Isolates

Accessing drift databases on multiple isolates.

As sqlite3 is a synchronous C library, accessing the database from the main isolate can cause blocking IO operations that lead to reduced responsiveness of your application. To resolve this problem, drift can spawn a long-running isolate to run SQL statements. When following the recommended getting started guide and using NativeDatabase.createInBackground, you automatically benefit from an isolate drift manages for you without needing additional setup. This page describes when advanced isolate setups are necessary, and how to approach them.

Introduction

While the default setup is probably suitable for most apps, some scenarios require complete additional over the way drift manages isolates. In particular, some of these

  • You want to use a drift isolate in compute(), Isolate.run or generally in other isolates.
  • You need to access a drift database in a background worker.
  • You want to control the way drift spawns isolates instead of using the default.

When you try to send a drift database instance across isolates, you will run into an exception about sending an invalid object:

Future<void> invalidIsolateUsage() async {
  final database = MyDatabase(NativeDatabase.memory());

  // Unfortunately, this doesn't work: Drift databases contain references to
  // async primitives like streams and futures that can't be serialized across
  // isolates like this.
  await Isolate.run(() async {
    await database.batch((batch) {
      // ...
    });
  });
}

Unfortunately, there is no magic change drift could implement to make sending databases over isolates feasible: There's simply too much mutable state needed to implement features like stream queries or high-level transaction APIs. However, with a little bit of additional setup, you can use drift APIs to obtain two database instances that are synchronized by an isolate channel drift manages for you. Writes on one database are readable on the other isolate and even update stream queries. So essentially, you get one logical database instance shared between isolates.

Simple sharing

Starting from drift 2.5, running a short-lived computation workload on a separate isolate is easily possible with computeWithDatabase. This function helps spawn an isolate and opens another database instance on that isolate. Since the background database instance needs to talk to the main instance on the foreground isolate for synchronization, you need a constructor on your database class that can take a custom QueryExecutor, the class used by drift to represent lower-level databases:


@DriftDatabase(tables: [SomeTable] /* ... */)
class MyDatabase extends _$MyDatabase {
  // A constructor like this can use the default connection as described in the
  // getting started guide, but also allows overriding the connection.
  MyDatabase([QueryExecutor? executor]) : super(executor ?? _openConnection());

  @override
  int get schemaVersion => 1;
}
Future<void> insertBulkData(MyDatabase database) async {
  // computeWithDatabase is an extension provided by package:drift/isolate.dart
  await database.computeWithDatabase(
    computation: (database) async {
      // Expensive computation that runs on its own isolate but talks to the
      // main database.
      final rows = await _complexAndExpensiveOperationToFetchRows();
      await database.batch((batch) {
        batch.insertAll(database.someTable, rows);
      });
    },
    connect: (connection) {
      // This function is responsible for creating a second instance of your
      // database class with a short-lived [connection].
      // For this to work, your database class needs to have a constructor that
      // allows taking a connection as described above.
      return MyDatabase(connection);
    },
  );
}

As the example shows, computeWithDatabase is an API useful to run heavy tasks, like inserting a large amount of batch data, into a database.

Internally, computeWithDatabase does the following:

  1. It sets up a pair of SendPort / ReceivePorts over which database calls are relayed.
  2. It spawns a new isolate with Isolate.run and creates a raw database connection based on those ports.
  3. The new isolate invokes the connect callback to create a second instance of your database class that talks to the main instance over isolate ports.
  4. The computation callback is invoked.
  5. Transparently, drift also takes care of winding down the connection afterwards.

If you don't want drift to spawn the isolate for you, for instance because you want to use compute instead of Isolate.run, you can also do that manually with the serializableConnection() API:

Future<void> customIsolateUsage(MyDatabase database) async {
  final connection = await database.serializableConnection();

  await Isolate.run(
    () async {
      // We can't share the [database] object across isolates, but the connection
      // is fine!
      final databaseForIsolate = MyDatabase(await connection.connect());

      try {
        await databaseForIsolate.batch((batch) {
          // (...)
        });
      } finally {
        databaseForIsolate.close();
      }
    },
    debugName: 'My custom database task',
  );
}

Manually managing drift isolates

Instead of using functions like NativeDatabase.createInBackground or computeWithDatabase, you can also create database connections that can be shared across isolates manually.

