Line data Source code
1 : pub(crate) mod split_state;
2 : use std::collections::HashMap;
3 : use std::io::Write;
4 : use std::str::FromStr;
5 : use std::sync::Arc;
6 : use std::time::{Duration, Instant};
7 :
8 : use diesel::deserialize::{FromSql, FromSqlRow};
9 : use diesel::expression::AsExpression;
10 : use diesel::pg::Pg;
11 : use diesel::prelude::*;
12 : use diesel::serialize::{IsNull, ToSql};
13 : use diesel_async::async_connection_wrapper::AsyncConnectionWrapper;
14 : use diesel_async::pooled_connection::bb8::Pool;
15 : use diesel_async::pooled_connection::{AsyncDieselConnectionManager, ManagerConfig};
16 : use diesel_async::{AsyncPgConnection, RunQueryDsl};
17 : use diesel_migrations::{EmbeddedMigrations, embed_migrations};
18 : use futures::FutureExt;
19 : use futures::future::BoxFuture;
20 : use itertools::Itertools;
21 : use pageserver_api::controller_api::{
22 : AvailabilityZone, MetadataHealthRecord, NodeLifecycle, NodeSchedulingPolicy, PlacementPolicy,
23 : SafekeeperDescribeResponse, ShardSchedulingPolicy, SkSchedulingPolicy,
24 : };
25 : use pageserver_api::models::{ShardImportStatus, TenantConfig};
26 : use pageserver_api::shard::{
27 : ShardConfigError, ShardCount, ShardIdentity, ShardNumber, ShardStripeSize, TenantShardId,
28 : };
29 : use rustls::client::WebPkiServerVerifier;
30 : use rustls::client::danger::{ServerCertVerified, ServerCertVerifier};
31 : use rustls::crypto::ring;
32 : use safekeeper_api::membership::SafekeeperGeneration;
33 : use scoped_futures::ScopedBoxFuture;
34 : use serde::{Deserialize, Serialize};
35 : use utils::generation::Generation;
36 : use utils::id::{NodeId, TenantId, TimelineId};
37 : use utils::lsn::Lsn;
38 :
39 : use self::split_state::SplitState;
40 : use crate::metrics::{
41 : DatabaseQueryErrorLabelGroup, DatabaseQueryLatencyLabelGroup, METRICS_REGISTRY,
42 : };
43 : use crate::node::Node;
44 : use crate::timeline_import::{
45 : TimelineImport, TimelineImportUpdateError, TimelineImportUpdateFollowUp,
46 : };
47 : const MIGRATIONS: EmbeddedMigrations = embed_migrations!("./migrations");
48 :
49 : /// ## What do we store?
50 : ///
51 : /// The storage controller service does not store most of its state durably.
52 : ///
53 : /// The essential things to store durably are:
54 : /// - generation numbers, as these must always advance monotonically to ensure data safety.
55 : /// - Tenant's PlacementPolicy and TenantConfig, as the source of truth for these is something external.
56 : /// - Node's scheduling policies, as the source of truth for these is something external.
57 : ///
58 : /// Other things we store durably as an implementation detail:
59 : /// - Node's host/port: this could be avoided it we made nodes emit a self-registering heartbeat,
60 : /// but it is operationally simpler to make this service the authority for which nodes
61 : /// it talks to.
62 : ///
63 : /// ## Performance/efficiency
64 : ///
65 : /// The storage controller service does not go via the database for most things: there are
66 : /// a couple of places where we must, and where efficiency matters:
67 : /// - Incrementing generation numbers: the Reconciler has to wait for this to complete
68 : /// before it can attach a tenant, so this acts as a bound on how fast things like
69 : /// failover can happen.
70 : /// - Pageserver re-attach: we will increment many shards' generations when this happens,
71 : /// so it is important to avoid e.g. issuing O(N) queries.
72 : ///
73 : /// Database calls relating to nodes have low performance requirements, as they are very rarely
74 : /// updated, and reads of nodes are always from memory, not the database. We only require that
75 : /// we can UPDATE a node's scheduling mode reasonably quickly to mark a bad node offline.
76 : pub struct Persistence {
77 : connection_pool: Pool<AsyncPgConnection>,
78 : }
79 :
80 : /// Legacy format, for use in JSON compat objects in test environment
81 0 : #[derive(Serialize, Deserialize)]
82 : struct JsonPersistence {
83 : tenants: HashMap<TenantShardId, TenantShardPersistence>,
84 : }
85 :
86 : #[derive(thiserror::Error, Debug)]
87 : pub(crate) enum DatabaseError {
88 : #[error(transparent)]
89 : Query(#[from] diesel::result::Error),
90 : #[error(transparent)]
91 : Connection(#[from] diesel::result::ConnectionError),
92 : #[error(transparent)]
93 : ConnectionPool(#[from] diesel_async::pooled_connection::bb8::RunError),
94 : #[error("Logical error: {0}")]
95 : Logical(String),
96 : #[error("Migration error: {0}")]
97 : Migration(String),
98 : #[error("CAS error: {0}")]
99 : Cas(String),
100 : }
101 :
102 : #[derive(measured::FixedCardinalityLabel, Copy, Clone)]
103 : pub(crate) enum DatabaseOperation {
104 : InsertNode,
105 : UpdateNode,
106 : DeleteNode,
107 : ListNodes,
108 : ListTombstones,
109 : BeginShardSplit,
110 : CompleteShardSplit,
111 : AbortShardSplit,
112 : Detach,
113 : ReAttach,
114 : IncrementGeneration,
115 : TenantGenerations,
116 : ShardGenerations,
117 : ListTenantShards,
118 : LoadTenant,
119 : InsertTenantShards,
120 : UpdateTenantShard,
121 : DeleteTenant,
122 : UpdateTenantConfig,
123 : UpdateMetadataHealth,
124 : ListMetadataHealth,
125 : ListMetadataHealthUnhealthy,
126 : ListMetadataHealthOutdated,
127 : ListSafekeepers,
128 : GetLeader,
129 : UpdateLeader,
130 : SetPreferredAzs,
131 : InsertTimeline,
132 : UpdateTimeline,
133 : UpdateTimelineMembership,
134 : UpdateCplaneNotifiedGeneration,
135 : UpdateSkSetNotifiedGeneration,
136 : GetTimeline,
137 : InsertTimelineReconcile,
138 : RemoveTimelineReconcile,
139 : ListTimelineReconcile,
140 : ListTimelineReconcileStartup,
141 : InsertTimelineImport,
142 : UpdateTimelineImport,
143 : DeleteTimelineImport,
144 : ListTimelineImports,
145 : IsTenantImportingTimeline,
146 : }
147 :
148 : #[must_use]
149 : pub(crate) enum AbortShardSplitStatus {
150 : /// We aborted the split in the database by reverting to the parent shards
151 : Aborted,
152 : /// The split had already been persisted.
153 : Complete,
154 : }
155 :
156 : pub(crate) type DatabaseResult<T> = Result<T, DatabaseError>;
157 :
158 : /// Some methods can operate on either a whole tenant or a single shard
159 : #[derive(Clone)]
160 : pub(crate) enum TenantFilter {
161 : Tenant(TenantId),
162 : Shard(TenantShardId),
163 : }
164 :
165 : /// Represents the results of looking up generation+pageserver for the shards of a tenant
166 : pub(crate) struct ShardGenerationState {
167 : pub(crate) tenant_shard_id: TenantShardId,
168 : pub(crate) generation: Option<Generation>,
169 : pub(crate) generation_pageserver: Option<NodeId>,
170 : }
171 :
172 : // A generous allowance for how many times we may retry serializable transactions
173 : // before giving up. This is not expected to be hit: it is a defensive measure in case we
174 : // somehow engineer a situation where duelling transactions might otherwise live-lock.
175 : const MAX_RETRIES: usize = 128;
176 :
177 : impl Persistence {
178 : // The default postgres connection limit is 100. We use up to 99, to leave one free for a human admin under
179 : // normal circumstances. This assumes we have exclusive use of the database cluster to which we connect.
180 : pub const MAX_CONNECTIONS: u32 = 99;
181 :
182 : // We don't want to keep a lot of connections alive: close them down promptly if they aren't being used.
183 : const IDLE_CONNECTION_TIMEOUT: Duration = Duration::from_secs(10);
184 : const MAX_CONNECTION_LIFETIME: Duration = Duration::from_secs(60);
185 :
186 0 : pub async fn new(database_url: String) -> Self {
187 0 : let mut mgr_config = ManagerConfig::default();
188 0 : mgr_config.custom_setup = Box::new(establish_connection_rustls);
189 :
190 0 : let manager = AsyncDieselConnectionManager::<AsyncPgConnection>::new_with_config(
191 0 : database_url,
192 0 : mgr_config,
193 : );
194 :
195 : // We will use a connection pool: this is primarily to _limit_ our connection count, rather than to optimize time
196 : // to execute queries (database queries are not generally on latency-sensitive paths).
197 0 : let connection_pool = Pool::builder()
198 0 : .max_size(Self::MAX_CONNECTIONS)
199 0 : .max_lifetime(Some(Self::MAX_CONNECTION_LIFETIME))
200 0 : .idle_timeout(Some(Self::IDLE_CONNECTION_TIMEOUT))
201 0 : // Always keep at least one connection ready to go
202 0 : .min_idle(Some(1))
203 0 : .test_on_check_out(true)
204 0 : .build(manager)
205 0 : .await
206 0 : .expect("Could not build connection pool");
207 :
208 0 : Self { connection_pool }
209 0 : }
210 :
211 : /// A helper for use during startup, where we would like to tolerate concurrent restarts of the
212 : /// database and the storage controller, therefore the database might not be available right away
213 0 : pub async fn await_connection(
214 0 : database_url: &str,
215 0 : timeout: Duration,
216 0 : ) -> Result<(), diesel::ConnectionError> {
217 0 : let started_at = Instant::now();
218 0 : log_postgres_connstr_info(database_url)
219 0 : .map_err(|e| diesel::ConnectionError::InvalidConnectionUrl(e.to_string()))?;
220 : loop {
221 0 : match establish_connection_rustls(database_url).await {
222 : Ok(_) => {
223 0 : tracing::info!("Connected to database.");
224 0 : return Ok(());
225 : }
226 0 : Err(e) => {
227 0 : if started_at.elapsed() > timeout {
228 0 : return Err(e);
229 : } else {
230 0 : tracing::info!("Database not yet available, waiting... ({e})");
231 0 : tokio::time::sleep(Duration::from_millis(100)).await;
232 : }
233 : }
234 : }
235 : }
236 0 : }
237 :
238 : /// Execute the diesel migrations that are built into this binary
239 0 : pub(crate) async fn migration_run(&self) -> DatabaseResult<()> {
240 : use diesel_migrations::{HarnessWithOutput, MigrationHarness};
241 :
242 : // Can't use self.with_conn here as we do spawn_blocking which requires static.
243 0 : let conn = self
244 0 : .connection_pool
245 0 : .dedicated_connection()
246 0 : .await
247 0 : .map_err(|e| DatabaseError::Migration(e.to_string()))?;
248 0 : let mut async_wrapper: AsyncConnectionWrapper<AsyncPgConnection> =
249 0 : AsyncConnectionWrapper::from(conn);
250 0 : tokio::task::spawn_blocking(move || {
251 0 : let mut retry_count = 0;
252 : loop {
253 0 : let result = HarnessWithOutput::write_to_stdout(&mut async_wrapper)
254 0 : .run_pending_migrations(MIGRATIONS)
255 0 : .map(|_| ())
256 0 : .map_err(|e| DatabaseError::Migration(e.to_string()));
257 0 : match result {
258 0 : Ok(r) => break Ok(r),
259 : Err(
260 0 : err @ DatabaseError::Query(diesel::result::Error::DatabaseError(
261 : diesel::result::DatabaseErrorKind::SerializationFailure,
262 : _,
263 : )),
264 : ) => {
265 0 : retry_count += 1;
266 0 : if retry_count > MAX_RETRIES {
267 0 : tracing::error!(
268 0 : "Exceeded max retries on SerializationFailure errors: {err:?}"
269 : );
270 0 : break Err(err);
271 : } else {
272 : // Retry on serialization errors: these are expected, because even though our
273 : // transactions don't fight for the same rows, they will occasionally collide
274 : // on index pages (e.g. increment_generation for unrelated shards can collide)
275 0 : tracing::debug!(
276 0 : "Retrying transaction on serialization failure {err:?}"
277 : );
278 0 : continue;
279 : }
280 : }
281 0 : Err(e) => break Err(e),
282 : }
283 : }
284 0 : })
285 0 : .await
286 0 : .map_err(|e| DatabaseError::Migration(e.to_string()))??;
287 0 : Ok(())
288 0 : }
289 :
290 : /// Wraps `with_conn` in order to collect latency and error metrics
291 0 : async fn with_measured_conn<'a, 'b, F, R>(
292 0 : &self,
293 0 : op: DatabaseOperation,
294 0 : func: F,
295 0 : ) -> DatabaseResult<R>
296 0 : where
297 0 : F: for<'r> Fn(&'r mut AsyncPgConnection) -> ScopedBoxFuture<'b, 'r, DatabaseResult<R>>
298 0 : + Send
299 0 : + std::marker::Sync
300 0 : + 'a,
301 0 : R: Send + 'b,
302 0 : {
303 0 : let latency = &METRICS_REGISTRY
304 0 : .metrics_group
305 0 : .storage_controller_database_query_latency;
306 0 : let _timer = latency.start_timer(DatabaseQueryLatencyLabelGroup { operation: op });
307 :
308 0 : let res = self.with_conn(func).await;
309 :
310 0 : if let Err(err) = &res {
311 0 : let error_counter = &METRICS_REGISTRY
312 0 : .metrics_group
313 0 : .storage_controller_database_query_error;
314 0 : error_counter.inc(DatabaseQueryErrorLabelGroup {
315 0 : error_type: err.error_label(),
316 0 : operation: op,
317 0 : })
318 0 : }
319 :
320 0 : res
321 0 : }
322 :
323 : /// Call the provided function with a Diesel database connection in a retry loop
324 0 : async fn with_conn<'a, 'b, F, R>(&self, func: F) -> DatabaseResult<R>
325 0 : where
326 0 : F: for<'r> Fn(&'r mut AsyncPgConnection) -> ScopedBoxFuture<'b, 'r, DatabaseResult<R>>
327 0 : + Send
328 0 : + std::marker::Sync
329 0 : + 'a,
330 0 : R: Send + 'b,
331 0 : {
332 0 : let mut retry_count = 0;
333 : loop {
334 0 : let mut conn = self.connection_pool.get().await?;
335 0 : match conn
336 0 : .build_transaction()
337 0 : .serializable()
338 0 : .run(|c| func(c))
339 0 : .await
340 : {
341 0 : Ok(r) => break Ok(r),
342 : Err(
343 0 : err @ DatabaseError::Query(diesel::result::Error::DatabaseError(
344 : diesel::result::DatabaseErrorKind::SerializationFailure,
345 : _,
346 : )),
347 : ) => {
348 0 : retry_count += 1;
349 0 : if retry_count > MAX_RETRIES {
350 0 : tracing::error!(
351 0 : "Exceeded max retries on SerializationFailure errors: {err:?}"
352 : );
353 0 : break Err(err);
354 : } else {
355 : // Retry on serialization errors: these are expected, because even though our
356 : // transactions don't fight for the same rows, they will occasionally collide
357 : // on index pages (e.g. increment_generation for unrelated shards can collide)
358 0 : tracing::debug!("Retrying transaction on serialization failure {err:?}");
359 0 : continue;
360 : }
361 : }
362 0 : Err(e) => break Err(e),
363 : }
364 : }
365 0 : }
366 :
367 : /// When a node is first registered, persist it before using it for anything
368 : /// If the provided node_id already exists, it will be error.
