueryLatencyLabelGroup { operation: op });
214 :
215 0 : let res = self.with_conn(func).await;
216 :
217 0 : if let Err(err) = &res {
218 0 : let error_counter = &METRICS_REGISTRY
219 0 : .metrics_group
220 0 : .storage_controller_database_query_error;
221 0 : error_counter.inc(DatabaseQueryErrorLabelGroup {
222 0 : error_type: err.error_label(),
223 0 : operation: op,
224 0 : })
225 0 : }
226 :
227 0 : res
228 0 : }
229 :
230 : /// Call the provided function in a tokio blocking thread, with a Diesel database connection.
231 0 : async fn with_conn<F, R>(&self, func: F) -> DatabaseResult<R>
232 0 : where
233 0 : F: Fn(&mut PgConnection) -> DatabaseResult<R> + Send + 'static,
234 0 : R: Send + 'static,
235 0 : {
236 : // A generous allowance for how many times we may retry serializable transactions
237 : // before giving up. This is not expected to be hit: it is a defensive measure in case we
238 : // somehow engineer a situation where duelling transactions might otherwise live-lock.
239 : const MAX_RETRIES: usize = 128;
240 :
241 0 : let mut conn = self.connection_pool.get()?;
242 0 : tokio::task::spawn_blocking(move || -> DatabaseResult<R> {
243 0 : let mut retry_count = 0;
244 : loop {
245 0 : match conn.build_transaction().serializable().run(|c| func(c)) {
246 0 : Ok(r) => break Ok(r),
247 : Err(
248 0 : err @ DatabaseError::Query(diesel::result::Error::DatabaseError(
249 0 : diesel::result::DatabaseErrorKind::SerializationFailure,
250 0 : _,
251 0 : )),
252 0 : ) => {
253 0 : retry_count += 1;
254 0 : if retry_count > MAX_RETRIES {
255 0 : tracing::error!(
256 0 : "Exceeded max retries on SerializationFailure errors: {err:?}"
257 : );
258 0 : break Err(err);
259 : } else {
260 : // Retry on serialization errors: these are expected, because even though our
261 : // transactions don't fight for the same rows, they will occasionally collide
262 : // on index pages (e.g. increment_generation for unrelated shards can collide)
263 0 : tracing::debug!(
264 0 : "Retrying transaction on serialization failure {err:?}"
265 : );
266 0 : continue;
267 : }
268 : }
269 0 : Err(e) => break Err(e),
270 : }
271 : }
272 0 : })
273 0 : .await
274 0 : .expect("Task panic")
275 0 : }
276 :
277 : /// When a node is first registered, persist it before using it for anything
278 0 : pub(crate) async fn insert_node(&self, node: &Node) -> DatabaseResult<()> {
279 0 : let np = node.to_persistent();
280 0 : self.with_measured_conn(
281 0 : DatabaseOperation::InsertNode,
282 0 : move |conn| -> DatabaseResult<()> {
283 0 : diesel::insert_into(crate::schema::nodes::table)
284 0 : .values(&np)
285 0 : .execute(conn)?;
286 0 : Ok(())
287 0 : },
288 0 : )
289 0 : .await
290 0 : }
291 :
292 : /// At startup, populate the list of nodes which our shards may be placed on
293 0 : pub(crate) async fn list_nodes(&self) -> DatabaseResult<Vec<NodePersistence>> {
294 0 : let nodes: Vec<NodePersistence> = self
295 0 : .with_measured_conn(
296 0 : DatabaseOperation::ListNodes,
297 0 : move |conn| -> DatabaseResult<_> {
298 0 : Ok(crate::schema::nodes::table.load::<NodePersistence>(conn)?)
299 0 : },
300 0 : )
301 0 : .await?;
302 :
303 0 : tracing::info!("list_nodes: loaded {} nodes", nodes.len());
304 :
305 0 : Ok(nodes)
306 0 : }
307 :
308 0 : pub(crate) async fn update_node(
309 0 : &self,
310 0 : input_node_id: NodeId,
311 0 : input_scheduling: NodeSchedulingPolicy,
312 0 : ) -> DatabaseResult<()> {
313 : use crate::schema::nodes::dsl::*;
314 0 : let updated = self
315 0 : .with_measured_conn(DatabaseOperation::UpdateNode, move |conn| {
316 0 : let updated = diesel::update(nodes)
317 0 : .filter(node_id.eq(input_node_id.0 as i64))
318 0 : .set((scheduling_policy.eq(String::from(input_scheduling)),))
319 0 : .execute(conn)?;
320 0 : Ok(updated)
321 0 : })
322 0 : .await?;
323 :
324 0 : if updated != 1 {
325 0 : Err(DatabaseError::Logical(format!(
326 0 : "Node {node_id:?} not found for update",
327 0 : )))
328 : } else {
329 0 : Ok(())
330 : }
331 0 : }
332 :
333 : /// At startup, load the high level state for shards, such as their config + policy. This will
334 : /// be enriched at runtime with state discovered on pageservers.
335 : ///
336 : /// We exclude shards configured to be detached. During startup, if we see any attached locations
337 : /// for such shards, they will automatically be detached as 'orphans'.
338 0 : pub(crate) async fn load_active_tenant_shards(
339 0 : &self,
340 0 : ) -> DatabaseResult<Vec<TenantShardPersistence>> {
341 : use crate::schema::tenant_shards::dsl::*;
342 0 : self.with_measured_conn(
343 0 : DatabaseOperation::ListTenantShards,
344 0 : move |conn| -> DatabaseResult<_> {
345 0 : let query = tenant_shards.filter(
346 0 : placement_policy.ne(serde_json::to_string(&PlacementPolicy::Detached).unwrap()),
347 0 : );
348 0 : let result = query.load::<TenantShardPersistence>(conn)?;
349 :
350 0 : Ok(result)
351 0 : },
352 0 : )
353 0 : .await
354 0 : }
355 :
356 : /// When restoring a previously detached tenant into memory, load it from the database
357 0 : pub(crate) async fn load_tenant(
358 0 : &self,
359 0 : filter_tenant_id: TenantId,
360 0 : ) -> DatabaseResult<Vec<TenantShardPersistence>> {
361 : use crate::schema::tenant_shards::dsl::*;
362 0 : self.with_measured_conn(
363 0 : DatabaseOperation::LoadTenant,
364 0 : move |conn| -> DatabaseResult<_> {
365 0 : let query = tenant_shards.filter(tenant_id.eq(filter_tenant_id.to_string()));
366 0 : let result = query.load::<TenantShardPersistence>(conn)?;
367 :
368 0 : Ok(result)
369 0 : },
370 0 : )
371 0 : .await
372 0 : }
373 :
374 : /// Tenants must be persisted before we schedule them for the first time. This enables us
375 : /// to correctly retain generation monotonicity, and the externally provided placement policy & config.