Drift exposes the DriftIsolate class, which is a reference to an internal database server you can access on other isolates. Creating a DriftIsolate server is possible with DriftIsolate.spawn():

Future<DriftIsolate> createIsolateWithSpawn() async {
  final token = RootIsolateToken.instance!;
  return await DriftIsolate.spawn(() {
    // This function runs in a new isolate, so we must first initialize the
    // messenger to use platform channels.
    BackgroundIsolateBinaryMessenger.ensureInitialized(token);

    // The callback to DriftIsolate.spawn() must return the database connection
    // to use.
    return LazyDatabase(() async {
      // Note that this runs on a background isolate, which only started to
      // support platform channels in Flutter 3.7. For earlier Flutter versions,
      // a workaround is described later in this article.
      final dbFolder = await getApplicationDocumentsDirectory();
      final path = p.join(dbFolder.path, 'app.db');

      return NativeDatabase(File(path));
    });
  });
}

If you want to spawn the isolate yourself, that is possible too:

Future<DriftIsolate> createIsolateManually() async {
  final receiveIsolate = ReceivePort('receive drift isolate handle');
  await Isolate.spawn<SendPort>((message) async {
    final server = DriftIsolate.inCurrent(() {
      // Again, this needs to return the LazyDatabase or the connection to use.
    });

    // Now, inform the original isolate about the created server:
    message.send(server);
  }, receiveIsolate.sendPort);

  final server = await receiveIsolate.first as DriftIsolate;
  receiveIsolate.close();
  return server;
}

After creating a DriftIsolate server, you can use connect() to connect to it from different isolates:

import 'package:drift/isolate.dart';

@DriftDatabase(tables: [SomeTable] /* ... */)
class MyDatabase extends _$MyDatabase {
  // A constructor like this can use the default connection as described in the
  // getting started guide, but also allows overriding the connection.
  MyDatabase([QueryExecutor? executor]) : super(executor ?? _openConnection());

  @override
  int get schemaVersion => 1;
}
void main() async {
  final isolate = await createIsolate();

  // After creating the isolate, calling connect() will return a connection
  // which can be used to create a database.
  // As long as the isolate is used by only one database (it is here), we can
  // use `singleClientMode` to dispose the isolate after closing the connection.
  final database = MyDatabase(await isolate.connect(singleClientMode: true));

  // you can now use your database exactly like you regularly would, it
  // transparently uses a background isolate to execute queries.
}

If you need to construct the database outside of an async context, you can use the DatabaseConnection.delayed constructor. In the example above, you could synchronously obtain a MyDatabase instance by using:

MyDatabase(
  DatabaseConnection.delayed(Future.sync(() async {
    final isolate = await createIsolate();
    return isolate.connect(singleClientMode: true);
  })),
);

This can be helpful when using drift in dependency injection frameworks, since you have a way to create the database instance synchronously. Internally, drift will connect when the first query is sent to the database.

Workaround for old Flutter versions

Before Flutter 3.7, platforms channels weren't available on background isolates. So, if functions like getApplicationDocumentsDirectory from path_provider are used to construct the path to the database, some tricks were necessary. This section describes a workaround to start the isolate running the database manually. This allows passing additional data that can be computed on the main isolate, using platform channels.

import 'package:path/path.dart' as p;
import 'package:path_provider/path_provider.dart';

Future<DriftIsolate> _createDriftIsolate() async {
  // this method is called from the main isolate. Since we can't use
  // getApplicationDocumentsDirectory on a background isolate, we calculate
  // the database path in the foreground isolate and then inform the
  // background isolate about the path.
  final dir = await getApplicationDocumentsDirectory();
  final path = p.join(dir.path, 'db.sqlite');
  final receivePort = ReceivePort();

  await Isolate.spawn(
    _startBackground,
    _IsolateStartRequest(receivePort.sendPort, path),
  );

  // _startBackground will send the DriftIsolate to this ReceivePort
  return await receivePort.first as DriftIsolate;
}

void _startBackground(_IsolateStartRequest request) {
  // this is the entry point from the background isolate! Let's create
  // the database from the path we received
  final executor = NativeDatabase(File(request.targetPath));
  // we're using DriftIsolate.inCurrent here as this method already runs on a
  // background isolate. If we used DriftIsolate.spawn, a third isolate would be
  // started which is not what we want!
  final driftIsolate = DriftIsolate.inCurrent(
    () => DatabaseConnection(executor),
  );
  // inform the starting isolate about this, so that it can call .connect()
  request.sendDriftIsolate.send(driftIsolate);
}

// used to bundle the SendPort and the target path, since isolate entry point
// functions can only take one parameter.
class _IsolateStartRequest {
  final SendPort sendDriftIsolate;
  final String targetPath;

  _IsolateStartRequest(this.sendDriftIsolate, this.targetPath);
}

Once again, you can use a DatabaseConnection.delayed() to obtain a database connection for your database class:

DatabaseConnection createDriftIsolateAndConnect() {
  return DatabaseConnection.delayed(Future.sync(() async {
    final isolate = await _createDriftIsolate();
    return await isolate.connect(singleClientMode: true);
  }));
}

Shutting down the isolate

Multiple clients can connect to a single DriftIsolate multiple times. So, by default, the isolate must outlive individual connections. Simply calling Database.close on one of the clients won't stop the isolate (which could interrupt other databases). Instead, use DriftIsolate.shutdownAll() to close the isolate and all clients. This call will release all resources used by the drift isolate.