369 : /// The common case is when a node marked for deletion wants to register.
370 0 : pub(crate) async fn insert_node(&self, node: &Node) -> DatabaseResult<()> {
371 0 : let np = &node.to_persistent();
372 0 : self.with_measured_conn(DatabaseOperation::InsertNode, move |conn| {
373 0 : Box::pin(async move {
374 0 : diesel::insert_into(crate::schema::nodes::table)
375 0 : .values(np)
376 0 : .execute(conn)
377 0 : .await?;
378 0 : Ok(())
379 0 : })
380 0 : })
381 0 : .await
382 0 : }
383 :
384 : /// At startup, populate the list of nodes which our shards may be placed on
385 0 : pub(crate) async fn list_nodes(&self) -> DatabaseResult<Vec<NodePersistence>> {
386 : use crate::schema::nodes::dsl::*;
387 :
388 0 : let result: Vec<NodePersistence> = self
389 0 : .with_measured_conn(DatabaseOperation::ListNodes, move |conn| {
390 0 : Box::pin(async move {
391 0 : Ok(crate::schema::nodes::table
392 0 : .filter(lifecycle.ne(String::from(NodeLifecycle::Deleted)))
393 0 : .load::<NodePersistence>(conn)
394 0 : .await?)
395 0 : })
396 0 : })
397 0 : .await?;
398 :
399 0 : tracing::info!("list_nodes: loaded {} nodes", result.len());
400 :
401 0 : Ok(result)
402 0 : }
403 :
404 0 : pub(crate) async fn list_tombstones(&self) -> DatabaseResult<Vec<NodePersistence>> {
405 : use crate::schema::nodes::dsl::*;
406 :
407 0 : let result: Vec<NodePersistence> = self
408 0 : .with_measured_conn(DatabaseOperation::ListTombstones, move |conn| {
409 0 : Box::pin(async move {
410 0 : Ok(crate::schema::nodes::table
411 0 : .filter(lifecycle.eq(String::from(NodeLifecycle::Deleted)))
412 0 : .load::<NodePersistence>(conn)
413 0 : .await?)
414 0 : })
415 0 : })
416 0 : .await?;
417 :
418 0 : tracing::info!("list_tombstones: loaded {} nodes", result.len());
419 :
420 0 : Ok(result)
421 0 : }
422 :
423 0 : pub(crate) async fn update_node<V>(
424 0 : &self,
425 0 : input_node_id: NodeId,
426 0 : values: V,
427 0 : ) -> DatabaseResult<()>
428 0 : where
429 0 : V: diesel::AsChangeset<Target = crate::schema::nodes::table> + Clone + Send + Sync,
430 0 : V::Changeset: diesel::query_builder::QueryFragment<diesel::pg::Pg> + Send, // valid Postgres SQL
431 0 : {
432 : use crate::schema::nodes::dsl::*;
433 0 : let updated = self
434 0 : .with_measured_conn(DatabaseOperation::UpdateNode, move |conn| {
435 0 : let values = values.clone();
436 0 : Box::pin(async move {
437 0 : let updated = diesel::update(nodes)
438 0 : .filter(node_id.eq(input_node_id.0 as i64))
439 0 : .filter(lifecycle.ne(String::from(NodeLifecycle::Deleted)))
440 0 : .set(values)
441 0 : .execute(conn)
442 0 : .await?;
443 0 : Ok(updated)
444 0 : })
445 0 : })
446 0 : .await?;
447 :
448 0 : if updated != 1 {
449 0 : Err(DatabaseError::Logical(format!(
450 0 : "Node {node_id:?} not found for update",
451 0 : )))
452 : } else {
453 0 : Ok(())
454 : }
455 0 : }
456 :
457 0 : pub(crate) async fn update_node_scheduling_policy(
458 0 : &self,
459 0 : input_node_id: NodeId,
460 0 : input_scheduling: NodeSchedulingPolicy,
461 0 : ) -> DatabaseResult<()> {
462 : use crate::schema::nodes::dsl::*;
463 0 : self.update_node(
464 0 : input_node_id,
465 0 : scheduling_policy.eq(String::from(input_scheduling)),
466 0 : )
467 0 : .await
468 0 : }
469 :
470 0 : pub(crate) async fn update_node_on_registration(
471 0 : &self,
472 0 : input_node_id: NodeId,
473 0 : input_https_port: Option<u16>,
474 0 : input_grpc_addr: Option<String>,
475 0 : input_grpc_port: Option<u16>,
476 0 : ) -> DatabaseResult<()> {
477 : use crate::schema::nodes::dsl::*;
478 0 : self.update_node(
479 0 : input_node_id,
480 : (
481 0 : listen_https_port.eq(input_https_port.map(|x| x as i32)),
482 0 : listen_grpc_addr.eq(input_grpc_addr),
483 0 : listen_grpc_port.eq(input_grpc_port.map(|x| x as i32)),
484 : ),
485 : )
486 0 : .await
487 0 : }
488 :
489 : /// Tombstone is a special state where the node is not deleted from the database,
490 : /// but it is not available for usage.
491 : /// The main reason for it is to prevent the flaky node to register.
492 0 : pub(crate) async fn set_tombstone(&self, del_node_id: NodeId) -> DatabaseResult<()> {
493 : use crate::schema::nodes::dsl::*;
494 0 : self.update_node(
495 0 : del_node_id,
496 0 : lifecycle.eq(String::from(NodeLifecycle::Deleted)),
497 0 : )
498 0 : .await
499 0 : }
500 :
501 0 : pub(crate) async fn delete_node(&self, del_node_id: NodeId) -> DatabaseResult<()> {
502 : use crate::schema::nodes::dsl::*;
503 0 : self.with_measured_conn(DatabaseOperation::DeleteNode, move |conn| {
504 0 : Box::pin(async move {
505 : // You can hard delete a node only if it has a tombstone.
506 : // So we need to check if the node has lifecycle set to deleted.
507 0 : let node_to_delete = nodes
508 0 : .filter(node_id.eq(del_node_id.0 as i64))
509 0 : .first::<NodePersistence>(conn)
510 0 : .await
511 0 : .optional()?;
512 :
513 0 : if let Some(np) = node_to_delete {
514 0 : let lc = NodeLifecycle::from_str(&np.lifecycle).map_err(|e| {
515 0 : DatabaseError::Logical(format!(
516 0 : "Node {del_node_id} has invalid lifecycle: {e}"
517 0 : ))
518 0 : })?;
519 :
520 0 : if lc != NodeLifecycle::Deleted {
521 0 : return Err(DatabaseError::Logical(format!(
522 0 : "Node {del_node_id} was not soft deleted before, cannot hard delete it"
523 0 : )));
524 0 : }
525 :
526 0 : diesel::delete(nodes)
527 0 : .filter(node_id.eq(del_node_id.0 as i64))
528 0 : .execute(conn)
529 0 : .await?;
530 0 : }
531 :
532 0 : Ok(())
533 0 : })
534 0 : })
535 0 : .await
536 0 : }
537 :
538 : /// At startup, load the high level state for shards, such as their config + policy. This will
539 : /// be enriched at runtime with state discovered on pageservers.
540 : ///
541 : /// We exclude shards configured to be detached. During startup, if we see any attached locations
542 : /// for such shards, they will automatically be detached as 'orphans'.
543 0 : pub(crate) async fn load_active_tenant_shards(
544 0 : &self,
545 0 : ) -> DatabaseResult<Vec<TenantShardPersistence>> {
546 : use crate::schema::tenant_shards::dsl::*;
547 0 : self.with_measured_conn(DatabaseOperation::ListTenantShards, move |conn| {
548 0 : Box::pin(async move {
549 0 : let query = tenant_shards.filter(
550 0 : placement_policy.ne(serde_json::to_string(&PlacementPolicy::Detached).unwrap()),
551 : );
552 0 : let result = query.load::<TenantShardPersistence>(conn).await?;
553 :
554 0 : Ok(result)
555 0 : })
556 0 : })
557 0 : .await
558 0 : }
559 :
560 : /// When restoring a previously detached tenant into memory, load it from the database
561 0 : pub(crate) async fn load_tenant(
562 0 : &self,
563 0 : filter_tenant_id: TenantId,
564 0 : ) -> DatabaseResult<Vec<TenantShardPersistence>> {
565 : use crate::schema::tenant_shards::dsl::*;
566 0 : self.with_measured_conn(DatabaseOperation::LoadTenant, move |conn| {
567 0 : Box::pin(async move {
568 0 : let query = tenant_shards.filter(tenant_id.eq(filter_tenant_id.to_string()));
569 0 : let result = query.load::<TenantShardPersistence>(conn).await?;
570 :
571 0 : Ok(result)
572 0 : })
573 0 : })
574 0 : .await
575 0 : }
576 :
577 : /// Tenants must be persisted before we schedule them for the first time. This enables us
578 : /// to correctly retain generation monotonicity, and the externally provided placement policy & config.
579 0 : pub(crate) async fn insert_tenant_shards(
580 0 : &self,
581 0 : shards: Vec<TenantShardPersistence>,
582 0 : ) -> DatabaseResult<()> {
583 : use crate::schema::{metadata_health, tenant_shards};
584 :
585 0 : let now = chrono::Utc::now();
586 :
587 0 : let metadata_health_records = shards
588 0 : .iter()
589 0 : .map(|t| MetadataHealthPersistence {
590 0 : tenant_id: t.tenant_id.clone(),
591 0 : shard_number: t.shard_number,
592 0 : shard_count: t.shard_count,
593 : healthy: true,
594 0 : last_scrubbed_at: now,
595 0 : })
596 0 : .collect::<Vec<_>>();
597 :
598 0 : let shards = &shards;
599 0 : let metadata_health_records = &metadata_health_records;
600 0 : self.with_measured_conn(DatabaseOperation::InsertTenantShards, move |conn| {
601 0 : Box::pin(async move {
602 0 : diesel::insert_into(tenant_shards::table)
603 0 : .values(shards)
604 0 : .execute(conn)
605 0 : .await?;
606 :
607 0 : diesel::insert_into(metadata_health::table)
608 0 : .values(metadata_health_records)
609 0 : .execute(conn)
610 0 : .await?;
611 0 : Ok(())
612 0 : })
613 0 : })
614 0 : .await
615 0 : }
616 :
617 : /// Ordering: call this _after_ deleting the tenant on pageservers, but _before_ dropping state for
618 : /// the tenant from memory on this server.
619 0 : pub(crate) async fn delete_tenant(&self, del_tenant_id: TenantId) -> DatabaseResult<()> {
620 : use crate::schema::tenant_shards::dsl::*;
621 0 : self.with_measured_conn(DatabaseOperation::DeleteTenant, move |conn| {
622 0 : Box::pin(async move {
623 : // `metadata_health` status (if exists) is also deleted based on the cascade behavior.
624 0 : diesel::delete(tenant_shards)
625 0 : .filter(tenant_id.eq(del_tenant_id.to_string()))
626 0 : .execute(conn)
627 0 : .await?;
628 0 : Ok(())
629 0 : })
630 0 : })
631 0 : .await
632 0 : }
633 :
634 : /// When a tenant invokes the /re-attach API, this function is responsible for doing an efficient
635 : /// batched increment of the generations of all tenants whose generation_pageserver is equal to
636 : /// the node that called /re-attach.
637 : #[tracing::instrument(skip_all, fields(node_id))]
638 : pub(crate) async fn re_attach(
639 : &self,
640 : input_node_id: NodeId,
641 : ) -> DatabaseResult<HashMap<TenantShardId, Generation>> {
642 : use crate::schema::nodes::dsl::{scheduling_policy, *};
643 : use crate::schema::tenant_shards::dsl::*;
644 : let updated = self
645 0 : .with_measured_conn(DatabaseOperation::ReAttach, move |conn| {
646 0 : Box::pin(async move {
647 0 : let node: Option<NodePersistence> = nodes
648 0 : .filter(node_id.eq(input_node_id.0 as i64))
649 0 : .first::<NodePersistence>(conn)
650 0 : .await
651 0 : .optional()?;
652 :
653 : // Check if the node is not marked as deleted
654 0 : match node {
655 0 : Some(node) if matches!(NodeLifecycle::from_str(&node.lifecycle), Ok(NodeLifecycle::Deleted)) => {
656 0 : return Err(DatabaseError::Logical(format!(
657 0 : "Node {input_node_id} is marked as deleted, re-attach is not allowed"
658 0 : )));
659 : }
660 0 : _ => {
661 0 : // go through
662 0 : }
663 : };
664 :
665 0 : let rows_updated = diesel::update(tenant_shards)
666 0 : .filter(generation_pageserver.eq(input_node_id.0 as i64))
667 0 : .set(generation.eq(generation + 1))
668 0 : .execute(conn)
669 0 : .await?;
670 :
671 0 : tracing::info!("Incremented {} tenants' generations", rows_updated);
672 :
673 : // TODO: UPDATE+SELECT in one query
674 :
675 0 : let updated = tenant_shards
676 0 : .filter(generation_pageserver.eq(input_node_id.0 as i64))
677 0 : .select(TenantShardPersistence::as_select())
678 0 : .load(conn)
679 0 : .await?;
680 :
681 0 : if let Some(node) = node {
682 0 : let old_scheduling_policy =
683 0 : NodeSchedulingPolicy::from_str(&node.scheduling_policy).unwrap();
684 0 : let new_scheduling_policy = match old_scheduling_policy {
685 0 : NodeSchedulingPolicy::Active => NodeSchedulingPolicy::Active,
686 0 : NodeSchedulingPolicy::PauseForRestart => NodeSchedulingPolicy::Active,
687 0 : NodeSchedulingPolicy::Draining => NodeSchedulingPolicy::Active,
688 0 : NodeSchedulingPolicy::Filling => NodeSchedulingPolicy::Active,
689 0 : NodeSchedulingPolicy::Pause => NodeSchedulingPolicy::Pause,
690 0 : NodeSchedulingPolicy::Deleting => NodeSchedulingPolicy::Pause,
691 : };
692 0 : diesel::update(nodes)
693 0 : .filter(node_id.eq(input_node_id.0 as i64))
694 0 : .set(scheduling_policy.eq(String::from(new_scheduling_policy)))
695 0 : .execute(conn)
696 0 : .await?;
697 0 : }
698 :
699 0 : Ok(updated)
700 0 : })
701 0 : })
702 : .await?;
703 :
704 : let mut result = HashMap::new();
705 : for tsp in updated {
706 : let tenant_shard_id = TenantShardId {
707 : tenant_id: TenantId::from_str(tsp.tenant_id.as_str())
708 0 : .map_err(|e| DatabaseError::Logical(format!("Malformed tenant id: {e}")))?,
709 : shard_number: ShardNumber(tsp.shard_number as u8),
710 : shard_count: ShardCount::new(tsp.shard_count as u8),
711 : };
712 :
713 : let Some(g) = tsp.generation else {
714 : // If the generation_pageserver column was non-NULL, then the generation column should also be non-NULL:
715 : // we only set generation_pageserver when setting generation.