376 0 : pub(crate) async fn insert_tenant_shards(
377 0 : &self,
378 0 : shards: Vec<TenantShardPersistence>,
379 0 : ) -> DatabaseResult<()> {
380 : use crate::schema::metadata_health;
381 : use crate::schema::tenant_shards;
382 :
383 0 : let now = chrono::Utc::now();
384 0 :
385 0 : let metadata_health_records = shards
386 0 : .iter()
387 0 : .map(|t| MetadataHealthPersistence {
388 0 : tenant_id: t.tenant_id.clone(),
389 0 : shard_number: t.shard_number,
390 0 : shard_count: t.shard_count,
391 0 : healthy: true,
392 0 : last_scrubbed_at: now,
393 0 : })
394 0 : .collect::<Vec<_>>();
395 0 :
396 0 : self.with_measured_conn(
397 0 : DatabaseOperation::InsertTenantShards,
398 0 : move |conn| -> DatabaseResult<()> {
399 0 : diesel::insert_into(tenant_shards::table)
400 0 : .values(&shards)
401 0 : .execute(conn)?;
402 :
403 0 : diesel::insert_into(metadata_health::table)
404 0 : .values(&metadata_health_records)
405 0 : .execute(conn)?;
406 0 : Ok(())
407 0 : },
408 0 : )
409 0 : .await
410 0 : }
411 :
412 : /// Ordering: call this _after_ deleting the tenant on pageservers, but _before_ dropping state for
413 : /// the tenant from memory on this server.
414 0 : pub(crate) async fn delete_tenant(&self, del_tenant_id: TenantId) -> DatabaseResult<()> {
415 : use crate::schema::tenant_shards::dsl::*;
416 0 : self.with_measured_conn(
417 0 : DatabaseOperation::DeleteTenant,
418 0 : move |conn| -> DatabaseResult<()> {
419 0 : // `metadata_health` status (if exists) is also deleted based on the cascade behavior.
420 0 : diesel::delete(tenant_shards)
421 0 : .filter(tenant_id.eq(del_tenant_id.to_string()))
422 0 : .execute(conn)?;
423 0 : Ok(())
424 0 : },
425 0 : )
426 0 : .await
427 0 : }
428 :
429 0 : pub(crate) async fn delete_node(&self, del_node_id: NodeId) -> DatabaseResult<()> {
430 : use crate::schema::nodes::dsl::*;
431 0 : self.with_measured_conn(
432 0 : DatabaseOperation::DeleteNode,
433 0 : move |conn| -> DatabaseResult<()> {
434 0 : diesel::delete(nodes)
435 0 : .filter(node_id.eq(del_node_id.0 as i64))
436 0 : .execute(conn)?;
437 :
438 0 : Ok(())
439 0 : },
440 0 : )
441 0 : .await
442 0 : }
443 :
444 : /// When a tenant invokes the /re-attach API, this function is responsible for doing an efficient
445 : /// batched increment of the generations of all tenants whose generation_pageserver is equal to
446 : /// the node that called /re-attach.
447 : #[tracing::instrument(skip_all, fields(node_id))]
448 : pub(crate) async fn re_attach(
449 : &self,
450 : input_node_id: NodeId,
451 : ) -> DatabaseResult<HashMap<TenantShardId, Generation>> {
452 : use crate::schema::nodes::dsl::scheduling_policy;
453 : use crate::schema::nodes::dsl::*;
454 : use crate::schema::tenant_shards::dsl::*;
455 : let updated = self
456 0 : .with_measured_conn(DatabaseOperation::ReAttach, move |conn| {
457 0 : let rows_updated = diesel::update(tenant_shards)
458 0 : .filter(generation_pageserver.eq(input_node_id.0 as i64))
459 0 : .set(generation.eq(generation + 1))
460 0 : .execute(conn)?;
461 :
462 0 : tracing::info!("Incremented {} tenants' generations", rows_updated);
463 :
464 : // TODO: UPDATE+SELECT in one query
465 :
466 0 : let updated = tenant_shards
467 0 : .filter(generation_pageserver.eq(input_node_id.0 as i64))
468 0 : .select(TenantShardPersistence::as_select())
469 0 : .load(conn)?;
470 :
471 : // If the node went through a drain and restart phase before re-attaching,
472 : // then reset it's node scheduling policy to active.
473 0 : diesel::update(nodes)
474 0 : .filter(node_id.eq(input_node_id.0 as i64))
475 0 : .filter(
476 0 : scheduling_policy
477 0 : .eq(String::from(NodeSchedulingPolicy::PauseForRestart))
478 0 : .or(scheduling_policy.eq(String::from(NodeSchedulingPolicy::Draining)))
479 0 : .or(scheduling_policy.eq(String::from(NodeSchedulingPolicy::Filling))),
480 0 : )
481 0 : .set(scheduling_policy.eq(String::from(NodeSchedulingPolicy::Active)))
482 0 : .execute(conn)?;
483 :
484 0 : Ok(updated)
485 0 : })
486 : .await?;
487 :
488 : let mut result = HashMap::new();
489 : for tsp in updated {
490 : let tenant_shard_id = TenantShardId {
491 : tenant_id: TenantId::from_str(tsp.tenant_id.as_str())
492 0 : .map_err(|e| DatabaseError::Logical(format!("Malformed tenant id: {e}")))?,
493 : shard_number: ShardNumber(tsp.shard_number as u8),
494 : shard_count: ShardCount::new(tsp.shard_count as u8),
495 : };
496 :
497 : let Some(g) = tsp.generation else {
498 : // If the generation_pageserver column was non-NULL, then the generation column should also be non-NULL:
499 : // we only set generation_pageserver when setting generation.
