fix(sqlite): one extenddb process per database file

README.md

@@ -136,6 +136,10 @@ extenddb healthcheck --endpoint https://127.0.0.1:18443 # explicit target
 
 This is a liveness check, which is what a container `HEALTHCHECK` wants: `/health` does not query the storage backend, so it reports healthy even if PostgreSQL becomes unreachable after startup. That is deliberate, since a liveness probe that failed on a database outage would restart every replica at once. A backend that is unreachable at startup does stop the server from listening, so that case is caught. There is no separate readiness endpoint yet.
 
+### One instance per database
+
+Run one `extenddb serve` per catalog. On SQLite this is enforced: the server takes an exclusive lock on `<database>.lock` next to the database file and a second server on the same file refuses to start (`init`, `migrate`, and `destroy` take the same lock while they run). On PostgreSQL and MongoDB nothing prevents a second instance, but running more than one against the same catalog is not supported in this release: every instance runs every background worker, credential and policy caches are per instance with no cross-instance invalidation (a revoked key stays valid on the other instance for up to the cache TTL), and `/health` does not check the database. See the [deployment guide](docs/manuals/11-deployment-guide.md#multi-instance-considerations).
+
 To make the generated self-signed certificate valid for the name clients use — an in-cluster service DNS name, for example — pass `--tls-san` to `init` (repeatable):
 
 ```bash

crates/storage-sqlite/Cargo.toml

@@ -35,3 +35,4 @@ rand = { workspace = true }
 bigdecimal = { workspace = true }
 crc32fast = { workspace = true }
 zeroize = { workspace = true }
+libc = { workspace = true }

crates/storage-sqlite/src/bootstrapper.rs

@@ -20,6 +20,7 @@ use sqlx::SqlitePool;
 use sqlx::sqlite::SqlitePoolOptions;
 
 use crate::schema::{self, CATALOG_VERSION};
+use crate::serve_lock::ServeLock;
 use crate::sqlite_util::sqlite_url;
 
 /// SQLite backend bootstrapper.
@@ -29,11 +30,41 @@ use crate::sqlite_util::sqlite_url;
 /// paths, not the hot serving path.
 pub struct SqliteBootstrapper {
     path: String,
+    /// The database-file lock held between [`Bootstrapper::acquire_migration_lock`]
+    /// and [`Bootstrapper::release_migration_lock`]; see `serve_lock`.
+    held_lock: std::sync::Mutex<Option<ServeLock>>,
 }
 
 impl SqliteBootstrapper {
     pub fn new(path: impl Into<String>) -> Self {
-        Self { path: path.into() }
+        Self {
+            path: path.into(),
+            held_lock: std::sync::Mutex::new(None),
+        }
+    }
+
+    /// Take the exclusive lock a running `extenddb serve` holds on this
+    /// database, or explain which process holds it. `Ok(None)` for an
+    /// in-memory database.
+    fn exclusive_lock(&self, what: &str) -> OpResult<Option<ServeLock>> {
+        // The lock file lives next to the database; `init` may be the one
+        // creating that directory (see `pool`).
+        if !self.is_memory()
+            && let Some(parent) = std::path::Path::new(&self.path).parent()
+            && !parent.as_os_str().is_empty()
+        {
+            std::fs::create_dir_all(parent).map_err(|e| {
+                OpError::Internal(format!(
+                    "create parent directory for SQLite database '{}': {e}",
+                    self.path
+                ))
+            })?;
+        }
+        ServeLock::acquire_for_location(&self.path).map_err(|e| {
+            OpError::Internal(format!(
+                "{what} needs exclusive use of the SQLite database and cannot have it: {e}"
+            ))
+        })
     }
 
     /// Build a `SqliteBootstrapper` from the config file and CLI args.
@@ -444,10 +475,37 @@ impl Bootstrapper for SqliteBootstrapper {
         Ok(row.map(|(v,)| v))
     }
 
+    /// `init` and `migrate` rewrite the schema, so they take the same
+    /// exclusive lock `extenddb serve` holds on the database file. Unlike the
+    /// trait's default contract this does not wait for the lock: the holder is
+    /// a server that runs until stopped, so waiting would hang, and the
+    /// operator is told to stop it instead.
+    async fn acquire_migration_lock(&self) -> OpResult<()> {
+        let lock = self.exclusive_lock("this command")?;
+        *self
+            .held_lock
+            .lock()
+            .map_err(|_| OpError::Internal("lock state poisoned".to_owned()))? = lock;
+        Ok(())
+    }
+
+    async fn release_migration_lock(&self) -> OpResult<()> {
+        self.held_lock
+            .lock()
+            .map_err(|_| OpError::Internal("lock state poisoned".to_owned()))?
+            .take();
+        Ok(())
+    }
+
     async fn drop_databases(&self, _data_db: &str) -> OpResult<()> {
         if self.is_memory() {
             return Ok(());
         }
+        // A server that still has the file open would keep serving an unlinked
+        // database, and a later `init` would create a second one at the same
+        // path. Refuse while any other process holds the file, and hold it
+        // ourselves until the files are gone.
+        let _lock = self.exclusive_lock("destroy")?;
         if std::path::Path::new(&self.path).exists() {