716 : return Err(DatabaseError::Logical(
717 : "Generation should always be set after incrementing".to_string(),
718 : ));
719 : };
720 : result.insert(tenant_shard_id, Generation::new(g as u32));
721 : }
722 :
723 : Ok(result)
724 : }
725 :
726 : /// Reconciler calls this immediately before attaching to a new pageserver, to acquire a unique, monotonically
727 : /// advancing generation number. We also store the NodeId for which the generation was issued, so that in
728 : /// [`Self::re_attach`] we can do a bulk UPDATE on the generations for that node.
729 0 : pub(crate) async fn increment_generation(
730 0 : &self,
731 0 : tenant_shard_id: TenantShardId,
732 0 : node_id: NodeId,
733 0 : ) -> anyhow::Result<Generation> {
734 : use crate::schema::tenant_shards::dsl::*;
735 0 : let updated = self
736 0 : .with_measured_conn(DatabaseOperation::IncrementGeneration, move |conn| {
737 0 : Box::pin(async move {
738 0 : let updated = diesel::update(tenant_shards)
739 0 : .filter(tenant_id.eq(tenant_shard_id.tenant_id.to_string()))
740 0 : .filter(shard_number.eq(tenant_shard_id.shard_number.0 as i32))
741 0 : .filter(shard_count.eq(tenant_shard_id.shard_count.literal() as i32))
742 0 : .set((
743 0 : generation.eq(generation + 1),
744 0 : generation_pageserver.eq(node_id.0 as i64),
745 0 : ))
746 0 : // TODO: only returning() the generation column
747 0 : .returning(TenantShardPersistence::as_returning())
748 0 : .get_result(conn)
749 0 : .await?;
750 :
751 0 : Ok(updated)
752 0 : })
753 0 : })
754 0 : .await?;
755 :
756 : // Generation is always non-null in the rseult: if the generation column had been NULL, then we
757 : // should have experienced an SQL Confilict error while executing a query that tries to increment it.
758 0 : debug_assert!(updated.generation.is_some());
759 0 : let Some(g) = updated.generation else {
760 0 : return Err(DatabaseError::Logical(
761 0 : "Generation should always be set after incrementing".to_string(),
762 0 : )
763 0 : .into());
764 : };
765 :
766 0 : Ok(Generation::new(g as u32))
767 0 : }
768 :
769 : /// When we want to call out to the running shards for a tenant, e.g. during timeline CRUD operations,
770 : /// we need to know where the shard is attached, _and_ the generation, so that we can re-check the generation
771 : /// afterwards to confirm that our timeline CRUD operation is truly persistent (it must have happened in the
772 : /// latest generation)
773 : ///
774 : /// If the tenant doesn't exist, an empty vector is returned.
775 : ///
776 : /// Output is sorted by shard number
777 0 : pub(crate) async fn tenant_generations(
778 0 : &self,
779 0 : filter_tenant_id: TenantId,
780 0 : ) -> Result<Vec<ShardGenerationState>, DatabaseError> {
781 : use crate::schema::tenant_shards::dsl::*;
782 0 : let rows = self
783 0 : .with_measured_conn(DatabaseOperation::TenantGenerations, move |conn| {
784 0 : Box::pin(async move {
785 0 : let result = tenant_shards
786 0 : .filter(tenant_id.eq(filter_tenant_id.to_string()))
787 0 : .select(TenantShardPersistence::as_select())
788 0 : .order(shard_number)
789 0 : .load(conn)
790 0 : .await?;
791 0 : Ok(result)
792 0 : })
793 0 : })
794 0 : .await?;
795 :
796 0 : Ok(rows
797 0 : .into_iter()
798 0 : .map(|p| ShardGenerationState {
799 0 : tenant_shard_id: p
800 0 : .get_tenant_shard_id()
801 0 : .expect("Corrupt tenant shard id in database"),
802 0 : generation: p.generation.map(|g| Generation::new(g as u32)),
803 0 : generation_pageserver: p.generation_pageserver.map(|n| NodeId(n as u64)),
804 0 : })
805 0 : .collect())
806 0 : }
807 :
808 : /// Read the generation number of specific tenant shards
809 : ///
810 : /// Output is unsorted. Output may not include values for all inputs, if they are missing in the database.
811 0 : pub(crate) async fn shard_generations(
812 0 : &self,
813 0 : mut tenant_shard_ids: impl Iterator<Item = &TenantShardId>,
814 0 : ) -> Result<Vec<(TenantShardId, Option<Generation>)>, DatabaseError> {
815 0 : let mut rows = Vec::with_capacity(tenant_shard_ids.size_hint().0);
816 :
817 : // We will chunk our input to avoid composing arbitrarily long `IN` clauses. Typically we are
818 : // called with a single digit number of IDs, but in principle we could be called with tens
819 : // of thousands (all the shards on one pageserver) from the generation validation API.
820 : loop {
821 : // A modest hardcoded chunk size to handle typical cases in a single query but never generate particularly
822 : // large query strings.
823 0 : let chunk_ids = tenant_shard_ids.by_ref().take(32);
824 :
825 : // Compose a comma separated list of tuples for matching on (tenant_id, shard_number, shard_count)
826 0 : let in_clause = chunk_ids
827 0 : .map(|tsid| {
828 0 : format!(
829 0 : "('{}', {}, {})",
830 : tsid.tenant_id, tsid.shard_number.0, tsid.shard_count.0
831 : )
832 0 : })
833 0 : .join(",");
834 :
835 : // We are done when our iterator gives us nothing to filter on
836 0 : if in_clause.is_empty() {
837 0 : break;
838 0 : }
839 :
840 0 : let in_clause = &in_clause;
841 0 : let chunk_rows = self
842 0 : .with_measured_conn(DatabaseOperation::ShardGenerations, move |conn| {
843 0 : Box::pin(async move {
844 : // diesel doesn't support multi-column IN queries, so we compose raw SQL. No escaping is required because
845 : // the inputs are strongly typed and cannot carry any user-supplied raw string content.
846 0 : let result : Vec<TenantShardPersistence> = diesel::sql_query(
847 0 : format!("SELECT * from tenant_shards where (tenant_id, shard_number, shard_count) in ({in_clause});").as_str()
848 0 : ).load(conn).await?;
849 :
850 0 : Ok(result)
851 0 : })
852 0 : })
853 0 : .await?;
854 0 : rows.extend(chunk_rows.into_iter())
855 : }
856 :
857 0 : Ok(rows
858 0 : .into_iter()
859 0 : .map(|tsp| {
860 : (
861 0 : tsp.get_tenant_shard_id()
862 0 : .expect("Bad tenant ID in database"),
863 0 : tsp.generation.map(|g| Generation::new(g as u32)),
864 : )
865 0 : })
866 0 : .collect())
867 0 : }
868 :
869 : #[allow(non_local_definitions)]
870 : /// For use when updating a persistent property of a tenant, such as its config or placement_policy.
871 : ///
872 : /// Do not use this for settting generation, unless in the special onboarding code path (/location_config)
873 : /// API: use [`Self::increment_generation`] instead. Setting the generation via this route is a one-time thing
874 : /// that we only do the first time a tenant is set to an attached policy via /location_config.
875 0 : pub(crate) async fn update_tenant_shard(
876 0 : &self,
877 0 : tenant: TenantFilter,
878 0 : input_placement_policy: Option<PlacementPolicy>,
879 0 : input_config: Option<TenantConfig>,
880 0 : input_generation: Option<Generation>,
881 0 : input_scheduling_policy: Option<ShardSchedulingPolicy>,
882 0 : ) -> DatabaseResult<()> {
883 : use crate::schema::tenant_shards::dsl::*;
884 :
885 0 : let tenant = &tenant;
886 0 : let input_placement_policy = &input_placement_policy;
887 0 : let input_config = &input_config;
888 0 : let input_generation = &input_generation;
889 0 : let input_scheduling_policy = &input_scheduling_policy;
890 0 : self.with_measured_conn(DatabaseOperation::UpdateTenantShard, move |conn| {
891 0 : Box::pin(async move {
892 0 : let query = match tenant {
893 0 : TenantFilter::Shard(tenant_shard_id) => diesel::update(tenant_shards)
894 0 : .filter(tenant_id.eq(tenant_shard_id.tenant_id.to_string()))
895 0 : .filter(shard_number.eq(tenant_shard_id.shard_number.0 as i32))
896 0 : .filter(shard_count.eq(tenant_shard_id.shard_count.literal() as i32))
897 0 : .into_boxed(),
898 0 : TenantFilter::Tenant(input_tenant_id) => diesel::update(tenant_shards)
899 0 : .filter(tenant_id.eq(input_tenant_id.to_string()))
900 0 : .into_boxed(),
901 : };
902 :
903 : // Clear generation_pageserver if we are moving into a state where we won't have
904 : // any attached pageservers.
905 0 : let input_generation_pageserver = match input_placement_policy {
906 0 : None | Some(PlacementPolicy::Attached(_)) => None,
907 0 : Some(PlacementPolicy::Detached | PlacementPolicy::Secondary) => Some(None),
908 : };
909 :
910 : #[derive(AsChangeset)]
911 : #[diesel(table_name = crate::schema::tenant_shards)]
912 : struct ShardUpdate {
913 : generation: Option<i32>,
914 : placement_policy: Option<String>,
915 : config: Option<String>,
916 : scheduling_policy: Option<String>,
917 : generation_pageserver: Option<Option<i64>>,
918 : }
919 :
920 0 : let update = ShardUpdate {
921 0 : generation: input_generation.map(|g| g.into().unwrap() as i32),
922 0 : placement_policy: input_placement_policy
923 0 : .as_ref()
924 0 : .map(|p| serde_json::to_string(&p).unwrap()),
925 0 : config: input_config
926 0 : .as_ref()
927 0 : .map(|c| serde_json::to_string(&c).unwrap()),
928 0 : scheduling_policy: input_scheduling_policy
929 0 : .map(|p| serde_json::to_string(&p).unwrap()),
930 0 : generation_pageserver: input_generation_pageserver,
931 : };
932 :
933 0 : query.set(update).execute(conn).await?;
934 :
935 0 : Ok(())
936 0 : })
937 0 : })
938 0 : .await?;
939 :
940 0 : Ok(())
941 0 : }
942 :
943 : /// Note that passing None for a shard clears the preferred AZ (rather than leaving it unmodified)
944 0 : pub(crate) async fn set_tenant_shard_preferred_azs(
945 0 : &self,
946 0 : preferred_azs: Vec<(TenantShardId, Option<AvailabilityZone>)>,
947 0 : ) -> DatabaseResult<Vec<(TenantShardId, Option<AvailabilityZone>)>> {
948 : use crate::schema::tenant_shards::dsl::*;
949 :
950 0 : let preferred_azs = preferred_azs.as_slice();
951 0 : self.with_measured_conn(DatabaseOperation::SetPreferredAzs, move |conn| {
952 0 : Box::pin(async move {
953 0 : let mut shards_updated = Vec::default();
954 :
955 0 : for (tenant_shard_id, preferred_az) in preferred_azs.iter() {
956 0 : let updated = diesel::update(tenant_shards)
957 0 : .filter(tenant_id.eq(tenant_shard_id.tenant_id.to_string()))
958 0 : .filter(shard_number.eq(tenant_shard_id.shard_number.0 as i32))
959 0 : .filter(shard_count.eq(tenant_shard_id.shard_count.literal() as i32))
960 0 : .set(preferred_az_id.eq(preferred_az.as_ref().map(|az| az.0.clone())))
961 0 : .execute(conn)
962 0 : .await?;
963 :
964 0 : if updated == 1 {
965 0 : shards_updated.push((*tenant_shard_id, preferred_az.clone()));
966 0 : }
967 : }
968 :
969 0 : Ok(shards_updated)
970 0 : })
971 0 : })
972 0 : .await
973 0 : }
974 :
975 0 : pub(crate) async fn detach(&self, tenant_shard_id: TenantShardId) -> anyhow::Result<()> {
976 : use crate::schema::tenant_shards::dsl::*;
977 0 : self.with_measured_conn(DatabaseOperation::Detach, move |conn| {
978 0 : Box::pin(async move {
979 0 : let updated = diesel::update(tenant_shards)
980 0 : .filter(tenant_id.eq(tenant_shard_id.tenant_id.to_string()))
981 0 : .filter(shard_number.eq(tenant_shard_id.shard_number.0 as i32))
982 0 : .filter(shard_count.eq(tenant_shard_id.shard_count.literal() as i32))
983 0 : .set((
984 0 : generation_pageserver.eq(Option::<i64>::None),
985 0 : placement_policy
986 0 : .eq(serde_json::to_string(&PlacementPolicy::Detached).unwrap()),
987 0 : ))
988 0 : .execute(conn)
989 0 : .await?;
990 :
991 0 : Ok(updated)
992 0 : })
993 0 : })
994 0 : .await?;
995 :
996 0 : Ok(())
997 0 : }
998 :
999 : // When we start shard splitting, we must durably mark the tenant so that
1000 : // on restart, we know that we must go through recovery.
1001 : //
1002 : // We create the child shards here, so that they will be available for increment_generation calls
1003 : // if some pageserver holding a child shard needs to restart before the overall tenant split is complete.
1004 0 : pub(crate) async fn begin_shard_split(
1005 0 : &self,
1006 0 : old_shard_count: ShardCount,
1007 0 : split_tenant_id: TenantId,
1008 0 : parent_to_children: Vec<(TenantShardId, Vec<TenantShardPersistence>)>,
1009 0 : ) -> DatabaseResult<()> {
1010 : use crate::schema::tenant_shards::dsl::*;
1011 0 : let parent_to_children = parent_to_children.as_slice();
1012 0 : self.with_measured_conn(DatabaseOperation::BeginShardSplit, move |conn| {
1013 0 : Box::pin(async move {
1014 : // Mark parent shards as splitting
1015 :
1016 0 : let updated = diesel::update(tenant_shards)
1017 0 : .filter(tenant_id.eq(split_tenant_id.to_string()))
1018 0 : .filter(shard_count.eq(old_shard_count.literal() as i32))
1019 0 : .set((splitting.eq(1),))
1020 0 : .execute(conn).await?;
1021 0 : if u8::try_from(updated)
1022 0 : .map_err(|_| DatabaseError::Logical(
1023 0 : format!("Overflow existing shard count {updated} while splitting"))
1024 0 : )? != old_shard_count.count() {
1025 : // Perhaps a deletion or another split raced with this attempt to split, mutating
1026 : // the parent shards that we intend to split. In this case the split request should fail.
1027 0 : return Err(DatabaseError::Logical(
1028 0 : format!("Unexpected existing shard count {updated} when preparing tenant for split (expected {})", old_shard_count.count())
1029 0 : ));
1030 0 : }
1031 :
1032 : // FIXME: spurious clone to sidestep closure move rules
1033 0 : let parent_to_children = parent_to_children.to_vec();
1034 :
1035 : // Insert child shards
1036 0 : for (parent_shard_id, children) in parent_to_children {
1037 0 : let mut parent = crate::schema::tenant_shards::table
1038 0 : .filter(tenant_id.eq(parent_shard_id.tenant_id.to_string()))
1039 0 : .filter(shard_number.eq(parent_shard_id.shard_number.0 as i32))
1040 0 : .filter(shard_count.eq(parent_shard_id.shard_count.literal() as i32))
1041 0 : .load::<TenantShardPersistence>(conn).await?;
1042 0 : let parent = if parent.len() != 1 {
1043 0 : return Err(DatabaseError::Logical(format!(
1044 0 : "Parent shard {parent_shard_id} not found"
1045 0 : )));
1046 : } else {
1047 0 : parent.pop().unwrap()
1048 : };
1049 0 : for mut shard in children {
1050 : // Carry the parent's generation into the child
1051 0 : shard.generation = parent.generation;
1052 :
1053 0 : debug_assert!(shard.splitting == SplitState::Splitting);
1054 0 : diesel::insert_into(tenant_shards)
1055 0 : .values(shard)
1056 0 : .execute(conn).await?;
1057 : }
1058 : }
1059 :
1060 0 : Ok(())
1061 0 : })
1062 0 : })
1063 0 : .await
1064 0 : }
1065 :
1066 : // When we finish shard splitting, we must atomically clean up the old shards
1067 : // and insert the new shards, and clear the splitting marker.
