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