500 : return Err(DatabaseError::Logical(
501 : "Generation should always be set after incrementing".to_string(),
502 : ));
503 : };
504 : result.insert(tenant_shard_id, Generation::new(g as u32));
505 : }
506 :
507 : Ok(result)
508 : }
509 :
510 : /// Reconciler calls this immediately before attaching to a new pageserver, to acquire a unique, monotonically
511 : /// advancing generation number. We also store the NodeId for which the generation was issued, so that in
512 : /// [`Self::re_attach`] we can do a bulk UPDATE on the generations for that node.
513 0 : pub(crate) async fn increment_generation(
514 0 : &self,
515 0 : tenant_shard_id: TenantShardId,
516 0 : node_id: NodeId,
517 0 : ) -> anyhow::Result<Generation> {
518 : use crate::schema::tenant_shards::dsl::*;
519 0 : let updated = self
520 0 : .with_measured_conn(DatabaseOperation::IncrementGeneration, move |conn| {
521 0 : let updated = diesel::update(tenant_shards)
522 0 : .filter(tenant_id.eq(tenant_shard_id.tenant_id.to_string()))
523 0 : .filter(shard_number.eq(tenant_shard_id.shard_number.0 as i32))
524 0 : .filter(shard_count.eq(tenant_shard_id.shard_count.literal() as i32))
525 0 : .set((
526 0 : generation.eq(generation + 1),
527 0 : generation_pageserver.eq(node_id.0 as i64),
528 0 : ))
529 0 : // TODO: only returning() the generation column
530 0 : .returning(TenantShardPersistence::as_returning())
531 0 : .get_result(conn)?;
532 :
533 0 : Ok(updated)
534 0 : })
535 0 : .await?;
536 :
537 : // Generation is always non-null in the rseult: if the generation column had been NULL, then we
538 : // should have experienced an SQL Confilict error while executing a query that tries to increment it.
539 0 : debug_assert!(updated.generation.is_some());
540 0 : let Some(g) = updated.generation else {
541 0 : return Err(DatabaseError::Logical(
542 0 : "Generation should always be set after incrementing".to_string(),
543 0 : )
544 0 : .into());
545 : };
546 :
547 0 : Ok(Generation::new(g as u32))
548 0 : }
549 :
550 : /// When we want to call out to the running shards for a tenant, e.g. during timeline CRUD operations,
551 : /// we need to know where the shard is attached, _and_ the generation, so that we can re-check the generation
552 : /// afterwards to confirm that our timeline CRUD operation is truly persistent (it must have happened in the
553 : /// latest generation)
554 : ///
555 : /// If the tenant doesn't exist, an empty vector is returned.
556 : ///
557 : /// Output is sorted by shard number
558 0 : pub(crate) async fn tenant_generations(
559 0 : &self,
560 0 : filter_tenant_id: TenantId,
561 0 : ) -> Result<Vec<ShardGenerationState>, DatabaseError> {
562 : use crate::schema::tenant_shards::dsl::*;
563 0 : let rows = self
564 0 : .with_measured_conn(DatabaseOperation::TenantGenerations, move |conn| {
565 0 : let result = tenant_shards
566 0 : .filter(tenant_id.eq(filter_tenant_id.to_string()))
567 0 : .select(TenantShardPersistence::as_select())
568 0 : .order(shard_number)
569 0 : .load(conn)?;
570 0 : Ok(result)
571 0 : })
572 0 : .await?;
573 :
574 0 : Ok(rows
575 0 : .into_iter()
576 0 : .map(|p| ShardGenerationState {
577 0 : tenant_shard_id: p
578 0 : .get_tenant_shard_id()
579 0 : .expect("Corrupt tenant shard id in database"),
580 0 : generation: p.generation.map(|g| Generation::new(g as u32)),
581 0 : generation_pageserver: p.generation_pageserver.map(|n| NodeId(n as u64)),
582 0 : })
583 0 : .collect())
584 0 : }
585 :
586 : /// Read the generation number of specific tenant shards
587 : ///
588 : /// Output is unsorted. Output may not include values for all inputs, if they are missing in the database.
589 0 : pub(crate) async fn shard_generations(
590 0 : &self,
591 0 : mut tenant_shard_ids: impl Iterator<Item = &TenantShardId>,
592 0 : ) -> Result<Vec<(TenantShardId, Option<Generation>)>, DatabaseError> {
593 0 : let mut rows = Vec::with_capacity(tenant_shard_ids.size_hint().0);
594 :
595 : // We will chunk our input to avoid composing arbitrarily long `IN` clauses. Typically we are
596 : // called with a single digit number of IDs, but in principle we could be called with tens
597 : // of thousands (all the shards on one pageserver) from the generation validation API.
598 0 : loop {
599 0 : // A modest hardcoded chunk size to handle typical cases in a single query but never generate particularly
600 0 : // large query strings.
601 0 : let chunk_ids = tenant_shard_ids.by_ref().take(32);
602 0 :
603 0 : // Compose a comma separated list of tuples for matching on (tenant_id, shard_number, shard_count)
604 0 : let in_clause = chunk_ids
605 0 : .map(|tsid| {
606 0 : format!(
607 0 : "('{}', {}, {})",
608 0 : tsid.tenant_id, tsid.shard_number.0, tsid.shard_count.0
609 0 : )
610 0 : })
611 0 : .join(",");
612 0 :
613 0 : // We are done when our iterator gives us nothing to filter on
614 0 : if in_clause.is_empty() {
615 0 : break;
616 0 : }
617 :
618 0 : let chunk_rows = self
619 0 : .with_measured_conn(DatabaseOperation::ShardGenerations, move |conn| {
620 : // diesel doesn't support multi-column IN queries, so we compose raw SQL. No escaping is required because
621 : // the inputs are strongly typed and cannot carry any user-supplied raw string content.
622 0 : let result : Vec<TenantShardPersistence> = diesel::sql_query(
623 0 : format!("SELECT * from tenant_shards where (tenant_id, shard_number, shard_count) in ({in_clause});").as_str()
624 0 : ).load(conn)?;
625 :
626 0 : Ok(result)
627 0 : })
628 0 : .await?;
629 0 : rows.extend(chunk_rows.into_iter())
630 : }
631 :
632 0 : Ok(rows
633 0 : .into_iter()
634 0 : .map(|tsp| {
635 0 : (
636 0 : tsp.get_tenant_shard_id()
637 0 : .expect("Bad tenant ID in database"),
638 0 : tsp.generation.map(|g| Generation::new(g as u32)),
639 0 : )
640 0 : })
641 0 : .collect())
642 0 : }
643 :
644 : #[allow(non_local_definitions)]
645 : /// For use when updating a persistent property of a tenant, such as its config or placement_policy.