1068 0 : pub(crate) async fn complete_shard_split(
1069 0 : &self,
1070 0 : split_tenant_id: TenantId,
1071 0 : old_shard_count: ShardCount,
1072 0 : new_shard_count: ShardCount,
1073 0 : ) -> DatabaseResult<()> {
1074 : use crate::schema::tenant_shards::dsl::*;
1075 0 : self.with_measured_conn(DatabaseOperation::CompleteShardSplit, move |conn| {
1076 0 : Box::pin(async move {
1077 : // Sanity: child shards must still exist, as we're deleting parent shards
1078 0 : let child_shards_query = tenant_shards
1079 0 : .filter(tenant_id.eq(split_tenant_id.to_string()))
1080 0 : .filter(shard_count.eq(new_shard_count.literal() as i32));
1081 0 : let child_shards = child_shards_query
1082 0 : .load::<TenantShardPersistence>(conn)
1083 0 : .await?;
1084 0 : if child_shards.len() != new_shard_count.count() as usize {
1085 0 : return Err(DatabaseError::Logical(format!(
1086 0 : "Unexpected child shard count {} while completing split to \
1087 0 : count {new_shard_count:?} on tenant {split_tenant_id}",
1088 0 : child_shards.len()
1089 0 : )));
1090 0 : }
1091 :
1092 : // Drop parent shards
1093 0 : diesel::delete(tenant_shards)
1094 0 : .filter(tenant_id.eq(split_tenant_id.to_string()))
1095 0 : .filter(shard_count.eq(old_shard_count.literal() as i32))
1096 0 : .execute(conn)
1097 0 : .await?;
1098 :
1099 : // Clear sharding flag
1100 0 : let updated = diesel::update(tenant_shards)
1101 0 : .filter(tenant_id.eq(split_tenant_id.to_string()))
1102 0 : .filter(shard_count.eq(new_shard_count.literal() as i32))
1103 0 : .set((splitting.eq(0),))
1104 0 : .execute(conn)
1105 0 : .await?;
1106 0 : assert!(updated == new_shard_count.count() as usize);
1107 :
1108 0 : Ok(())
1109 0 : })
1110 0 : })
1111 0 : .await
1112 0 : }
1113 :
1114 : /// Used when the remote part of a shard split failed: we will revert the database state to have only
1115 : /// the parent shards, with SplitState::Idle.
1116 0 : pub(crate) async fn abort_shard_split(
1117 0 : &self,
1118 0 : split_tenant_id: TenantId,
1119 0 : new_shard_count: ShardCount,
1120 0 : ) -> DatabaseResult<AbortShardSplitStatus> {
1121 : use crate::schema::tenant_shards::dsl::*;
1122 0 : self.with_measured_conn(DatabaseOperation::AbortShardSplit, move |conn| {
1123 0 : Box::pin(async move {
1124 : // Clear the splitting state on parent shards
1125 0 : let updated = diesel::update(tenant_shards)
1126 0 : .filter(tenant_id.eq(split_tenant_id.to_string()))
1127 0 : .filter(shard_count.ne(new_shard_count.literal() as i32))
1128 0 : .set((splitting.eq(0),))
1129 0 : .execute(conn)
1130 0 : .await?;
1131 :
1132 : // Parent shards are already gone: we cannot abort.
1133 0 : if updated == 0 {
1134 0 : return Ok(AbortShardSplitStatus::Complete);
1135 0 : }
1136 :
1137 : // Sanity check: if parent shards were present, their cardinality should
1138 : // be less than the number of child shards.
1139 0 : if updated >= new_shard_count.count() as usize {
1140 0 : return Err(DatabaseError::Logical(format!(
1141 0 : "Unexpected parent shard count {updated} while aborting split to \
1142 0 : count {new_shard_count:?} on tenant {split_tenant_id}"
1143 0 : )));
1144 0 : }
1145 :
1146 : // Erase child shards
1147 0 : diesel::delete(tenant_shards)
1148 0 : .filter(tenant_id.eq(split_tenant_id.to_string()))
1149 0 : .filter(shard_count.eq(new_shard_count.literal() as i32))
1150 0 : .execute(conn)
1151 0 : .await?;
1152 :
1153 0 : Ok(AbortShardSplitStatus::Aborted)
1154 0 : })
1155 0 : })
1156 0 : .await
1157 0 : }
1158 :
1159 : /// Stores all the latest metadata health updates durably. Updates existing entry on conflict.
1160 : ///
1161 : /// **Correctness:** `metadata_health_updates` should all belong the tenant shards managed by the storage controller.
1162 : #[allow(dead_code)]
1163 0 : pub(crate) async fn update_metadata_health_records(
1164 0 : &self,
1165 0 : healthy_records: Vec<MetadataHealthPersistence>,
1166 0 : unhealthy_records: Vec<MetadataHealthPersistence>,
1167 0 : now: chrono::DateTime<chrono::Utc>,
1168 0 : ) -> DatabaseResult<()> {
1169 : use crate::schema::metadata_health::dsl::*;
1170 :
1171 0 : let healthy_records = healthy_records.as_slice();
1172 0 : let unhealthy_records = unhealthy_records.as_slice();
1173 0 : self.with_measured_conn(DatabaseOperation::UpdateMetadataHealth, move |conn| {
1174 0 : Box::pin(async move {
1175 0 : diesel::insert_into(metadata_health)
1176 0 : .values(healthy_records)
1177 0 : .on_conflict((tenant_id, shard_number, shard_count))
1178 0 : .do_update()
1179 0 : .set((healthy.eq(true), last_scrubbed_at.eq(now)))
1180 0 : .execute(conn)
1181 0 : .await?;
1182 :
1183 0 : diesel::insert_into(metadata_health)
1184 0 : .values(unhealthy_records)
1185 0 : .on_conflict((tenant_id, shard_number, shard_count))
1186 0 : .do_update()
1187 0 : .set((healthy.eq(false), last_scrubbed_at.eq(now)))
1188 0 : .execute(conn)
1189 0 : .await?;
1190 0 : Ok(())
1191 0 : })
1192 0 : })
1193 0 : .await
1194 0 : }
1195 :
1196 : /// Lists all the metadata health records.
1197 : #[allow(dead_code)]
1198 0 : pub(crate) async fn list_metadata_health_records(
1199 0 : &self,
1200 0 : ) -> DatabaseResult<Vec<MetadataHealthPersistence>> {
1201 0 : self.with_measured_conn(DatabaseOperation::ListMetadataHealth, move |conn| {
1202 0 : Box::pin(async {
1203 0 : Ok(crate::schema::metadata_health::table
1204 0 : .load::<MetadataHealthPersistence>(conn)
1205 0 : .await?)
1206 0 : })
1207 0 : })
1208 0 : .await
1209 0 : }
1210 :
1211 : /// Lists all the metadata health records that is unhealthy.
1212 : #[allow(dead_code)]
1213 0 : pub(crate) async fn list_unhealthy_metadata_health_records(
1214 0 : &self,
1215 0 : ) -> DatabaseResult<Vec<MetadataHealthPersistence>> {
1216 : use crate::schema::metadata_health::dsl::*;
1217 0 : self.with_measured_conn(
1218 0 : DatabaseOperation::ListMetadataHealthUnhealthy,
1219 0 : move |conn| {
1220 0 : Box::pin(async {
1221 : DatabaseResult::Ok(
1222 0 : crate::schema::metadata_health::table
1223 0 : .filter(healthy.eq(false))
1224 0 : .load::<MetadataHealthPersistence>(conn)
1225 0 : .await?,
1226 : )
1227 0 : })
1228 0 : },
1229 : )
1230 0 : .await
1231 0 : }
1232 :
1233 : /// Lists all the metadata health records that have not been updated since an `earlier` time.
1234 : #[allow(dead_code)]
1235 0 : pub(crate) async fn list_outdated_metadata_health_records(
1236 0 : &self,
1237 0 : earlier: chrono::DateTime<chrono::Utc>,
1238 0 : ) -> DatabaseResult<Vec<MetadataHealthPersistence>> {
1239 : use crate::schema::metadata_health::dsl::*;
1240 :
1241 0 : self.with_measured_conn(DatabaseOperation::ListMetadataHealthOutdated, move |conn| {
1242 0 : Box::pin(async move {
1243 0 : let query = metadata_health.filter(last_scrubbed_at.lt(earlier));
1244 0 : let res = query.load::<MetadataHealthPersistence>(conn).await?;
1245 :
1246 0 : Ok(res)
1247 0 : })
1248 0 : })
1249 0 : .await
1250 0 : }
1251 :
1252 : /// Get the current entry from the `leader` table if one exists.
1253 : /// It is an error for the table to contain more than one entry.
1254 0 : pub(crate) async fn get_leader(&self) -> DatabaseResult<Option<ControllerPersistence>> {
1255 0 : let mut leader: Vec<ControllerPersistence> = self
1256 0 : .with_measured_conn(DatabaseOperation::GetLeader, move |conn| {
1257 0 : Box::pin(async move {
1258 0 : Ok(crate::schema::controllers::table
1259 0 : .load::<ControllerPersistence>(conn)
1260 0 : .await?)
1261 0 : })
1262 0 : })
1263 0 : .await?;
1264 :
1265 0 : if leader.len() > 1 {
1266 0 : return Err(DatabaseError::Logical(format!(
1267 0 : "More than one entry present in the leader table: {leader:?}"
1268 0 : )));
1269 0 : }
1270 :
1271 0 : Ok(leader.pop())
1272 0 : }
1273 :
1274 : /// Update the new leader with compare-exchange semantics. If `prev` does not
1275 : /// match the current leader entry, then the update is treated as a failure.
1276 : /// When `prev` is not specified, the update is forced.
1277 0 : pub(crate) async fn update_leader(
1278 0 : &self,
1279 0 : prev: Option<ControllerPersistence>,
1280 0 : new: ControllerPersistence,
1281 0 : ) -> DatabaseResult<()> {
1282 : use crate::schema::controllers::dsl::*;
1283 :
1284 0 : let updated = self
1285 0 : .with_measured_conn(DatabaseOperation::UpdateLeader, move |conn| {
1286 0 : let prev = prev.clone();
1287 0 : let new = new.clone();
1288 0 : Box::pin(async move {
1289 0 : let updated = match &prev {
1290 0 : Some(prev) => {
1291 0 : diesel::update(controllers)
1292 0 : .filter(address.eq(prev.address.clone()))
1293 0 : .filter(started_at.eq(prev.started_at))
1294 0 : .set((
1295 0 : address.eq(new.address.clone()),
1296 0 : started_at.eq(new.started_at),
1297 0 : ))
1298 0 : .execute(conn)
1299 0 : .await?
1300 : }
1301 : None => {
1302 0 : diesel::insert_into(controllers)
1303 0 : .values(new.clone())
1304 0 : .execute(conn)
1305 0 : .await?
1306 : }
1307 : };
1308 :
1309 0 : Ok(updated)
1310 0 : })
1311 0 : })
1312 0 : .await?;
1313 :
1314 0 : if updated == 0 {
1315 0 : return Err(DatabaseError::Logical(
1316 0 : "Leader table update failed".to_string(),
1317 0 : ));
1318 0 : }
1319 :
1320 0 : Ok(())
1321 0 : }
1322 :
1323 : /// At startup, populate the list of nodes which our shards may be placed on
1324 0 : pub(crate) async fn list_safekeepers(&self) -> DatabaseResult<Vec<SafekeeperPersistence>> {
1325 0 : let safekeepers: Vec<SafekeeperPersistence> = self
1326 0 : .with_measured_conn(DatabaseOperation::ListNodes, move |conn| {
1327 0 : Box::pin(async move {
1328 0 : Ok(crate::schema::safekeepers::table
1329 0 : .load::<SafekeeperPersistence>(conn)
1330 0 : .await?)
1331 0 : })
1332 0 : })
1333 0 : .await?;
1334 :
1335 0 : tracing::info!("list_safekeepers: loaded {} nodes", safekeepers.len());
1336 :
1337 0 : Ok(safekeepers)
1338 0 : }
1339 :
1340 0 : pub(crate) async fn safekeeper_upsert(
1341 0 : &self,
1342 0 : record: SafekeeperUpsert,
1343 0 : ) -> Result<(), DatabaseError> {
1344 : use crate::schema::safekeepers::dsl::*;
1345 :
1346 0 : self.with_conn(move |conn| {
1347 0 : let record = record.clone();
1348 0 : Box::pin(async move {
1349 0 : let bind = record
1350 0 : .as_insert_or_update()
1351 0 : .map_err(|e| DatabaseError::Logical(format!("{e}")))?;
1352 :
1353 0 : let inserted_updated = diesel::insert_into(safekeepers)
1354 0 : .values(&bind)
1355 0 : .on_conflict(id)
1356 0 : .do_update()
1357 0 : .set(&bind)
1358 0 : .execute(conn)
1359 0 : .await?;
1360 :
1361 0 : if inserted_updated != 1 {
1362 0 : return Err(DatabaseError::Logical(format!(
1363 0 : "unexpected number of rows ({inserted_updated})"
1364 0 : )));
1365 0 : }
1366 :
1367 0 : Ok(())
1368 0 : })
1369 0 : })
1370 0 : .await
1371 0 : }
1372 :
1373 0 : pub(crate) async fn set_safekeeper_scheduling_policy(
1374 0 : &self,
1375 0 : id_: i64,
1376 0 : scheduling_policy_: SkSchedulingPolicy,
1377 0 : ) -> Result<(), DatabaseError> {
1378 : use crate::schema::safekeepers::dsl::*;
1379 :
1380 0 : self.with_conn(move |conn| {
1381 0 : Box::pin(async move {
1382 : #[derive(Insertable, AsChangeset)]
1383 : #[diesel(table_name = crate::schema::safekeepers)]
1384 : struct UpdateSkSchedulingPolicy<'a> {
1385 : id: i64,
1386 : scheduling_policy: &'a str,
1387 : }
1388 0 : let scheduling_policy_ = String::from(scheduling_policy_);
1389 :
1390 0 : let rows_affected = diesel::update(safekeepers.filter(id.eq(id_)))
1391 0 : .set(scheduling_policy.eq(scheduling_policy_))
1392 0 : .execute(conn)
1393 0 : .await?;
1394 :
1395 0 : if rows_affected != 1 {
1396 0 : return Err(DatabaseError::Logical(format!(
1397 0 : "unexpected number of rows ({rows_affected})",
1398 0 : )));
1399 0 : }
1400 :
1401 0 : Ok(())
1402 0 : })
1403 0 : })
1404 0 : .await
1405 0 : }
1406 :
1407 : /// Activate the given safekeeper, ensuring that there is no TOCTOU.
