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