646 : ///
647 : /// Do not use this for settting generation, unless in the special onboarding code path (/location_config)
648 : /// API: use [`Self::increment_generation`] instead. Setting the generation via this route is a one-time thing
649 : /// that we only do the first time a tenant is set to an attached policy via /location_config.
650 0 : pub(crate) async fn update_tenant_shard(
651 0 : &self,
652 0 : tenant: TenantFilter,
653 0 : input_placement_policy: Option<PlacementPolicy>,
654 0 : input_config: Option<TenantConfig>,
655 0 : input_generation: Option<Generation>,
656 0 : input_scheduling_policy: Option<ShardSchedulingPolicy>,
657 0 : ) -> DatabaseResult<()> {
658 : use crate::schema::tenant_shards::dsl::*;
659 :
660 0 : self.with_measured_conn(DatabaseOperation::UpdateTenantShard, move |conn| {
661 0 : let query = match tenant {
662 0 : TenantFilter::Shard(tenant_shard_id) => diesel::update(tenant_shards)
663 0 : .filter(tenant_id.eq(tenant_shard_id.tenant_id.to_string()))
664 0 : .filter(shard_number.eq(tenant_shard_id.shard_number.0 as i32))
665 0 : .filter(shard_count.eq(tenant_shard_id.shard_count.literal() as i32))
666 0 : .into_boxed(),
667 0 : TenantFilter::Tenant(input_tenant_id) => diesel::update(tenant_shards)
668 0 : .filter(tenant_id.eq(input_tenant_id.to_string()))
669 0 : .into_boxed(),
670 : };
671 :
672 : // Clear generation_pageserver if we are moving into a state where we won't have
673 : // any attached pageservers.
674 0 : let input_generation_pageserver = match input_placement_policy {
675 0 : None | Some(PlacementPolicy::Attached(_)) => None,
676 0 : Some(PlacementPolicy::Detached | PlacementPolicy::Secondary) => Some(None),
677 : };
678 :
679 0 : #[derive(AsChangeset)]
680 : #[diesel(table_name = crate::schema::tenant_shards)]
681 : struct ShardUpdate {
682 : generation: Option<i32>,
683 : placement_policy: Option<String>,
684 : config: Option<String>,
685 : scheduling_policy: Option<String>,
686 : generation_pageserver: Option<Option<i64>>,
687 : }
688 :
689 0 : let update = ShardUpdate {
690 0 : generation: input_generation.map(|g| g.into().unwrap() as i32),
691 0 : placement_policy: input_placement_policy
692 0 : .as_ref()
693 0 : .map(|p| serde_json::to_string(&p).unwrap()),
694 0 : config: input_config
695 0 : .as_ref()
696 0 : .map(|c| serde_json::to_string(&c).unwrap()),
697 0 : scheduling_policy: input_scheduling_policy
698 0 : .map(|p| serde_json::to_string(&p).unwrap()),
699 0 : generation_pageserver: input_generation_pageserver,
700 0 : };
701 0 :
702 0 : query.set(update).execute(conn)?;
703 :
704 0 : Ok(())
705 0 : })
706 0 : .await?;
707 :
708 0 : Ok(())
709 0 : }
710 :
711 0 : pub(crate) async fn set_tenant_shard_preferred_azs(
712 0 : &self,
713 0 : preferred_azs: Vec<(TenantShardId, AvailabilityZone)>,
714 0 : ) -> DatabaseResult<Vec<(TenantShardId, AvailabilityZone)>> {
715 : use crate::schema::tenant_shards::dsl::*;
716 :
717 0 : self.with_measured_conn(DatabaseOperation::SetPreferredAzs, move |conn| {
718 0 : let mut shards_updated = Vec::default();
719 :
720 0 : for (tenant_shard_id, preferred_az) in preferred_azs.iter() {
721 0 : let updated = diesel::update(tenant_shards)
722 0 : .filter(tenant_id.eq(tenant_shard_id.tenant_id.to_string()))
723 0 : .filter(shard_number.eq(tenant_shard_id.shard_number.0 as i32))
724 0 : .filter(shard_count.eq(tenant_shard_id.shard_count.literal() as i32))
725 0 : .set(preferred_az_id.eq(preferred_az.0.clone()))
726 0 : .execute(conn)?;
727 :
728 0 : if updated == 1 {
729 0 : shards_updated.push((*tenant_shard_id, preferred_az.clone()));
730 0 : }
731 : }
732 :
733 0 : Ok(shards_updated)
734 0 : })
735 0 : .await
736 0 : }
737 :
738 0 : pub(crate) async fn detach(&self, tenant_shard_id: TenantShardId) -> anyhow::Result<()> {
739 : use crate::schema::tenant_shards::dsl::*;
740 0 : self.with_measured_conn(DatabaseOperation::Detach, move |conn| {
741 0 : let updated = diesel::update(tenant_shards)
742 0 : .filter(tenant_id.eq(tenant_shard_id.tenant_id.to_string()))
743 0 : .filter(shard_number.eq(tenant_shard_id.shard_number.0 as i32))
744 0 : .filter(shard_count.eq(tenant_shard_id.shard_count.literal() as i32))
745 0 : .set((
746 0 : generation_pageserver.eq(Option::<i64>::None),
747 0 : placement_policy.eq(serde_json::to_string(&PlacementPolicy::Detached).unwrap()),
748 0 : ))
749 0 : .execute(conn)?;
750 :
751 0 : Ok(updated)
752 0 : })
753 0 : .await?;
754 :
755 0 : Ok(())
756 0 : }
757 :
758 : // When we start shard splitting, we must durably mark the tenant so that
759 : // on restart, we know that we must go through recovery.
760 : //
761 : // We create the child shards here, so that they will be available for increment_generation calls
762 : // if some pageserver holding a child shard needs to restart before the overall tenant split is complete.