             std::fs::remove_file(&self.path)
                 .map_err(|e| OpError::Internal(format!("remove database file: {e}")))?;
@@ -498,3 +556,56 @@ impl Bootstrapper for SqliteBootstrapper {
         )
     }
 }
+
+#[cfg(all(test, unix))]
+mod lock_tests {
+    use super::SqliteBootstrapper;
+    use crate::serve_lock::ServeLock;
+    use extenddb_storage::bootstrapper::Bootstrapper;
+
+    /// `destroy` and `migrate` are refused while a server holds the database,
+    /// and a server is refused while `migrate` holds it.
+    #[tokio::test]
+    async fn destroy_and_migrate_are_refused_while_a_server_holds_the_file() {
+        let dir = std::env::temp_dir().join(format!(
+            "extenddb-bootstrap-lock-{}",
+            uuid::Uuid::new_v4().simple()
+        ));
+        std::fs::create_dir_all(&dir).expect("dir");
+        let db = dir.join("db.sqlite");
+        std::fs::write(&db, b"").expect("db file");
+        let bootstrap = SqliteBootstrapper::new(db.to_string_lossy().into_owned());
+
+        let server = ServeLock::acquire(&db).expect("server lock");
+        let err = bootstrap
+            .drop_databases("unused")
+            .await
+            .expect_err("destroy under a server");
+        assert!(
+            format!("{err:?}").contains("another extenddb process"),
+            "{err:?}"
+        );
+        assert!(
+            db.exists(),
+            "destroy must not unlink a file a server has open"
+        );
+        let err = bootstrap
+            .acquire_migration_lock()
+            .await
+            .expect_err("migrate under a server");
+        assert!(
+            format!("{err:?}").contains("another extenddb process"),
+            "{err:?}"
+        );
+        drop(server);
+
+        bootstrap.acquire_migration_lock().await.expect("free now");
+        assert!(ServeLock::acquire(&db).is_err(), "serve during migrate");
+        bootstrap.release_migration_lock().await.expect("release");
+        ServeLock::acquire(&db).expect("free after release");
+
+        bootstrap.drop_databases("unused").await.expect("destroy");
+        assert!(!db.exists());
+        let _ = std::fs::remove_dir_all(&dir);
+    }
+}

crates/storage-sqlite/src/lib.rs

@@ -36,6 +36,7 @@ mod metadata;
 mod number_key;
 mod operations;
 mod schema;
+mod serve_lock;
 mod sqlite_util;
 mod store;
 mod stream;
@@ -188,7 +189,16 @@ fn sqlite_server_components_factory(
     let pool_size = config.max_connections();
     let region = region.to_owned();
     Box::pin(async move {
-        let engine = SqliteEngine::new(&db_path, pool_size, &region, MAX_ITEM_SIZE_BYTES)
+        // One server per database file, taken before the database is opened:
+        // startup recovery below assumes no other server is using it. A file
+        // database only; an in-memory one belongs to this process alone.
+        let serve_lock = serve_lock::ServeLock::acquire_for_location(&db_path)
+            .map_err(BackendError::InitializationFailed)?
+            .map(|lock| {
+                tracing::debug!("holding serve lock {}", lock.path().display());
+                std::sync::Arc::new(lock)
+            });
+        let mut engine = SqliteEngine::new(&db_path, pool_size, &region, MAX_ITEM_SIZE_BYTES)
             .await
             .map_err(|e| BackendError::ConnectionFailed {
                 backend: "sqlite".to_owned(),
@@ -259,6 +269,7 @@ fn sqlite_server_components_factory(
             Err(e) => tracing::error!("Failed to reconcile incomplete vector indexes: {e}"),
         }
 
+        engine.serve_lock = serve_lock;
         let control_plane_notify = engine.control_plane_notify();
         let engine = Arc::new(engine);
 
@@ -304,3 +315,40 @@ fn sqlite_server_components_factory(
         })
     })
 }
+
+#[cfg(all(test, unix))]
+mod serve_lock_tests {
+    /// The server factory itself takes the lock: a second server on the same
+    /// file fails to start while the first is up, and starts once it is gone.
+    #[tokio::test]
+    async fn a_second_server_on_one_file_is_refused() {
+        let dir = std::env::temp_dir().join(format!(
+            "extenddb-serve-lock-factory-{}",
+            uuid::Uuid::new_v4().simple()
+        ));
+        std::fs::create_dir_all(&dir).expect("dir");
+        let config = crate::config::SqliteConfig {
+            path: dir.join("db.sqlite").to_string_lossy().into_owned(),
+            pool_size: 2,
+        };
+        let mut options = extenddb_storage::server_components::ServerComponentsOptions::default();
+        options.bootstrap_if_uninitialized = true;
+        let first = super::sqlite_server_components_factory(&config, "us-east-1", options)
+            .await
+            .expect("first server");
+        let Err(err) = super::sqlite_server_components_factory(&config, "us-east-1", options).await
+        else {
+            panic!("a second server on the same file started");
+        };
+        assert!(
+            err.to_string().contains("another extenddb process"),
+            "{err}"
+        );
+        drop(first);
+        let again = super::sqlite_server_components_factory(&config, "us-east-1", options)
+            .await
+            .expect("starts once the first is gone");
+        drop(again);
+        let _ = std::fs::remove_dir_all(&dir);
+    }
+}