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