763 0 : pub(crate) async fn begin_shard_split(
764 0 : &self,
765 0 : old_shard_count: ShardCount,
766 0 : split_tenant_id: TenantId,
767 0 : parent_to_children: Vec<(TenantShardId, Vec<TenantShardPersistence>)>,
768 0 : ) -> DatabaseResult<()> {
769 : use crate::schema::tenant_shards::dsl::*;
770 0 : self.with_measured_conn(DatabaseOperation::BeginShardSplit, move |conn| -> DatabaseResult<()> {
771 : // Mark parent shards as splitting
772 :
773 0 : let updated = diesel::update(tenant_shards)
774 0 : .filter(tenant_id.eq(split_tenant_id.to_string()))
775 0 : .filter(shard_count.eq(old_shard_count.literal() as i32))
776 0 : .set((splitting.eq(1),))
777 0 : .execute(conn)?;
778 0 : if u8::try_from(updated)
779 0 : .map_err(|_| DatabaseError::Logical(
780 0 : format!("Overflow existing shard count {} while splitting", updated))
781 0 : )? != old_shard_count.count() {
782 : // Perhaps a deletion or another split raced with this attempt to split, mutating
783 : // the parent shards that we intend to split. In this case the split request should fail.
784 0 : return Err(DatabaseError::Logical(
785 0 : format!("Unexpected existing shard count {updated} when preparing tenant for split (expected {})", old_shard_count.count())
786 0 : ));
787 0 : }
788 0 :
789 0 : // FIXME: spurious clone to sidestep closure move rules
790 0 : let parent_to_children = parent_to_children.clone();
791 :
792 : // Insert child shards
793 0 : for (parent_shard_id, children) in parent_to_children {
794 0 : let mut parent = crate::schema::tenant_shards::table
795 0 : .filter(tenant_id.eq(parent_shard_id.tenant_id.to_string()))
796 0 : .filter(shard_number.eq(parent_shard_id.shard_number.0 as i32))
797 0 : .filter(shard_count.eq(parent_shard_id.shard_count.literal() as i32))
798 0 : .load::<TenantShardPersistence>(conn)?;
799 0 : let parent = if parent.len() != 1 {
800 0 : return Err(DatabaseError::Logical(format!(
801 0 : "Parent shard {parent_shard_id} not found"
802 0 : )));
803 : } else {
804 0 : parent.pop().unwrap()
805 : };
806 0 : for mut shard in children {
807 : // Carry the parent's generation into the child
808 0 : shard.generation = parent.generation;
809 0 :
810 0 : debug_assert!(shard.splitting == SplitState::Splitting);
811 0 : diesel::insert_into(tenant_shards)
812 0 : .values(shard)
813 0 : .execute(conn)?;
814 : }
815 : }
816 :
817 0 : Ok(())
818 0 : })
819 0 : .await
820 0 : }
821 :
822 : // When we finish shard splitting, we must atomically clean up the old shards
823 : // and insert the new shards, and clear the splitting marker.
824 0 : pub(crate) async fn complete_shard_split(
825 0 : &self,
826 0 : split_tenant_id: TenantId,
827 0 : old_shard_count: ShardCount,
828 0 : ) -> DatabaseResult<()> {
829 : use crate::schema::tenant_shards::dsl::*;
830 0 : self.with_measured_conn(
831 0 : DatabaseOperation::CompleteShardSplit,
832 0 : move |conn| -> DatabaseResult<()> {
833 0 : // Drop parent shards
834 0 : diesel::delete(tenant_shards)
835 0 : .filter(tenant_id.eq(split_tenant_id.to_string()))
836 0 : .filter(shard_count.eq(old_shard_count.literal() as i32))
837 0 : .execute(conn)?;
838 :
839 : // Clear sharding flag
840 0 : let updated = diesel::update(tenant_shards)
841 0 : .filter(tenant_id.eq(split_tenant_id.to_string()))
842 0 : .set((splitting.eq(0),))
843 0 : .execute(conn)?;
844 0 : debug_assert!(updated > 0);
845 :
846 0 : Ok(())
847 0 : },
848 0 : )
849 0 : .await
850 0 : }
851 :
852 : /// Used when the remote part of a shard split failed: we will revert the database state to have only
853 : /// the parent shards, with SplitState::Idle.
854 0 : pub(crate) async fn abort_shard_split(
855 0 : &self,
856 0 : split_tenant_id: TenantId,
857 0 : new_shard_count: ShardCount,
858 0 : ) -> DatabaseResult<AbortShardSplitStatus> {
859 : use crate::schema::tenant_shards::dsl::*;
860 0 : self.with_measured_conn(
861 0 : DatabaseOperation::AbortShardSplit,
862 0 : move |conn| -> DatabaseResult<AbortShardSplitStatus> {
863 : // Clear the splitting state on parent shards
864 0 : let updated = diesel::update(tenant_shards)
865 0 : .filter(tenant_id.eq(split_tenant_id.to_string()))
866 0 : .filter(shard_count.ne(new_shard_count.literal() as i32))
867 0 : .set((splitting.eq(0),))
868 0 : .execute(conn)?;
869 :
870 : // Parent shards are already gone: we cannot abort.
871 0 : if updated == 0 {
872 0 : return Ok(AbortShardSplitStatus::Complete);
873 0 : }
874 0 :
875 0 : // Sanity check: if parent shards were present, their cardinality should
876 0 : // be less than the number of child shards.
877 0 : if updated >= new_shard_count.count() as usize {
878 0 : return Err(DatabaseError::Logical(format!(
879 0 : "Unexpected parent shard count {updated} while aborting split to \
880 0 : count {new_shard_count:?} on tenant {split_tenant_id}"
881 0 : )));
882 0 : }
883 0 :
884 0 : // Erase child shards
885 0 : diesel::delete(tenant_shards)
886 0 : .filter(tenant_id.eq(split_tenant_id.to_string()))
887 0 : .filter(shard_count.eq(new_shard_count.literal() as i32))
888 0 : .execute(conn)?;
889 :
890 0 : Ok(AbortShardSplitStatus::Aborted)
891 0 : },
892 0 : )
893 0 : .await
894 0 : }
895 :
896 : /// Stores all the latest metadata health updates durably. Updates existing entry on conflict.
897 : ///
898 : /// **Correctness:** `metadata_health_updates` should all belong the tenant shards managed by the storage controller.