1408 : /// Returns `Some` if the safekeeper has indeed been activating (or already active). Other states return `None`.
1409 0 : pub(crate) async fn activate_safekeeper(&self, id_: i64) -> Result<Option<()>, DatabaseError> {
1410 : use crate::schema::safekeepers::dsl::*;
1411 :
1412 0 : self.with_conn(move |conn| {
1413 0 : Box::pin(async move {
1414 : #[derive(Insertable, AsChangeset)]
1415 : #[diesel(table_name = crate::schema::safekeepers)]
1416 : struct UpdateSkSchedulingPolicy<'a> {
1417 : id: i64,
1418 : scheduling_policy: &'a str,
1419 : }
1420 0 : let scheduling_policy_active = String::from(SkSchedulingPolicy::Active);
1421 0 : let scheduling_policy_activating = String::from(SkSchedulingPolicy::Activating);
1422 :
1423 0 : let rows_affected = diesel::update(
1424 0 : safekeepers.filter(id.eq(id_)).filter(
1425 0 : scheduling_policy
1426 0 : .eq(scheduling_policy_activating)
1427 0 : .or(scheduling_policy.eq(&scheduling_policy_active)),
1428 0 : ),
1429 0 : )
1430 0 : .set(scheduling_policy.eq(&scheduling_policy_active))
1431 0 : .execute(conn)
1432 0 : .await?;
1433 :
1434 0 : if rows_affected == 0 {
1435 0 : return Ok(Some(()));
1436 0 : }
1437 0 : if rows_affected != 1 {
1438 0 : return Err(DatabaseError::Logical(format!(
1439 0 : "unexpected number of rows ({rows_affected})",
1440 0 : )));
1441 0 : }
1442 :
1443 0 : Ok(Some(()))
1444 0 : })
1445 0 : })
1446 0 : .await
1447 0 : }
1448 :
1449 : /// Persist timeline. Returns if the timeline was newly inserted. If it wasn't, we haven't done any writes.
1450 0 : pub(crate) async fn insert_timeline(&self, entry: TimelinePersistence) -> DatabaseResult<bool> {
1451 : use crate::schema::timelines;
1452 :
1453 0 : let entry = &entry;
1454 0 : self.with_measured_conn(DatabaseOperation::InsertTimeline, move |conn| {
1455 0 : Box::pin(async move {
1456 0 : let inserted_updated = diesel::insert_into(timelines::table)
1457 0 : .values(entry)
1458 0 : .on_conflict((timelines::tenant_id, timelines::timeline_id))
1459 0 : .do_nothing()
1460 0 : .execute(conn)
1461 0 : .await?;
1462 :
1463 0 : match inserted_updated {
1464 0 : 0 => Ok(false),
1465 0 : 1 => Ok(true),
1466 0 : _ => Err(DatabaseError::Logical(format!(
1467 0 : "unexpected number of rows ({inserted_updated})"
1468 0 : ))),
1469 : }
1470 0 : })
1471 0 : })
1472 0 : .await
1473 0 : }
1474 :
1475 : /// Update an already present timeline.
1476 : /// VERY UNSAFE FUNCTION: this overrides in-progress migrations. Don't use this unless neccessary.
1477 0 : pub(crate) async fn update_timeline_unsafe(
1478 0 : &self,
1479 0 : entry: TimelineUpdate,
1480 0 : ) -> DatabaseResult<bool> {
1481 : use crate::schema::timelines;
1482 :
1483 0 : let entry = &entry;
1484 0 : self.with_measured_conn(DatabaseOperation::UpdateTimeline, move |conn| {
1485 0 : Box::pin(async move {
1486 0 : let inserted_updated = diesel::update(timelines::table)
1487 0 : .filter(timelines::tenant_id.eq(&entry.tenant_id))
1488 0 : .filter(timelines::timeline_id.eq(&entry.timeline_id))
1489 0 : .set(entry)
1490 0 : .execute(conn)
1491 0 : .await?;
1492 :
1493 0 : match inserted_updated {
1494 0 : 0 => Ok(false),
1495 0 : 1 => Ok(true),
1496 0 : _ => Err(DatabaseError::Logical(format!(
1497 0 : "unexpected number of rows ({inserted_updated})"
1498 0 : ))),
1499 : }
1500 0 : })
1501 0 : })
1502 0 : .await
1503 0 : }
1504 :
1505 : /// Update timeline membership configuration in the database.
1506 : /// Perform a compare-and-swap (CAS) operation on the timeline's generation.
1507 : /// The `new_generation` must be the next (+1) generation after the one in the database.
1508 : /// Also inserts reconcile_requests to safekeeper_timeline_pending_ops table in the same
1509 : /// transaction.
1510 0 : pub(crate) async fn update_timeline_membership(
1511 0 : &self,
1512 0 : tenant_id: TenantId,
1513 0 : timeline_id: TimelineId,
1514 0 : new_generation: SafekeeperGeneration,
1515 0 : sk_set: &[NodeId],
1516 0 : new_sk_set: Option<&[NodeId]>,
1517 0 : reconcile_requests: &[TimelinePendingOpPersistence],
1518 0 : ) -> DatabaseResult<()> {
1519 : use crate::schema::safekeeper_timeline_pending_ops as stpo;
1520 : use crate::schema::timelines;
1521 : use diesel::query_dsl::methods::FilterDsl;
1522 :
1523 0 : let prev_generation = new_generation.previous().unwrap();
1524 :
1525 0 : let tenant_id = &tenant_id;
1526 0 : let timeline_id = &timeline_id;
1527 0 : self.with_measured_conn(DatabaseOperation::UpdateTimelineMembership, move |conn| {
1528 0 : Box::pin(async move {
1529 0 : let updated = diesel::update(timelines::table)
1530 0 : .filter(timelines::tenant_id.eq(&tenant_id.to_string()))
1531 0 : .filter(timelines::timeline_id.eq(&timeline_id.to_string()))
1532 0 : .filter(timelines::generation.eq(prev_generation.into_inner() as i32))
1533 0 : .set((
1534 0 : timelines::generation.eq(new_generation.into_inner() as i32),
1535 0 : timelines::sk_set
1536 0 : .eq(sk_set.iter().map(|id| id.0 as i64).collect::<Vec<_>>()),
1537 0 : timelines::new_sk_set.eq(new_sk_set
1538 0 : .map(|set| set.iter().map(|id| id.0 as i64).collect::<Vec<_>>())),
1539 : ))
1540 0 : .execute(conn)
1541 0 : .await?;
1542 :
1543 0 : match updated {
1544 : 0 => {
1545 : // TODO(diko): It makes sense to select the current generation
1546 : // and include it in the error message for better debuggability.
1547 0 : return Err(DatabaseError::Cas(
1548 0 : "Failed to update membership configuration".to_string(),
1549 0 : ));
1550 : }
1551 0 : 1 => {}
1552 : _ => {
1553 0 : return Err(DatabaseError::Logical(format!(
1554 0 : "unexpected number of rows ({updated})"
1555 0 : )));
1556 : }
1557 : };
1558 :
1559 0 : for req in reconcile_requests {
1560 0 : let inserted_updated = diesel::insert_into(stpo::table)
1561 0 : .values(req)
1562 0 : .on_conflict((stpo::tenant_id, stpo::timeline_id, stpo::sk_id))
1563 0 : .do_update()
1564 0 : .set(req)
1565 0 : .filter(stpo::generation.lt(req.generation))
1566 0 : .execute(conn)
1567 0 : .await?;
1568 :
1569 0 : if inserted_updated > 1 {
1570 0 : return Err(DatabaseError::Logical(format!(
1571 0 : "unexpected number of rows ({inserted_updated})"
1572 0 : )));
1573 0 : }
1574 : }
1575 :
1576 0 : Ok(())
1577 0 : })
1578 0 : })
1579 0 : .await
1580 0 : }
1581 :
1582 : /// Update the cplane notified generation for a timeline.
1583 : /// Perform a compare-and-swap (CAS) operation on the timeline's cplane notified generation.
1584 : /// The update will fail if the specified generation is less than the cplane notified generation
1585 : /// in the database.
1586 0 : pub(crate) async fn update_cplane_notified_generation(
1587 0 : &self,
1588 0 : tenant_id: TenantId,
1589 0 : timeline_id: TimelineId,
1590 0 : generation: SafekeeperGeneration,
1591 0 : ) -> DatabaseResult<()> {
1592 : use crate::schema::timelines::dsl;
1593 :
1594 0 : let tenant_id = &tenant_id;
1595 0 : let timeline_id = &timeline_id;
1596 0 : self.with_measured_conn(
1597 0 : DatabaseOperation::UpdateCplaneNotifiedGeneration,
1598 0 : move |conn| {
1599 0 : Box::pin(async move {
1600 0 : let updated = diesel::update(dsl::timelines)
1601 0 : .filter(dsl::tenant_id.eq(&tenant_id.to_string()))
1602 0 : .filter(dsl::timeline_id.eq(&timeline_id.to_string()))
1603 0 : .filter(dsl::cplane_notified_generation.le(generation.into_inner() as i32))
1604 0 : .set(dsl::cplane_notified_generation.eq(generation.into_inner() as i32))
1605 0 : .execute(conn)
1606 0 : .await?;
1607 :
1608 0 : match updated {
1609 0 : 0 => Err(DatabaseError::Cas(
1610 0 : "Failed to update cplane notified generation".to_string(),
1611 0 : )),
1612 0 : 1 => Ok(()),
1613 0 : _ => Err(DatabaseError::Logical(format!(
1614 0 : "unexpected number of rows ({updated})"
1615 0 : ))),
1616 : }
1617 0 : })
1618 0 : },
1619 : )
1620 0 : .await
1621 0 : }
1622 :
1623 : /// Update the sk set notified generation for a timeline.
1624 : /// Perform a compare-and-swap (CAS) operation on the timeline's sk set notified generation.
1625 : /// The update will fail if the specified generation is less than the sk set notified generation
1626 : /// in the database.
1627 0 : pub(crate) async fn update_sk_set_notified_generation(
1628 0 : &self,
1629 0 : tenant_id: TenantId,
1630 0 : timeline_id: TimelineId,
1631 0 : generation: SafekeeperGeneration,
1632 0 : ) -> DatabaseResult<()> {
1633 : use crate::schema::timelines::dsl;
1634 :
1635 0 : let tenant_id = &tenant_id;
1636 0 : let timeline_id = &timeline_id;
1637 0 : self.with_measured_conn(
1638 0 : DatabaseOperation::UpdateSkSetNotifiedGeneration,
1639 0 : move |conn| {
1640 0 : Box::pin(async move {
1641 0 : let updated = diesel::update(dsl::timelines)
1642 0 : .filter(dsl::tenant_id.eq(&tenant_id.to_string()))
1643 0 : .filter(dsl::timeline_id.eq(&timeline_id.to_string()))
1644 0 : .filter(dsl::sk_set_notified_generation.le(generation.into_inner() as i32))
1645 0 : .set(dsl::sk_set_notified_generation.eq(generation.into_inner() as i32))
1646 0 : .execute(conn)
1647 0 : .await?;
1648 :
1649 0 : match updated {
1650 0 : 0 => Err(DatabaseError::Cas(
1651 0 : "Failed to update sk set notified generation".to_string(),
1652 0 : )),
1653 0 : 1 => Ok(()),
1654 0 : _ => Err(DatabaseError::Logical(format!(
1655 0 : "unexpected number of rows ({updated})"
1656 0 : ))),
1657 : }
1658 0 : })
1659 0 : },
1660 : )
1661 0 : .await
1662 0 : }
1663 :
1664 : /// Load timeline from db. Returns `None` if not present.
1665 0 : pub(crate) async fn get_timeline(
1666 0 : &self,
1667 0 : tenant_id: TenantId,
1668 0 : timeline_id: TimelineId,
1669 0 : ) -> DatabaseResult<Option<TimelinePersistence>> {
1670 : use crate::schema::timelines::dsl;
1671 :
1672 0 : let tenant_id = &tenant_id;
1673 0 : let timeline_id = &timeline_id;
1674 0 : let timeline_from_db = self
1675 0 : .with_measured_conn(DatabaseOperation::GetTimeline, move |conn| {
1676 0 : Box::pin(async move {
1677 0 : let mut from_db: Vec<TimelineFromDb> = dsl::timelines
1678 0 : .filter(
1679 0 : dsl::tenant_id
1680 0 : .eq(&tenant_id.to_string())
1681 0 : .and(dsl::timeline_id.eq(&timeline_id.to_string())),
1682 0 : )
1683 0 : .load(conn)
1684 0 : .await?;
1685 0 : if from_db.is_empty() {
1686 0 : return Ok(None);
1687 0 : }
1688 0 : if from_db.len() != 1 {
1689 0 : return Err(DatabaseError::Logical(format!(
1690 0 : "unexpected number of rows ({})",
1691 0 : from_db.len()
1692 0 : )));
1693 0 : }
1694 :
1695 0 : Ok(Some(from_db.pop().unwrap().into_persistence()))
1696 0 : })
1697 0 : })
1698 0 : .await?;
1699 :
1700 0 : Ok(timeline_from_db)
1701 0 : }
1702 :
1703 : /// Set `delete_at` for the given timeline
1704 0 : pub(crate) async fn timeline_set_deleted_at(
1705 0 : &self,
1706 0 : tenant_id: TenantId,
1707 0 : timeline_id: TimelineId,
1708 0 : ) -> DatabaseResult<()> {
1709 : use crate::schema::timelines;
1710 :
1711 0 : let deletion_time = chrono::Local::now().to_utc();
1712 0 : self.with_measured_conn(DatabaseOperation::InsertTimeline, move |conn| {
1713 0 : Box::pin(async move {
1714 0 : let updated = diesel::update(timelines::table)
1715 0 : .filter(timelines::tenant_id.eq(tenant_id.to_string()))
1716 0 : .filter(timelines::timeline_id.eq(timeline_id.to_string()))
1717 0 : .set(timelines::deleted_at.eq(Some(deletion_time)))
1718 0 : .execute(conn)
1719 0 : .await?;
1720 :
1721 0 : match updated {
1722 0 : 0 => Ok(()),
1723 0 : 1 => Ok(()),
1724 0 : _ => Err(DatabaseError::Logical(format!(
1725 0 : "unexpected number of rows ({updated})"
1726 0 : ))),
1727 : }
1728 0 : })
1729 0 : })
1730 0 : .await
1731 0 : }
1732 :
1733 : /// Load timeline from db. Returns `None` if not present.
1734 : ///
1735 : /// Only works if `deleted_at` is set, so you should call [`Self::timeline_set_deleted_at`] before.
1736 0 : pub(crate) async fn delete_timeline(
1737 0 : &self,
1738 0 : tenant_id: TenantId,
1739 0 : timeline_id: TimelineId,
1740 0 : ) -> DatabaseResult<()> {
1741 : use crate::schema::timelines::dsl;
1742 :
1743 0 : let tenant_id = &tenant_id;
1744 0 : let timeline_id = &timeline_id;
1745 0 : self.with_measured_conn(DatabaseOperation::GetTimeline, move |conn| {
1746 0 : Box::pin(async move {
1747 0 : diesel::delete(dsl::timelines)
1748 0 : .filter(dsl::tenant_id.eq(&tenant_id.to_string()))
1749 0 : .filter(dsl::timeline_id.eq(&timeline_id.to_string()))
1750 0 : .filter(dsl::deleted_at.is_not_null())
1751 0 : .execute(conn)
1752 0 : .await?;
1753 0 : Ok(())
1754 0 : })
1755 0 : })
1756 0 : .await?;
1757 :
1758 0 : Ok(())
1759 0 : }
1760 :
1761 : /// Loads a list of all timelines from database.
