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