899 : #[allow(dead_code)]
900 0 : pub(crate) async fn update_metadata_health_records(
901 0 : &self,
902 0 : healthy_records: Vec<MetadataHealthPersistence>,
903 0 : unhealthy_records: Vec<MetadataHealthPersistence>,
904 0 : now: chrono::DateTime<chrono::Utc>,
905 0 : ) -> DatabaseResult<()> {
906 : use crate::schema::metadata_health::dsl::*;
907 :
908 0 : self.with_measured_conn(
909 0 : DatabaseOperation::UpdateMetadataHealth,
910 0 : move |conn| -> DatabaseResult<_> {
911 0 : diesel::insert_into(metadata_health)
912 0 : .values(&healthy_records)
913 0 : .on_conflict((tenant_id, shard_number, shard_count))
914 0 : .do_update()
915 0 : .set((healthy.eq(true), last_scrubbed_at.eq(now)))
916 0 : .execute(conn)?;
917 :
918 0 : diesel::insert_into(metadata_health)
919 0 : .values(&unhealthy_records)
920 0 : .on_conflict((tenant_id, shard_number, shard_count))
921 0 : .do_update()
922 0 : .set((healthy.eq(false), last_scrubbed_at.eq(now)))
923 0 : .execute(conn)?;
924 0 : Ok(())
925 0 : },
926 0 : )
927 0 : .await
928 0 : }
929 :
930 : /// Lists all the metadata health records.
931 : #[allow(dead_code)]
932 0 : pub(crate) async fn list_metadata_health_records(
933 0 : &self,
934 0 : ) -> DatabaseResult<Vec<MetadataHealthPersistence>> {
935 0 : self.with_measured_conn(
936 0 : DatabaseOperation::ListMetadataHealth,
937 0 : move |conn| -> DatabaseResult<_> {
938 0 : Ok(
939 0 : crate::schema::metadata_health::table
940 0 : .load::<MetadataHealthPersistence>(conn)?,
941 : )
942 0 : },
943 0 : )
944 0 : .await
945 0 : }
946 :
947 : /// Lists all the metadata health records that is unhealthy.
948 : #[allow(dead_code)]
949 0 : pub(crate) async fn list_unhealthy_metadata_health_records(
950 0 : &self,
951 0 : ) -> DatabaseResult<Vec<MetadataHealthPersistence>> {
952 : use crate::schema::metadata_health::dsl::*;
953 0 : self.with_measured_conn(
954 0 : DatabaseOperation::ListMetadataHealthUnhealthy,
955 0 : move |conn| -> DatabaseResult<_> {
956 0 : Ok(crate::schema::metadata_health::table
957 0 : .filter(healthy.eq(false))
958 0 : .load::<MetadataHealthPersistence>(conn)?)
959 0 : },
960 0 : )
961 0 : .await
962 0 : }
963 :
964 : /// Lists all the metadata health records that have not been updated since an `earlier` time.
965 : #[allow(dead_code)]
966 0 : pub(crate) async fn list_outdated_metadata_health_records(
967 0 : &self,
968 0 : earlier: chrono::DateTime<chrono::Utc>,
969 0 : ) -> DatabaseResult<Vec<MetadataHealthPersistence>> {
970 : use crate::schema::metadata_health::dsl::*;
971 :
972 0 : self.with_measured_conn(
973 0 : DatabaseOperation::ListMetadataHealthOutdated,
974 0 : move |conn| -> DatabaseResult<_> {
975 0 : let query = metadata_health.filter(last_scrubbed_at.lt(earlier));
976 0 : let res = query.load::<MetadataHealthPersistence>(conn)?;
977 :
978 0 : Ok(res)
979 0 : },
980 0 : )
981 0 : .await
982 0 : }
983 :
984 : /// Get the current entry from the `leader` table if one exists.
985 : /// It is an error for the table to contain more than one entry.
986 0 : pub(crate) async fn get_leader(&self) -> DatabaseResult<Option<ControllerPersistence>> {
987 0 : let mut leader: Vec<ControllerPersistence> = self
988 0 : .with_measured_conn(
989 0 : DatabaseOperation::GetLeader,
990 0 : move |conn| -> DatabaseResult<_> {
991 0 : Ok(crate::schema::controllers::table.load::<ControllerPersistence>(conn)?)
992 0 : },
993 0 : )
994 0 : .await?;
995 :
996 0 : if leader.len() > 1 {
997 0 : return Err(DatabaseError::Logical(format!(
998 0 : "More than one entry present in the leader table: {leader:?}"
999 0 : )));
1000 0 : }
1001 0 :
1002 0 : Ok(leader.pop())
1003 0 : }
1004 :
1005 : /// Update the new leader with compare-exchange semantics. If `prev` does not
1006 : /// match the current leader entry, then the update is treated as a failure.
1007 : /// When `prev` is not specified, the update is forced.
1008 0 : pub(crate) async fn update_leader(
1009 0 : &self,
1010 0 : prev: Option<ControllerPersistence>,
1011 0 : new: ControllerPersistence,
1012 0 : ) -> DatabaseResult<()> {
1013 : use crate::schema::controllers::dsl::*;
1014 :
1015 0 : let updated = self
1016 0 : .with_measured_conn(
1017 0 : DatabaseOperation::UpdateLeader,
1018 0 : move |conn| -> DatabaseResult<usize> {
1019 0 : let updated = match &prev {
1020 0 : Some(prev) => diesel::update(controllers)
1021 0 : .filter(address.eq(prev.address.clone()))
1022 0 : .filter(started_at.eq(prev.started_at))
1023 0 : .set((
1024 0 : address.eq(new.address.clone()),
1025 0 : started_at.eq(new.started_at),
1026 0 : ))
1027 0 : .execute(conn)?,
1028 0 : None => diesel::insert_into(controllers)
1029 0 : .values(new.clone())
1030 0 : .execute(conn)?,
1031 : };
1032 :
1033 0 : Ok(updated)
1034 0 : },
1035 0 : )
1036 0 : .await?;
1037 :
1038 0 : if updated == 0 {
1039 0 : return Err(DatabaseError::Logical(
1040 0 : "Leader table update failed".to_string(),
1041 0 : ));
1042 0 : }
1043 0 :
1044 0 : Ok(())
1045 0 : }
1046 :
1047 : /// At startup, populate the list of nodes which our shards may be placed on
1048 0 : pub(crate) async fn list_safekeepers(&self) -> DatabaseResult<Vec<SafekeeperPersistence>> {
1049 0 : let safekeepers: Vec<SafekeeperPersistence> = self
1050 0 : .with_measured_conn(
1051 0 : DatabaseOperation::ListNodes,
1052 0 : move |conn| -> DatabaseResult<_> {
1053 0 : Ok(crate::schema::safekeepers::table.load::<SafekeeperPersistence>(conn)?)