1762 0 : pub(crate) async fn list_timelines_for_tenant(
1763 0 : &self,
1764 0 : tenant_id: TenantId,
1765 0 : ) -> DatabaseResult<Vec<TimelinePersistence>> {
1766 : use crate::schema::timelines::dsl;
1767 :
1768 0 : let tenant_id = &tenant_id;
1769 0 : let timelines = self
1770 0 : .with_measured_conn(DatabaseOperation::GetTimeline, move |conn| {
1771 0 : Box::pin(async move {
1772 0 : let timelines: Vec<TimelineFromDb> = dsl::timelines
1773 0 : .filter(dsl::tenant_id.eq(&tenant_id.to_string()))
1774 0 : .load(conn)
1775 0 : .await?;
1776 0 : Ok(timelines)
1777 0 : })
1778 0 : })
1779 0 : .await?;
1780 :
1781 0 : let timelines = timelines
1782 0 : .into_iter()
1783 0 : .map(TimelineFromDb::into_persistence)
1784 0 : .collect();
1785 0 : Ok(timelines)
1786 0 : }
1787 :
1788 : /// Persist pending op. Returns if it was newly inserted. If it wasn't, we haven't done any writes.
1789 0 : pub(crate) async fn insert_pending_op(
1790 0 : &self,
1791 0 : entry: TimelinePendingOpPersistence,
1792 0 : ) -> DatabaseResult<bool> {
1793 : use crate::schema::safekeeper_timeline_pending_ops as skpo;
1794 : // This overrides the `filter` fn used in other functions, so contain the mayhem via a function-local use
1795 : use diesel::query_dsl::methods::FilterDsl;
1796 :
1797 0 : let entry = &entry;
1798 0 : self.with_measured_conn(DatabaseOperation::InsertTimelineReconcile, move |conn| {
1799 0 : Box::pin(async move {
1800 : // For simplicity it makes sense to keep only the last operation
1801 : // per (tenant, timeline, sk) tuple: if we migrated a timeline
1802 : // from node and adding it back it is not necessary to remove
1803 : // data on it. Hence, generation is not part of primary key and
1804 : // we override any rows with lower generations here.
1805 0 : let inserted_updated = diesel::insert_into(skpo::table)
1806 0 : .values(entry)
1807 0 : .on_conflict((skpo::tenant_id, skpo::timeline_id, skpo::sk_id))
1808 0 : .do_update()
1809 0 : .set(entry)
1810 0 : .filter(skpo::generation.lt(entry.generation))
1811 0 : .execute(conn)
1812 0 : .await?;
1813 :
1814 0 : match inserted_updated {
1815 0 : 0 => Ok(false),
1816 0 : 1 => Ok(true),
1817 0 : _ => Err(DatabaseError::Logical(format!(
1818 0 : "unexpected number of rows ({inserted_updated})"
1819 0 : ))),
1820 : }
1821 0 : })
1822 0 : })
1823 0 : .await
1824 0 : }
1825 : /// Remove persisted pending op.
1826 0 : pub(crate) async fn remove_pending_op(
1827 0 : &self,
1828 0 : tenant_id: TenantId,
1829 0 : timeline_id: Option<TimelineId>,
1830 0 : sk_id: NodeId,
1831 0 : generation: u32,
1832 0 : ) -> DatabaseResult<()> {
1833 : use crate::schema::safekeeper_timeline_pending_ops::dsl;
1834 :
1835 0 : let tenant_id = &tenant_id;
1836 0 : let timeline_id = &timeline_id;
1837 0 : self.with_measured_conn(DatabaseOperation::RemoveTimelineReconcile, move |conn| {
1838 0 : let timeline_id_str = timeline_id.map(|tid| tid.to_string()).unwrap_or_default();
1839 0 : Box::pin(async move {
1840 0 : diesel::delete(dsl::safekeeper_timeline_pending_ops)
1841 0 : .filter(dsl::tenant_id.eq(tenant_id.to_string()))
1842 0 : .filter(dsl::timeline_id.eq(timeline_id_str))
1843 0 : .filter(dsl::sk_id.eq(sk_id.0 as i64))
1844 0 : .filter(dsl::generation.eq(generation as i32))
1845 0 : .execute(conn)
1846 0 : .await?;
1847 0 : Ok(())
1848 0 : })
1849 0 : })
1850 0 : .await
1851 0 : }
1852 :
1853 : /// Load pending operations from db, joined together with timeline data.
1854 0 : pub(crate) async fn list_pending_ops_with_timelines(
1855 0 : &self,
1856 0 : ) -> DatabaseResult<Vec<(TimelinePendingOpPersistence, Option<TimelinePersistence>)>> {
1857 : use crate::schema::safekeeper_timeline_pending_ops::dsl;
1858 : use crate::schema::timelines;
1859 :
1860 0 : let timeline_from_db = self
1861 0 : .with_measured_conn(
1862 0 : DatabaseOperation::ListTimelineReconcileStartup,
1863 0 : move |conn| {
1864 0 : Box::pin(async move {
1865 0 : let from_db: Vec<(TimelinePendingOpPersistence, Option<TimelineFromDb>)> =
1866 0 : dsl::safekeeper_timeline_pending_ops
1867 0 : .left_join(
1868 0 : timelines::table.on(timelines::tenant_id
1869 0 : .eq(dsl::tenant_id)
1870 0 : .and(timelines::timeline_id.eq(dsl::timeline_id))),
1871 0 : )
1872 0 : .select((
1873 0 : TimelinePendingOpPersistence::as_select(),
1874 0 : Option::<TimelineFromDb>::as_select(),
1875 0 : ))
1876 0 : .load(conn)
1877 0 : .await?;
1878 0 : Ok(from_db)
1879 0 : })
1880 0 : },
1881 : )
1882 0 : .await?;
1883 :
1884 0 : Ok(timeline_from_db
1885 0 : .into_iter()
1886 0 : .map(|(op, tl_opt)| (op, tl_opt.map(|tl_opt| tl_opt.into_persistence())))
1887 0 : .collect())
1888 0 : }
1889 : /// List pending operations for a given timeline (including tenant-global ones)
1890 0 : pub(crate) async fn list_pending_ops_for_timeline(
1891 0 : &self,
1892 0 : tenant_id: TenantId,
1893 0 : timeline_id: TimelineId,
1894 0 : ) -> DatabaseResult<Vec<TimelinePendingOpPersistence>> {
1895 : use crate::schema::safekeeper_timeline_pending_ops::dsl;
1896 :
1897 0 : let timelines_from_db = self
1898 0 : .with_measured_conn(DatabaseOperation::ListTimelineReconcile, move |conn| {
1899 0 : Box::pin(async move {
1900 0 : let from_db: Vec<TimelinePendingOpPersistence> =
1901 0 : dsl::safekeeper_timeline_pending_ops
1902 0 : .filter(dsl::tenant_id.eq(tenant_id.to_string()))
1903 0 : .filter(
1904 0 : dsl::timeline_id
1905 0 : .eq(timeline_id.to_string())
1906 0 : .or(dsl::timeline_id.eq("")),
1907 0 : )
1908 0 : .load(conn)
1909 0 : .await?;
1910 0 : Ok(from_db)
1911 0 : })
1912 0 : })
1913 0 : .await?;
1914 :
1915 0 : Ok(timelines_from_db)
1916 0 : }
1917 :
1918 : /// Delete all pending ops for the given timeline.
1919 : ///
1920 : /// Use this only at timeline deletion, otherwise use generation based APIs
1921 0 : pub(crate) async fn remove_pending_ops_for_timeline(
1922 0 : &self,
1923 0 : tenant_id: TenantId,
1924 0 : timeline_id: Option<TimelineId>,
1925 0 : ) -> DatabaseResult<()> {
1926 : use crate::schema::safekeeper_timeline_pending_ops::dsl;
1927 :
1928 0 : let tenant_id = &tenant_id;
1929 0 : let timeline_id = &timeline_id;
1930 0 : self.with_measured_conn(DatabaseOperation::RemoveTimelineReconcile, move |conn| {
1931 0 : let timeline_id_str = timeline_id.map(|tid| tid.to_string()).unwrap_or_default();
1932 0 : Box::pin(async move {
1933 0 : diesel::delete(dsl::safekeeper_timeline_pending_ops)
1934 0 : .filter(dsl::tenant_id.eq(tenant_id.to_string()))
1935 0 : .filter(dsl::timeline_id.eq(timeline_id_str))
1936 0 : .execute(conn)
1937 0 : .await?;
1938 0 : Ok(())
1939 0 : })
1940 0 : })
1941 0 : .await?;
1942 :
1943 0 : Ok(())
1944 0 : }
1945 :
1946 0 : pub(crate) async fn insert_timeline_import(
1947 0 : &self,
1948 0 : import: TimelineImportPersistence,
1949 0 : ) -> DatabaseResult<bool> {
1950 0 : self.with_measured_conn(DatabaseOperation::InsertTimelineImport, move |conn| {
1951 0 : Box::pin({
1952 0 : let import = import.clone();
1953 0 : async move {
1954 0 : let inserted = diesel::insert_into(crate::schema::timeline_imports::table)
1955 0 : .values(import)
1956 0 : .execute(conn)
1957 0 : .await?;
1958 0 : Ok(inserted == 1)
1959 0 : }
1960 : })
1961 0 : })
1962 0 : .await
1963 0 : }
1964 :
1965 0 : pub(crate) async fn list_timeline_imports(&self) -> DatabaseResult<Vec<TimelineImport>> {
1966 : use crate::schema::timeline_imports::dsl;
1967 0 : let persistent = self
1968 0 : .with_measured_conn(DatabaseOperation::ListTimelineImports, move |conn| {
1969 0 : Box::pin(async move {
1970 0 : let from_db: Vec<TimelineImportPersistence> =
1971 0 : dsl::timeline_imports.load(conn).await?;
1972 0 : Ok(from_db)
1973 0 : })
1974 0 : })
1975 0 : .await?;
1976 :
1977 0 : let imports: Result<Vec<TimelineImport>, _> = persistent
1978 0 : .into_iter()
1979 0 : .map(TimelineImport::from_persistent)
1980 0 : .collect();
1981 0 : match imports {
1982 0 : Ok(ok) => Ok(ok.into_iter().collect()),
1983 0 : Err(err) => Err(DatabaseError::Logical(format!(
1984 0 : "failed to deserialize import: {err}"
1985 0 : ))),
1986 : }
1987 0 : }
1988 :
1989 0 : pub(crate) async fn get_timeline_import(
1990 0 : &self,
1991 0 : tenant_id: TenantId,
1992 0 : timeline_id: TimelineId,
1993 0 : ) -> DatabaseResult<Option<TimelineImport>> {
1994 : use crate::schema::timeline_imports::dsl;
1995 0 : let persistent_import = self
1996 0 : .with_measured_conn(DatabaseOperation::ListTimelineImports, move |conn| {
1997 0 : Box::pin(async move {
1998 0 : let mut from_db: Vec<TimelineImportPersistence> = dsl::timeline_imports
1999 0 : .filter(dsl::tenant_id.eq(tenant_id.to_string()))
2000 0 : .filter(dsl::timeline_id.eq(timeline_id.to_string()))
2001 0 : .load(conn)
2002 0 : .await?;
2003 :
2004 0 : if from_db.len() > 1 {
2005 0 : return Err(DatabaseError::Logical(format!(
2006 0 : "unexpected number of rows ({})",
2007 0 : from_db.len()
2008 0 : )));
2009 0 : }
2010 :
2011 0 : Ok(from_db.pop())
2012 0 : })
2013 0 : })
2014 0 : .await?;
2015 :
2016 0 : persistent_import
2017 0 : .map(TimelineImport::from_persistent)
2018 0 : .transpose()
2019 0 : .map_err(|err| DatabaseError::Logical(format!("failed to deserialize import: {err}")))
2020 0 : }
2021 :
2022 0 : pub(crate) async fn delete_timeline_import(
2023 0 : &self,
2024 0 : tenant_id: TenantId,
2025 0 : timeline_id: TimelineId,
2026 0 : ) -> DatabaseResult<()> {
2027 : use crate::schema::timeline_imports::dsl;
2028 :
2029 0 : self.with_measured_conn(DatabaseOperation::DeleteTimelineImport, move |conn| {
2030 0 : Box::pin(async move {
2031 0 : diesel::delete(crate::schema::timeline_imports::table)
2032 0 : .filter(
2033 0 : dsl::tenant_id
2034 0 : .eq(tenant_id.to_string())
2035 0 : .and(dsl::timeline_id.eq(timeline_id.to_string())),
2036 0 : )
2037 0 : .execute(conn)
2038 0 : .await?;
2039 :
2040 0 : Ok(())
2041 0 : })
2042 0 : })
2043 0 : .await
2044 0 : }
2045 :
2046 : /// Idempotently update the status of one shard for an ongoing timeline import
2047 : ///
2048 : /// If the update was persisted to the database, then the current state of the
2049 : /// import is returned to the caller. In case of logical errors a bespoke
2050 : /// [`TimelineImportUpdateError`] instance is returned. Other database errors
2051 : /// are covered by the outer [`DatabaseError`].