In many cases, you know that only a single client will connect to the DriftIsolate (for instance because you're spawning a new DriftIsolate when opening a database). In this case, you can set the singleClientMode: true parameter on connect(). With this parameter, closing the single connection will also fully dispose the drift isolate.

Common operation modes

You can use a DriftIsolate across multiple isolates you control and connect from any of them.

One executor isolate, one foreground isolate: This is the most common usage mode. You would call DriftIsolate.spawn from the main method in your Flutter or Dart app. Similar to the example above, you could then use drift from the main isolate by connecting with DriftIsolate.connect and passing that connection to a generated database class. In this case, using the DriftIsolate APIs may be an overkill - NativeDatabase.createInBackground will do the exact same thing for you.

One executor isolate, multiple client isolates: The DriftIsolate handle can be sent across multiple isolates, each of which can use DriftIsolate.connect on their own. This is useful to implement a setup where you have three or more threads:

  • The drift executor isolate
  • A foreground isolate, probably for Flutter
  • Another background isolate, which could be used for networking or other long-running expensive tasks.

You can then read data from the foreground isolate or start query streams, similar to the example above. The background isolate would also call DriftIsolate.connect and create its own instance of the generated database class. Writes to one database will be visible to the other isolate and also update query streams.

The receiving end can reconstruct a DriftIsolate from a SendPort by using the DriftIsolate.fromConnectPort constructor. That DriftIsolate behaves exactly like the original one, but we only had to send a primitive SendPort and not a complex Dart object.

How does this work? Are there any limitations?

All drift features are supported on background isolates and work out of the box. This includes

  • Transactions
  • Auto-updating queries (even if the table was updated from another isolate)
  • Batched updates and inserts
  • Custom statements or those generated from an sql api

Please note that, while using a background isolate can reduce lag on the UI thread, the overall database is going to be slightly slower! There's a overhead involved in sending data between isolates, and that's exactly what drift has to do internally. If you're not running into dropped frames because of drift, using a background isolate is probably not necessary for your app. However, isolate performance has dramatically improved in recent Dart and Flutter versions.

Internally, drift uses the following model to implement this api:

  • A server isolate: A single isolate that executes all queries and broadcasts tables updates. This is the isolate created by DriftIsolate.spawn. It supports any number of clients via an rpc-like connection model. Connections are established via SendPorts and ReceivePorts. Internally, the DriftIsolate class only contains a reference to a SendPort that can be used to establish a connection to the background isolate. This lets users share the DriftIsolate object across many isolates and connect multiple times. The actual server logic that listens on the port is in a private RunningDriftServer class.
  • Client isolates: Any number of clients in any number of isolates can connect to a DriftIsolate. The client acts as a drift backend, which means that all queries are built on the client isolate. The raw sql string and parameters are then sent to the server isolate, which will enqueue the operation and execute it eventually. Implementing the isolate commands at a low level allows users to re-use all their code used without the isolate api.

Independent isolates

All setups mentioned here assume that there will be one main isolate responsible for spawning a DriftIsolate that it (and other isolates) can then connect to.

In Flutter apps, this model may not always fit your use case. For instance, your app may use background tasks or receive FCM notifications while closed. These tasks will run in a background FlutterEngine managed by native platform code, so there's no clear communication scheme between isolates. Still, you may want to share a live drift database between your UI engine and potential background engines, even without them directly knowing about each other.

An IsolateNameServer from dart:ui can be used to transparently share a drift isolate between such workers. You can store the connectPort of a DriftIsolate under a specific name to look it up later. Other clients can use DriftIsolate.fromConnectPort to obtain a DriftIsolate from the name server, if one has been registered.

Please note that, at the moment, Flutter still has some inherent problems with spawning isolates from background engines that complicate this setup. Further, the IsolateNameServer is not cleared on a (stateless) hot reload, even though the isolates are stopped and registered ports become invalid. There is no reliable way to check if a SendPort is bound to an active ReceivePort or not.

Possible implementations of this pattern and associated problems are described in this issue.