1054 0 : },
1055 0 : )
1056 0 : .await?;
1057 :
1058 0 : tracing::info!("list_safekeepers: loaded {} nodes", safekeepers.len());
1059 :
1060 0 : Ok(safekeepers)
1061 0 : }
1062 :
1063 0 : pub(crate) async fn safekeeper_get(
1064 0 : &self,
1065 0 : id: i64,
1066 0 : ) -> Result<SafekeeperPersistence, DatabaseError> {
1067 : use crate::schema::safekeepers::dsl::{id as id_column, safekeepers};
1068 0 : self.with_conn(move |conn| -> DatabaseResult<SafekeeperPersistence> {
1069 0 : Ok(safekeepers
1070 0 : .filter(id_column.eq(&id))
1071 0 : .select(SafekeeperPersistence::as_select())
1072 0 : .get_result(conn)?)
1073 0 : })
1074 0 : .await
1075 0 : }
1076 :
1077 0 : pub(crate) async fn safekeeper_upsert(
1078 0 : &self,
1079 0 : record: SafekeeperUpsert,
1080 0 : ) -> Result<(), DatabaseError> {
1081 : use crate::schema::safekeepers::dsl::*;
1082 :
1083 0 : self.with_conn(move |conn| -> DatabaseResult<()> {
1084 0 : let bind = record
1085 0 : .as_insert_or_update()
1086 0 : .map_err(|e| DatabaseError::Logical(format!("{e}")))?;
1087 :
1088 0 : let inserted_updated = diesel::insert_into(safekeepers)
1089 0 : .values(&bind)
1090 0 : .on_conflict(id)
1091 0 : .do_update()
1092 0 : .set(&bind)
1093 0 : .execute(conn)?;
1094 :
1095 0 : if inserted_updated != 1 {
1096 0 : return Err(DatabaseError::Logical(format!(
1097 0 : "unexpected number of rows ({})",
1098 0 : inserted_updated
1099 0 : )));
1100 0 : }
1101 0 :
1102 0 : Ok(())
1103 0 : })
1104 0 : .await
1105 0 : }
1106 : }
1107 :
1108 : /// Parts of [`crate::tenant_shard::TenantShard`] that are stored durably
1109 : #[derive(
1110 0 : QueryableByName, Queryable, Selectable, Insertable, Serialize, Deserialize, Clone, Eq, PartialEq,
1111 : )]
1112 : #[diesel(table_name = crate::schema::tenant_shards)]
1113 : pub(crate) struct TenantShardPersistence {
1114 : #[serde(default)]
1115 : pub(crate) tenant_id: String,
1116 : #[serde(default)]
1117 : pub(crate) shard_number: i32,
1118 : #[serde(default)]
1119 : pub(crate) shard_count: i32,
1120 : #[serde(default)]
1121 : pub(crate) shard_stripe_size: i32,
1122 :
1123 : // Latest generation number: next time we attach, increment this
1124 : // and use the incremented number when attaching.
1125 : //
1126 : // Generation is only None when first onboarding a tenant, where it may
1127 : // be in PlacementPolicy::Secondary and therefore have no valid generation state.
1128 : pub(crate) generation: Option<i32>,
1129 :
1130 : // Currently attached pageserver
1131 : #[serde(rename = "pageserver")]
1132 : pub(crate) generation_pageserver: Option<i64>,
1133 :
1134 : #[serde(default)]
1135 : pub(crate) placement_policy: String,
1136 : #[serde(default)]
1137 : pub(crate) splitting: SplitState,
1138 : #[serde(default)]
1139 : pub(crate) config: String,
1140 : #[serde(default)]
1141 : pub(crate) scheduling_policy: String,
1142 :
1143 : // Hint that we should attempt to schedule this tenant shard the given
1144 : // availability zone in order to minimise the chances of cross-AZ communication
1145 : // with compute.
1146 : pub(crate) preferred_az_id: Option<String>,
1147 : }
1148 :
1149 : impl TenantShardPersistence {
1150 0 : pub(crate) fn get_shard_identity(&self) -> Result<ShardIdentity, ShardConfigError> {
1151 0 : if self.shard_count == 0 {
1152 0 : Ok(ShardIdentity::unsharded())
1153 : } else {
1154 0 : Ok(ShardIdentity::new(
1155 0 : ShardNumber(self.shard_number as u8),
1156 0 : ShardCount::new(self.shard_count as u8),
1157 0 : ShardStripeSize(self.shard_stripe_size as u32),
1158 0 : )?)
1159 : }
1160 0 : }
1161 :
1162 0 : pub(crate) fn get_tenant_shard_id(&self) -> Result<TenantShardId, hex::FromHexError> {
1163 0 : Ok(TenantShardId {
1164 0 : tenant_id: TenantId::from_str(self.tenant_id.as_str())?,
1165 0 : shard_number: ShardNumber(self.shard_number as u8),
1166 0 : shard_count: ShardCount::new(self.shard_count as u8),
1167 : })
1168 0 : }
1169 : }
1170 :
1171 : /// Parts of [`crate::node::Node`] that are stored durably
1172 0 : #[derive(Serialize, Deserialize, Queryable, Selectable, Insertable, Eq, PartialEq)]
1173 : #[diesel(table_name = crate::schema::nodes)]
1174 : pub(crate) struct NodePersistence {
1175 : pub(crate) node_id: i64,
1176 : pub(crate) scheduling_policy: String,
1177 : pub(crate) listen_http_addr: String,
1178 : pub(crate) listen_http_port: i32,
1179 : pub(crate) listen_pg_addr: String,
1180 : pub(crate) listen_pg_port: i32,
1181 : pub(crate) availability_zone_id: String,
1182 : }
1183 :
1184 : /// Tenant metadata health status that are stored durably.