2052 0 : pub(crate) async fn update_timeline_import(
2053 0 : &self,
2054 0 : tenant_shard_id: TenantShardId,
2055 0 : timeline_id: TimelineId,
2056 0 : shard_status: ShardImportStatus,
2057 0 : ) -> DatabaseResult<Result<Option<TimelineImport>, TimelineImportUpdateError>> {
2058 : use crate::schema::timeline_imports::dsl;
2059 :
2060 0 : self.with_measured_conn(DatabaseOperation::UpdateTimelineImport, move |conn| {
2061 0 : Box::pin({
2062 0 : let shard_status = shard_status.clone();
2063 0 : async move {
2064 : // Load the current state from the database
2065 0 : let mut from_db: Vec<TimelineImportPersistence> = dsl::timeline_imports
2066 0 : .filter(
2067 0 : dsl::tenant_id
2068 0 : .eq(tenant_shard_id.tenant_id.to_string())
2069 0 : .and(dsl::timeline_id.eq(timeline_id.to_string())),
2070 0 : )
2071 0 : .load(conn)
2072 0 : .await?;
2073 :
2074 0 : assert!(from_db.len() <= 1);
2075 :
2076 0 : let mut status = match from_db.pop() {
2077 0 : Some(some) => TimelineImport::from_persistent(some).unwrap(),
2078 : None => {
2079 0 : return Ok(Err(TimelineImportUpdateError::ImportNotFound {
2080 0 : tenant_id: tenant_shard_id.tenant_id,
2081 0 : timeline_id,
2082 0 : }));
2083 : }
2084 : };
2085 :
2086 : // Perform the update in-memory
2087 0 : let follow_up = match status.update(tenant_shard_id.to_index(), shard_status) {
2088 0 : Ok(ok) => ok,
2089 0 : Err(err) => {
2090 0 : return Ok(Err(err));
2091 : }
2092 : };
2093 :
2094 0 : let new_persistent = status.to_persistent();
2095 :
2096 : // Write back if required (in the same transaction)
2097 0 : match follow_up {
2098 : TimelineImportUpdateFollowUp::Persist => {
2099 0 : let updated = diesel::update(dsl::timeline_imports)
2100 0 : .filter(
2101 0 : dsl::tenant_id
2102 0 : .eq(tenant_shard_id.tenant_id.to_string())
2103 0 : .and(dsl::timeline_id.eq(timeline_id.to_string())),
2104 0 : )
2105 0 : .set(dsl::shard_statuses.eq(new_persistent.shard_statuses))
2106 0 : .execute(conn)
2107 0 : .await?;
2108 :
2109 0 : if updated != 1 {
2110 0 : return Ok(Err(TimelineImportUpdateError::ImportNotFound {
2111 0 : tenant_id: tenant_shard_id.tenant_id,
2112 0 : timeline_id,
2113 0 : }));
2114 0 : }
2115 :
2116 0 : Ok(Ok(Some(status)))
2117 : }
2118 0 : TimelineImportUpdateFollowUp::None => Ok(Ok(None)),
2119 : }
2120 0 : }
2121 : })
2122 0 : })
2123 0 : .await
2124 0 : }
2125 :
2126 0 : pub(crate) async fn is_tenant_importing_timeline(
2127 0 : &self,
2128 0 : tenant_id: TenantId,
2129 0 : ) -> DatabaseResult<bool> {
2130 : use crate::schema::timeline_imports::dsl;
2131 0 : self.with_measured_conn(DatabaseOperation::IsTenantImportingTimeline, move |conn| {
2132 0 : Box::pin(async move {
2133 0 : let imports: i64 = dsl::timeline_imports
2134 0 : .filter(dsl::tenant_id.eq(tenant_id.to_string()))
2135 0 : .count()
2136 0 : .get_result(conn)
2137 0 : .await?;
2138 :
2139 0 : Ok(imports > 0)
2140 0 : })
2141 0 : })
2142 0 : .await
2143 0 : }
2144 : }
2145 :
2146 0 : pub(crate) fn load_certs() -> anyhow::Result<Arc<rustls::RootCertStore>> {
2147 0 : let der_certs = rustls_native_certs::load_native_certs();
2148 :
2149 0 : if !der_certs.errors.is_empty() {
2150 0 : anyhow::bail!("could not parse certificates: {:?}", der_certs.errors);
2151 0 : }
2152 :
2153 0 : let mut store = rustls::RootCertStore::empty();
2154 0 : store.add_parsable_certificates(der_certs.certs);
2155 0 : Ok(Arc::new(store))
2156 0 : }
2157 :
2158 : #[derive(Debug)]
2159 : /// A verifier that accepts all certificates (but logs an error still)
2160 : struct AcceptAll(Arc<WebPkiServerVerifier>);
2161 : impl ServerCertVerifier for AcceptAll {
2162 0 : fn verify_server_cert(
2163 0 : &self,
2164 0 : end_entity: &rustls::pki_types::CertificateDer<'_>,
2165 0 : intermediates: &[rustls::pki_types::CertificateDer<'_>],
2166 0 : server_name: &rustls::pki_types::ServerName<'_>,
2167 0 : ocsp_response: &[u8],
2168 0 : now: rustls::pki_types::UnixTime,
2169 0 : ) -> Result<ServerCertVerified, rustls::Error> {
2170 0 : let r =
2171 0 : self.0
2172 0 : .verify_server_cert(end_entity, intermediates, server_name, ocsp_response, now);
2173 0 : if let Err(err) = r {
2174 0 : tracing::info!(
2175 : ?server_name,
2176 0 : "ignoring db connection TLS validation error: {err:?}"
2177 : );
2178 0 : return Ok(ServerCertVerified::assertion());
2179 0 : }
2180 0 : r
2181 0 : }
2182 0 : fn verify_tls12_signature(
2183 0 : &self,
2184 0 : message: &[u8],
2185 0 : cert: &rustls::pki_types::CertificateDer<'_>,
2186 0 : dss: &rustls::DigitallySignedStruct,
2187 0 : ) -> Result<rustls::client::danger::HandshakeSignatureValid, rustls::Error> {
2188 0 : self.0.verify_tls12_signature(message, cert, dss)
2189 0 : }
2190 0 : fn verify_tls13_signature(
2191 0 : &self,
2192 0 : message: &[u8],
2193 0 : cert: &rustls::pki_types::CertificateDer<'_>,
2194 0 : dss: &rustls::DigitallySignedStruct,
2195 0 : ) -> Result<rustls::client::danger::HandshakeSignatureValid, rustls::Error> {
2196 0 : self.0.verify_tls13_signature(message, cert, dss)
2197 0 : }
2198 0 : fn supported_verify_schemes(&self) -> Vec<rustls::SignatureScheme> {
2199 0 : self.0.supported_verify_schemes()
2200 0 : }
2201 : }
2202 :
2203 : /// Loads the root certificates and constructs a client config suitable for connecting.
2204 : /// This function is blocking.
2205 0 : fn client_config_with_root_certs() -> anyhow::Result<rustls::ClientConfig> {
2206 0 : let client_config =
2207 0 : rustls::ClientConfig::builder_with_provider(Arc::new(ring::default_provider()))
2208 0 : .with_safe_default_protocol_versions()
2209 0 : .expect("ring should support the default protocol versions");
2210 : static DO_CERT_CHECKS: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
2211 0 : let do_cert_checks =
2212 0 : DO_CERT_CHECKS.get_or_init(|| std::env::var("STORCON_DB_CERT_CHECKS").is_ok());
2213 0 : Ok(if *do_cert_checks {
2214 0 : client_config
2215 0 : .with_root_certificates(load_certs()?)
2216 0 : .with_no_client_auth()
2217 : } else {
2218 0 : let verifier = AcceptAll(
2219 0 : WebPkiServerVerifier::builder_with_provider(
2220 0 : load_certs()?,
2221 0 : Arc::new(ring::default_provider()),
2222 : )
2223 0 : .build()?,
2224 : );
2225 0 : client_config
2226 0 : .dangerous()
2227 0 : .with_custom_certificate_verifier(Arc::new(verifier))
2228 0 : .with_no_client_auth()
2229 : })
2230 0 : }
2231 :
2232 0 : fn establish_connection_rustls(config: &str) -> BoxFuture<ConnectionResult<AsyncPgConnection>> {
2233 0 : let fut = async {
2234 : // We first set up the way we want rustls to work.
2235 0 : let rustls_config = client_config_with_root_certs()
2236 0 : .map_err(|err| ConnectionError::BadConnection(format!("{err:?}")))?;
2237 0 : let tls = tokio_postgres_rustls::MakeRustlsConnect::new(rustls_config);
2238 0 : let (client, conn) = tokio_postgres::connect(config, tls)
2239 0 : .await
2240 0 : .map_err(|e| ConnectionError::BadConnection(e.to_string()))?;
2241 :
2242 0 : AsyncPgConnection::try_from_client_and_connection(client, conn).await
2243 0 : };
2244 0 : fut.boxed()
2245 0 : }
2246 :
2247 : #[cfg_attr(test, test)]
2248 1 : fn test_config_debug_censors_password() {
2249 1 : let has_pw =
2250 1 : "host=/var/lib/postgresql,localhost port=1234 user=specialuser password='NOT ALLOWED TAG'";
2251 1 : let has_pw_cfg = has_pw.parse::<tokio_postgres::Config>().unwrap();
2252 1 : assert!(format!("{has_pw_cfg:?}").contains("specialuser"));
2253 : // Ensure that the password is not leaked by the debug impl
2254 1 : assert!(!format!("{has_pw_cfg:?}").contains("NOT ALLOWED TAG"));
2255 1 : }
2256 :
2257 0 : fn log_postgres_connstr_info(config_str: &str) -> anyhow::Result<()> {
2258 0 : let config = config_str
2259 0 : .parse::<tokio_postgres::Config>()
2260 0 : .map_err(|_e| anyhow::anyhow!("Couldn't parse config str"))?;
2261 : // We use debug formatting here, and use a unit test to ensure that we don't leak the password.
2262 : // To make extra sure the test gets ran, run it every time the function is called
2263 : // (this is rather cold code, we can afford it).
2264 : #[cfg(not(test))]
2265 0 : test_config_debug_censors_password();
2266 0 : tracing::info!("database connection config: {config:?}");
2267 0 : Ok(())
2268 0 : }
2269 :
2270 : /// Parts of [`crate::tenant_shard::TenantShard`] that are stored durably
2271 : #[derive(
2272 0 : QueryableByName, Queryable, Selectable, Insertable, Serialize, Deserialize, Clone, Eq, PartialEq,
2273 : )]
2274 : #[diesel(table_name = crate::schema::tenant_shards)]
2275 : pub(crate) struct TenantShardPersistence {
2276 : #[serde(default)]
2277 : pub(crate) tenant_id: String,
2278 : #[serde(default)]
2279 : pub(crate) shard_number: i32,
2280 : #[serde(default)]
2281 : pub(crate) shard_count: i32,
2282 : #[serde(default)]
2283 : pub(crate) shard_stripe_size: i32,
2284 :
2285 : // Latest generation number: next time we attach, increment this
2286 : // and use the incremented number when attaching.
2287 : //
2288 : // Generation is only None when first onboarding a tenant, where it may
2289 : // be in PlacementPolicy::Secondary and therefore have no valid generation state.
2290 : pub(crate) generation: Option<i32>,
2291 :
2292 : // Currently attached pageserver
2293 : #[serde(rename = "pageserver")]
2294 : pub(crate) generation_pageserver: Option<i64>,
2295 :
2296 : #[serde(default)]
2297 : pub(crate) placement_policy: String,
2298 : #[serde(default)]
2299 : pub(crate) splitting: SplitState,
2300 : #[serde(default)]
2301 : pub(crate) config: String,
2302 : #[serde(default)]
2303 : pub(crate) scheduling_policy: String,
2304 :
2305 : // Hint that we should attempt to schedule this tenant shard the given
2306 : // availability zone in order to minimise the chances of cross-AZ communication
2307 : // with compute.
2308 : pub(crate) preferred_az_id: Option<String>,
2309 : }
2310 :
2311 : impl TenantShardPersistence {
2312 0 : fn get_shard_count(&self) -> Result<ShardCount, ShardConfigError> {
2313 0 : self.shard_count
2314 0 : .try_into()
2315 0 : .map(ShardCount)
2316 0 : .map_err(|_| ShardConfigError::InvalidCount)
2317 0 : }
2318 :
2319 0 : fn get_shard_number(&self) -> Result<ShardNumber, ShardConfigError> {
2320 0 : self.shard_number
2321 0 : .try_into()
2322 0 : .map(ShardNumber)
2323 0 : .map_err(|_| ShardConfigError::InvalidNumber)
2324 0 : }
2325 :
2326 0 : fn get_stripe_size(&self) -> Result<ShardStripeSize, ShardConfigError> {
2327 0 : self.shard_stripe_size
2328 0 : .try_into()
2329 0 : .map(ShardStripeSize)
2330 0 : .map_err(|_| ShardConfigError::InvalidStripeSize)
2331 0 : }
2332 :
2333 0 : pub(crate) fn get_shard_identity(&self) -> Result<ShardIdentity, ShardConfigError> {
2334 0 : if self.shard_count == 0 {
2335 : // NB: carry over the stripe size from the persisted record, to avoid consistency check
2336 : // failures if the persisted value differs from the default stripe size. The stripe size
2337 : // doesn't really matter for unsharded tenants anyway.
2338 0 : Ok(ShardIdentity::unsharded_with_stripe_size(
2339 0 : self.get_stripe_size()?,
2340 : ))
2341 : } else {
2342 0 : Ok(ShardIdentity::new(
2343 0 : self.get_shard_number()?,
2344 0 : self.get_shard_count()?,
2345 0 : self.get_stripe_size()?,
2346 0 : )?)
2347 : }
2348 0 : }
2349 :
2350 0 : pub(crate) fn get_tenant_shard_id(&self) -> anyhow::Result<TenantShardId> {
2351 : Ok(TenantShardId {
2352 0 : tenant_id: TenantId::from_str(self.tenant_id.as_str())?,
2353 0 : shard_number: self.get_shard_number()?,
2354 0 : shard_count: self.get_shard_count()?,
2355 : })
2356 0 : }
2357 : }
2358 :
2359 : /// Parts of [`crate::node::Node`] that are stored durably
2360 0 : #[derive(Serialize, Deserialize, Queryable, Selectable, Insertable, Eq, PartialEq)]
2361 : #[diesel(table_name = crate::schema::nodes)]
2362 : pub(crate) struct NodePersistence {
2363 : pub(crate) node_id: i64,
2364 : pub(crate) scheduling_policy: String,
2365 : pub(crate) listen_http_addr: String,
2366 : pub(crate) listen_http_port: i32,
2367 : pub(crate) listen_pg_addr: String,
2368 : pub(crate) listen_pg_port: i32,
2369 : pub(crate) availability_zone_id: String,
2370 : pub(crate) listen_https_port: Option<i32>,
2371 : pub(crate) lifecycle: String,
2372 : pub(crate) listen_grpc_addr: Option<String>,
2373 : pub(crate) listen_grpc_port: Option<i32>,
2374 : }
2375 :
2376 : /// Tenant metadata health status that are stored durably.
2377 0 : #[derive(Queryable, Selectable, Insertable, Serialize, Deserialize, Clone, Eq, PartialEq)]
2378 : #[diesel(table_name = crate::schema::metadata_health)]
2379 : pub(crate) struct MetadataHealthPersistence {
2380 : #[serde(default)]
2381 : pub(crate) tenant_id: String,
2382 : #[serde(default)]
2383 : pub(crate) shard_number: i32,
2384 : #[serde(default)]
2385 : pub(crate) shard_count: i32,
2386 :
2387 : pub(crate) healthy: bool,
2388 : pub(crate) last_scrubbed_at: chrono::DateTime<chrono::Utc>,
2389 : }
2390 :
2391 : impl MetadataHealthPersistence {
2392 0 : pub fn new(
2393 0 : tenant_shard_id: TenantShardId,
2394 0 : healthy: bool,
2395 0 : last_scrubbed_at: chrono::DateTime<chrono::Utc>,
2396 0 : ) -> Self {
2397 0 : let tenant_id = tenant_shard_id.tenant_id.to_string();
2398 0 : let shard_number = tenant_shard_id.shard_number.0 as i32;
2399 0 : let shard_count = tenant_shard_id.shard_count.literal() as i32;
2400 :
2401 0 : MetadataHealthPersistence {
2402 0 : tenant_id,
2403 0 : shard_number,
2404 0 : shard_count,
2405 0 : healthy,
2406 0 : last_scrubbed_at,
2407 0 : }
2408 0 : }
2409 :
2410 : #[allow(dead_code)]
2411 0 : pub(crate) fn get_tenant_shard_id(&self) -> Result<TenantShardId, hex::FromHexError> {
2412 : Ok(TenantShardId {
2413 0 : tenant_id: TenantId::from_str(self.tenant_id.as_str())?,
2414 0 : shard_number: ShardNumber(self.shard_number as u8),
2415 0 : shard_count: ShardCount::new(self.shard_count as u8),
2416 : })
2417 0 : }
2418 : }
2419 :
2420 : impl From<MetadataHealthPersistence> for MetadataHealthRecord {
2421 0 : fn from(value: MetadataHealthPersistence) -> Self {
2422 0 : MetadataHealthRecord {
2423 0 : tenant_shard_id: value
2424 0 : .get_tenant_shard_id()
2425 0 : .expect("stored tenant id should be valid"),
2426 0 : healthy: value.healthy,
2427 0 : last_scrubbed_at: value.last_scrubbed_at,
2428 0 : }
2429 0 : }
2430 : }
2431 :
2432 : #[derive(
2433 0 : Serialize, Deserialize, Queryable, Selectable, Insertable, Eq, PartialEq, Debug, Clone,
2434 : )]
2435 : #[diesel(table_name = crate::schema::controllers)]
2436 : pub(crate) struct ControllerPersistence {
2437 : pub(crate) address: String,
2438 : pub(crate) started_at: chrono::DateTime<chrono::Utc>,
2439 : }
2440 :
2441 : // What we store in the database
2442 0 : #[derive(Serialize, Deserialize, Queryable, Selectable, Eq, PartialEq, Debug, Clone)]
2443 : #[diesel(table_name = crate::schema::safekeepers)]
2444 : pub(crate) struct SafekeeperPersistence {
2445 : pub(crate) id: i64,
2446 : pub(crate) region_id: String,
2447 : /// 1 is special, it means just created (not currently posted to storcon).
2448 : /// Zero or negative is not really expected.
2449 : /// Otherwise the number from `release-$(number_of_commits_on_branch)` tag.
2450 : pub(crate) version: i64,
2451 : pub(crate) host: String,
2452 : pub(crate) port: i32,
2453 : pub(crate) http_port: i32,
2454 : pub(crate) availability_zone_id: String,
2455 : pub(crate) scheduling_policy: SkSchedulingPolicyFromSql,
2456 : pub(crate) https_port: Option<i32>,
2457 : }
2458 :
2459 : /// Wrapper struct around [`SkSchedulingPolicy`] because both it and [`FromSql`] are from foreign crates,
2460 : /// and we don't want to make [`safekeeper_api`] depend on [`diesel`].
2461 0 : #[derive(Serialize, Deserialize, FromSqlRow, Eq, PartialEq, Debug, Copy, Clone)]
2462 : pub(crate) struct SkSchedulingPolicyFromSql(pub(crate) SkSchedulingPolicy);
2463 :
2464 : impl From<SkSchedulingPolicy> for SkSchedulingPolicyFromSql {
2465 0 : fn from(value: SkSchedulingPolicy) -> Self {
2466 0 : SkSchedulingPolicyFromSql(value)
2467 0 : }
2468 : }
2469 :
2470 : impl FromSql<diesel::sql_types::VarChar, Pg> for SkSchedulingPolicyFromSql {
2471 0 : fn from_sql(
2472 0 : bytes: <Pg as diesel::backend::Backend>::RawValue<'_>,
2473 0 : ) -> diesel::deserialize::Result<Self> {
2474 0 : let bytes = bytes.as_bytes();
2475 0 : match core::str::from_utf8(bytes) {
2476 0 : Ok(s) => match SkSchedulingPolicy::from_str(s) {
2477 0 : Ok(policy) => Ok(SkSchedulingPolicyFromSql(policy)),
2478 0 : Err(e) => Err(format!("can't parse: {e}").into()),
2479 : },
2480 0 : Err(e) => Err(format!("invalid UTF-8 for scheduling policy: {e}").into()),
2481 : }
2482 0 : }
2483 : }
2484 :
2485 : impl SafekeeperPersistence {
2486 0 : pub(crate) fn from_upsert(
2487 0 : upsert: SafekeeperUpsert,
2488 0 : scheduling_policy: SkSchedulingPolicy,
2489 0 : ) -> Self {
2490 0 : crate::persistence::SafekeeperPersistence {
2491 0 : id: upsert.id,
2492 0 : region_id: upsert.region_id,
2493 0 : version: upsert.version,
2494 0 : host: upsert.host,
2495 0 : port: upsert.port,
2496 0 : http_port: upsert.http_port,
2497 0 : https_port: upsert.https_port,
2498 0 : availability_zone_id: upsert.availability_zone_id,
2499 0 : scheduling_policy: SkSchedulingPolicyFromSql(scheduling_policy),
2500 0 : }
2501 0 : }
2502 0 : pub(crate) fn as_describe_response(&self) -> Result<SafekeeperDescribeResponse, DatabaseError> {
2503 0 : Ok(SafekeeperDescribeResponse {
2504 0 : id: NodeId(self.id as u64),
2505 0 : region_id: self.region_id.clone(),
2506 0 : version: self.version,
2507 0 : host: self.host.clone(),
2508 0 : port: self.port,
2509 0 : http_port: self.http_port,
2510 0 : https_port: self.https_port,
2511 0 : availability_zone_id: self.availability_zone_id.clone(),
2512 0 : scheduling_policy: self.scheduling_policy.0,
2513 0 : })
2514 0 : }
2515 : }
2516 :
2517 : /// What we expect from the upsert http api
2518 0 : #[derive(Serialize, Deserialize, Eq, PartialEq, Debug, Clone)]
2519 : pub(crate) struct SafekeeperUpsert {
2520 : pub(crate) id: i64,
2521 : pub(crate) region_id: String,
2522 : /// 1 is special, it means just created (not currently posted to storcon).
2523 : /// Zero or negative is not really expected.
2524 : /// Otherwise the number from `release-$(number_of_commits_on_branch)` tag.
2525 : pub(crate) version: i64,
2526 : pub(crate) host: String,
2527 : pub(crate) port: i32,
2528 : /// The active flag will not be stored in the database and will be ignored.
2529 : pub(crate) active: Option<bool>,
2530 : pub(crate) http_port: i32,
2531 : pub(crate) https_port: Option<i32>,
2532 : pub(crate) availability_zone_id: String,
2533 : }
2534 :
2535 : impl SafekeeperUpsert {
2536 0 : fn as_insert_or_update(&self) -> anyhow::Result<InsertUpdateSafekeeper<'_>> {
2537 0 : if self.version < 0 {
2538 0 : anyhow::bail!("negative version: {}", self.version);
2539 0 : }
2540 0 : Ok(InsertUpdateSafekeeper {
2541 0 : id: self.id,
2542 0 : region_id: &self.region_id,
2543 0 : version: self.version,
2544 0 : host: &self.host,
2545 0 : port: self.port,
2546 0 : http_port: self.http_port,
2547 0 : https_port: self.https_port,
2548 0 : availability_zone_id: &self.availability_zone_id,
2549 0 : // None means a wish to not update this column. We expose abilities to update it via other means.
2550 0 : scheduling_policy: None,
2551 0 : })
2552 0 : }
2553 : }
2554 :
2555 : #[derive(Insertable, AsChangeset)]
2556 : #[diesel(table_name = crate::schema::safekeepers)]
2557 : struct InsertUpdateSafekeeper<'a> {
2558 : id: i64,
2559 : region_id: &'a str,
2560 : version: i64,
2561 : host: &'a str,
2562 : port: i32,
2563 : http_port: i32,
2564 : https_port: Option<i32>,
2565 : availability_zone_id: &'a str,
2566 : scheduling_policy: Option<&'a str>,
2567 : }
2568 :
2569 0 : #[derive(Serialize, Deserialize, FromSqlRow, AsExpression, Eq, PartialEq, Debug, Copy, Clone)]
2570 : #[diesel(sql_type = crate::schema::sql_types::PgLsn)]
2571 : pub(crate) struct LsnWrapper(pub(crate) Lsn);
2572 :
2573 : impl From<Lsn> for LsnWrapper {
2574 0 : fn from(value: Lsn) -> Self {
2575 0 : LsnWrapper(value)
2576 0 : }
2577 : }
2578 :
2579 : impl FromSql<crate::schema::sql_types::PgLsn, Pg> for LsnWrapper {
2580 0 : fn from_sql(
2581 0 : bytes: <Pg as diesel::backend::Backend>::RawValue<'_>,
2582 0 : ) -> diesel::deserialize::Result<Self> {
2583 0 : let byte_arr: diesel::deserialize::Result<[u8; 8]> = bytes
2584 0 : .as_bytes()
2585 0 : .try_into()
2586 0 : .map_err(|_| "Can't obtain lsn from sql".into());
2587 0 : Ok(LsnWrapper(Lsn(u64::from_be_bytes(byte_arr?))))
2588 0 : }
2589 : }
2590 :
2591 : impl ToSql<crate::schema::sql_types::PgLsn, Pg> for LsnWrapper {
2592 0 : fn to_sql<'b>(
2593 0 : &'b self,
2594 0 : out: &mut diesel::serialize::Output<'b, '_, Pg>,
2595 0 : ) -> diesel::serialize::Result {
2596 0 : out.write_all(&u64::to_be_bytes(self.0.0))
2597 0 : .map(|_| IsNull::No)
2598 0 : .map_err(Into::into)
2599 0 : }
2600 : }
2601 :
2602 : #[derive(Insertable, AsChangeset, Clone)]
2603 : #[diesel(table_name = crate::schema::timelines)]
2604 : pub(crate) struct TimelinePersistence {
2605 : pub(crate) tenant_id: String,
2606 : pub(crate) timeline_id: String,
2607 : pub(crate) start_lsn: LsnWrapper,
2608 : pub(crate) generation: i32,
2609 : pub(crate) sk_set: Vec<i64>,
2610 : pub(crate) new_sk_set: Option<Vec<i64>>,
2611 : pub(crate) cplane_notified_generation: i32,
2612 : pub(crate) deleted_at: Option<chrono::DateTime<chrono::Utc>>,
2613 : pub(crate) sk_set_notified_generation: i32,
2614 : }
2615 :
2616 : /// This is separate from [TimelinePersistence] only because postgres allows NULLs
2617 : /// in arrays and there is no way to forbid that at schema level. Hence diesel
2618 : /// wants `sk_set` to be `Vec<Option<i64>>` instead of `Vec<i64>` for
2619 : /// Queryable/Selectable. It does however allow insertions without redundant
2620 : /// Option(s), so [TimelinePersistence] doesn't have them.
2621 0 : #[derive(Queryable, Selectable)]
2622 : #[diesel(table_name = crate::schema::timelines)]
2623 : pub(crate) struct TimelineFromDb {
2624 : pub(crate) tenant_id: String,
2625 : pub(crate) timeline_id: String,
2626 : pub(crate) start_lsn: LsnWrapper,
2627 : pub(crate) generation: i32,
2628 : pub(crate) sk_set: Vec<Option<i64>>,
2629 : pub(crate) new_sk_set: Option<Vec<Option<i64>>>,
2630 : pub(crate) cplane_notified_generation: i32,
2631 : pub(crate) deleted_at: Option<chrono::DateTime<chrono::Utc>>,
2632 : pub(crate) sk_set_notified_generation: i32,
2633 : }
2634 :
2635 : impl TimelineFromDb {
2636 0 : fn into_persistence(self) -> TimelinePersistence {
2637 : // We should never encounter null entries in the sets, but we need to filter them out.
2638 : // There is no way to forbid this in the schema that diesel recognizes (to our knowledge).
2639 0 : let sk_set = self.sk_set.into_iter().flatten().collect::<Vec<_>>();
2640 0 : let new_sk_set = self
2641 0 : .new_sk_set
2642 0 : .map(|s| s.into_iter().flatten().collect::<Vec<_>>());
2643 0 : TimelinePersistence {
2644 0 : tenant_id: self.tenant_id,
2645 0 : timeline_id: self.timeline_id,
2646 0 : start_lsn: self.start_lsn,
2647 0 : generation: self.generation,
2648 0 : sk_set,
2649 0 : new_sk_set,
2650 0 : cplane_notified_generation: self.cplane_notified_generation,
2651 0 : deleted_at: self.deleted_at,
2652 0 : sk_set_notified_generation: self.sk_set_notified_generation,
2653 0 : }
2654 0 : }
2655 : }
2656 :
2657 : // This is separate from TimelinePersistence because we don't want to touch generation and deleted_at values for the update.
2658 : #[derive(AsChangeset)]
2659 : #[diesel(table_name = crate::schema::timelines)]
2660 : #[diesel(treat_none_as_null = true)]
2661 : pub(crate) struct TimelineUpdate {
2662 : pub(crate) tenant_id: String,
2663 : pub(crate) timeline_id: String,
2664 : pub(crate) start_lsn: LsnWrapper,
2665 : pub(crate) sk_set: Vec<i64>,
2666 : pub(crate) new_sk_set: Option<Vec<i64>>,
2667 : }
2668 :
2669 0 : #[derive(Insertable, AsChangeset, Queryable, Selectable, Clone)]
2670 : #[diesel(table_name = crate::schema::safekeeper_timeline_pending_ops)]
2671 : pub(crate) struct TimelinePendingOpPersistence {
2672 : pub(crate) sk_id: i64,
2673 : pub(crate) tenant_id: String,
2674 : pub(crate) timeline_id: String,
2675 : pub(crate) generation: i32,
2676 : pub(crate) op_kind: SafekeeperTimelineOpKind,
2677 : }
2678 :
2679 0 : #[derive(Serialize, Deserialize, FromSqlRow, AsExpression, Eq, PartialEq, Debug, Copy, Clone)]
2680 : #[diesel(sql_type = diesel::sql_types::VarChar)]
2681 : pub(crate) enum SafekeeperTimelineOpKind {
2682 : Pull,
2683 : Exclude,
2684 : Delete,
2685 : }
2686 :
2687 : impl FromSql<diesel::sql_types::VarChar, Pg> for SafekeeperTimelineOpKind {
2688 0 : fn from_sql(
2689 0 : bytes: <Pg as diesel::backend::Backend>::RawValue<'_>,
2690 0 : ) -> diesel::deserialize::Result<Self> {
2691 0 : let bytes = bytes.as_bytes();
2692 0 : match core::str::from_utf8(bytes) {
2693 0 : Ok(s) => match s {
2694 0 : "pull" => Ok(SafekeeperTimelineOpKind::Pull),
2695 0 : "exclude" => Ok(SafekeeperTimelineOpKind::Exclude),
2696 0 : "delete" => Ok(SafekeeperTimelineOpKind::Delete),
2697 0 : _ => Err(format!("can't parse: {s}").into()),
2698 : },
2699 0 : Err(e) => Err(format!("invalid UTF-8 for op_kind: {e}").into()),
2700 : }
2701 0 : }
2702 : }
2703 :
2704 : impl ToSql<diesel::sql_types::VarChar, Pg> for SafekeeperTimelineOpKind {
2705 0 : fn to_sql<'b>(
2706 0 : &'b self,
2707 0 : out: &mut diesel::serialize::Output<'b, '_, Pg>,
2708 0 : ) -> diesel::serialize::Result {
2709 0 : let kind_str = match self {
2710 0 : SafekeeperTimelineOpKind::Pull => "pull",
2711 0 : SafekeeperTimelineOpKind::Exclude => "exclude",
2712 0 : SafekeeperTimelineOpKind::Delete => "delete",
2713 : };
2714 0 : out.write_all(kind_str.as_bytes())
2715 0 : .map(|_| IsNull::No)
2716 0 : .map_err(Into::into)
2717 0 : }
2718 : }
2719 :
2720 0 : #[derive(Serialize, Deserialize, Queryable, Selectable, Insertable, Eq, PartialEq, Clone)]
2721 : #[diesel(table_name = crate::schema::timeline_imports)]
2722 : pub(crate) struct TimelineImportPersistence {
2723 : pub(crate) tenant_id: String,
2724 : pub(crate) timeline_id: String,
2725 : pub(crate) shard_statuses: serde_json::Value,
2726 : }
|