1185 0 : #[derive(Queryable, Selectable, Insertable, Serialize, Deserialize, Clone, Eq, PartialEq)]
1186 : #[diesel(table_name = crate::schema::metadata_health)]
1187 : pub(crate) struct MetadataHealthPersistence {
1188 : #[serde(default)]
1189 : pub(crate) tenant_id: String,
1190 : #[serde(default)]
1191 : pub(crate) shard_number: i32,
1192 : #[serde(default)]
1193 : pub(crate) shard_count: i32,
1194 :
1195 : pub(crate) healthy: bool,
1196 : pub(crate) last_scrubbed_at: chrono::DateTime<chrono::Utc>,
1197 : }
1198 :
1199 : impl MetadataHealthPersistence {
1200 0 : pub fn new(
1201 0 : tenant_shard_id: TenantShardId,
1202 0 : healthy: bool,
1203 0 : last_scrubbed_at: chrono::DateTime<chrono::Utc>,
1204 0 : ) -> Self {
1205 0 : let tenant_id = tenant_shard_id.tenant_id.to_string();
1206 0 : let shard_number = tenant_shard_id.shard_number.0 as i32;
1207 0 : let shard_count = tenant_shard_id.shard_count.literal() as i32;
1208 0 :
1209 0 : MetadataHealthPersistence {
1210 0 : tenant_id,
1211 0 : shard_number,
1212 0 : shard_count,
1213 0 : healthy,
1214 0 : last_scrubbed_at,
1215 0 : }
1216 0 : }
1217 :
1218 : #[allow(dead_code)]
1219 0 : pub(crate) fn get_tenant_shard_id(&self) -> Result<TenantShardId, hex::FromHexError> {
1220 0 : Ok(TenantShardId {
1221 0 : tenant_id: TenantId::from_str(self.tenant_id.as_str())?,
1222 0 : shard_number: ShardNumber(self.shard_number as u8),
1223 0 : shard_count: ShardCount::new(self.shard_count as u8),
1224 : })
1225 0 : }
1226 : }
1227 :
1228 : impl From<MetadataHealthPersistence> for MetadataHealthRecord {
1229 0 : fn from(value: MetadataHealthPersistence) -> Self {
1230 0 : MetadataHealthRecord {
1231 0 : tenant_shard_id: value
1232 0 : .get_tenant_shard_id()
1233 0 : .expect("stored tenant id should be valid"),
1234 0 : healthy: value.healthy,
1235 0 : last_scrubbed_at: value.last_scrubbed_at,
1236 0 : }
1237 0 : }
1238 : }
1239 :
1240 : #[derive(
1241 0 : Serialize, Deserialize, Queryable, Selectable, Insertable, Eq, PartialEq, Debug, Clone,
1242 : )]
1243 : #[diesel(table_name = crate::schema::controllers)]
1244 : pub(crate) struct ControllerPersistence {
1245 : pub(crate) address: String,
1246 : pub(crate) started_at: chrono::DateTime<chrono::Utc>,
1247 : }
1248 :
1249 : // What we store in the database
1250 0 : #[derive(Serialize, Deserialize, Queryable, Selectable, Eq, PartialEq, Debug, Clone)]
1251 : #[diesel(table_name = crate::schema::safekeepers)]
1252 : pub(crate) struct SafekeeperPersistence {
1253 : pub(crate) id: i64,
1254 : pub(crate) region_id: String,
1255 : /// 1 is special, it means just created (not currently posted to storcon).
1256 : /// Zero or negative is not really expected.
1257 : /// Otherwise the number from `release-$(number_of_commits_on_branch)` tag.
1258 : pub(crate) version: i64,
1259 : pub(crate) host: String,
1260 : pub(crate) port: i32,
1261 : pub(crate) active: bool,
1262 : pub(crate) http_port: i32,
1263 : pub(crate) availability_zone_id: String,
1264 : pub(crate) scheduling_policy: String,
1265 : }
1266 :
1267 : impl SafekeeperPersistence {
1268 0 : pub(crate) fn as_describe_response(&self) -> Result<SafekeeperDescribeResponse, DatabaseError> {
1269 0 : let scheduling_policy =
1270 0 : SkSchedulingPolicy::from_str(&self.scheduling_policy).map_err(|e| {
1271 0 : DatabaseError::Logical(format!("can't construct SkSchedulingPolicy: {e:?}"))
1272 0 : })?;
1273 : // omit the `active` flag on purpose: it is deprecated.
1274 0 : Ok(SafekeeperDescribeResponse {
1275 0 : id: NodeId(self.id as u64),
1276 0 : region_id: self.region_id.clone(),
1277 0 : version: self.version,
1278 0 : host: self.host.clone(),
1279 0 : port: self.port,
1280 0 : http_port: self.http_port,
1281 0 : availability_zone_id: self.availability_zone_id.clone(),
1282 0 : scheduling_policy,
1283 0 : })
1284 0 : }
1285 : }
1286 :
1287 : /// What we expect from the upsert http api
1288 0 : #[derive(Serialize, Deserialize, Eq, PartialEq, Debug, Clone)]
1289 : pub(crate) struct SafekeeperUpsert {
1290 : pub(crate) id: i64,
1291 : pub(crate) region_id: String,
1292 : /// 1 is special, it means just created (not currently posted to storcon).
1293 : /// Zero or negative is not really expected.
1294 : /// Otherwise the number from `release-$(number_of_commits_on_branch)` tag.
1295 : pub(crate) version: i64,
1296 : pub(crate) host: String,
1297 : pub(crate) port: i32,
1298 : pub(crate) active: bool,
1299 : pub(crate) http_port: i32,
1300 : pub(crate) availability_zone_id: String,
1301 : }
1302 :
1303 : impl SafekeeperUpsert {
1304 0 : fn as_insert_or_update(&self) -> anyhow::Result<InsertUpdateSafekeeper<'_>> {
1305 0 : if self.version < 0 {
1306 0 : anyhow::bail!("negative version: {}", self.version);
1307 0 : }
1308 0 : Ok(InsertUpdateSafekeeper {
1309 0 : id: self.id,
1310 0 : region_id: &self.region_id,
1311 0 : version: self.version,
1312 0 : host: &self.host,
1313 0 : port: self.port,
1314 0 : active: self.active,
1315 0 : http_port: self.http_port,
1316 0 : availability_zone_id: &self.availability_zone_id,
1317 0 : // None means a wish to not update this column. We expose abilities to update it via other means.
1318 0 : scheduling_policy: None,
1319 0 : })
1320 0 : }
1321 : }
1322 :
1323 0 : #[derive(Insertable, AsChangeset)]
1324 : #[diesel(table_name = crate::schema::safekeepers)]
1325 : struct InsertUpdateSafekeeper<'a> {
1326 : id: i64,
1327 : region_id: &'a str,
1328 : version: i64,
1329 : host: &'a str,
1330 : port: i32,
1331 : active: bool,
1332 : http_port: i32,
1333 : availability_zone_id: &'a str,
1334 : scheduling_policy: Option<&'a str>,
1335 : }
|
