/libraft

Raft distributed consensus protocol in Java

Primary LanguageJavaBSD 3-Clause "New" or "Revised" LicenseBSD-3-Clause

libraft Build Status Coverage Status

libraft is a Java library that implements the [Raft distributed consensus protocol] (https://ramcloud.stanford.edu/wiki/download/attachments/11370504/raft.pdf "In Search of an Understandable Consensus Algorithm").

Release History

  • 0.1.0: Initial release. (11 Dec, 2013)
  • 0.1.1: Bug-fix release. (25 Jan, 2014)

For more information check the detailed [release history] (https://github.com/allengeorge/libraft/wiki/Release-History "libraft release history").

Features

libraft completely implements all checked features below. Unchecked features are planned but do not have an implementation timeline.

  • Leader election
  • Log replication
  • [o] Log compaction (snapshots)
  • [o] Online cluster reconfiguration

Overview

libraft consists of 2 components:

  • libraft-core
  • libraft-agent

Correctness, safety, and understandability were major factors during design and implementation. Performance was an explicit non-goal.

libraft-core

libraft-core contains an implementation of the Raft algorithm. It also defines interfaces for the components (such as timers, packet senders, log and metadata storage, etc.) it needs to operate. These interfaces make very few demands on the underlying implementation and should be easy to integrate into an existing stack.

libraft-agent

libraft-agent:

  1. Provides reference implementations for the interfaces defined in libraft-core.
  2. Provides a facade over libraft-core that can be instantiated to provide consensus services.

It uses JDBC for persistence, TCP connections for networking, and json for the wire format. It wires these implementations up to the algorithm classes from libraft-core. libraft-agent also defines a simple [json configuration format] (#configuring-libraft "Configuring libraft") with which users can configure their cluster. Finally, it provides a single class - RaftAgent - with which applications can control the lifecycle of a Raft server and interact with the Raft cluster.

Samples

libraft also comes with a fully-implemented sample application in libraft-samples/kayvee. KayVee is a distributed, consistent, key-value store that uses RaftAgent to durably replicate key-value pairs to a cluster. It demonstrates how libraft can be integrated into an application to provide consensus services.

Getting libraft

The simplest way to use libraft is through Gradle or Maven - simply add libraft-agent as a dependency. Replace ${libraft.version.latest} in the code below with the latest version on Maven Central.

Gradle

dependencies {
    compile 'io.libraft:libraft-agent:${libraft.version.latest}'
}

Maven

<dependencies>
    <dependency>
        <groupId>io.libraft</groupId>
        <artifactId>libraft-agent</artifactId>
        <version>${libraft.version.latest}</version>
    </dependency>
</dependencies>

Alternatively, if you only want the algorithm components and intend to build your own implementations for its interfaces, only libraft-core is necessary. Simply add libraft-core as a dependency, and replace ${libraft.version.latest} in the code below with the latest version on Maven Central.

Gradle

dependencies {
    compile 'io.libraft:libraft-core:${libraft.version.latest}'
}

Maven

<dependencies>
    <dependency>
        <groupId>io.libraft</groupId>
        <artifactId>libraft-core</artifactId>
        <version>${libraft.version.latest}</version>
    </dependency>
</dependencies>

Configuring libraft

RaftAgent can be configured in one of two ways:

  1. A json configuration file.
  2. A RaftConfiguration instance.

Both methods are equivalent and have the same properties.

Properties

libraft-agent exposes the following configuration properties. bolded properties are required.

  • minElectionTimeout: minimum election timeout for a Raft server
  • additionalElectionTimeoutRange: maximum additional time added to minElectionTimeout to get the applied election timeout. Election timeout is defined using the following formula: electionTimeout = minElectionTimeout + randomInRange(0, additionalElectionTimeoutRange).
  • rpcTimeout: maximum time a Raft server will wait for a response to an RPC request
  • heartbeatInterval: maximum time between messages from a Raft leader
  • connectTimeout: maximum time a Raft server will wait to establish a connection to another Raft server
  • minReconnectInterval: minimum interval a Raft server will wait before reconnecting to another Raft server
  • additionalReconnectIntervalRange: maximum additional time added to minReconnectInterval to get the applied reconnect interval. Reconnect interval is defined using the following formula: reconnectInterval = minReconnectInterval + randomInRange(0, additionalReconnectIntervalRange).
  • database: Raft database configuration block
    • driverClass: fully-qualified class name of the JDBC driver
    • url: JDBC connection URL
    • user: database user id
    • password: database password (may be empty or omitted)
  • cluster: Raft cluster configuration block
    • self: unique id of the local Raft server
    • members: Raft cluster configuration. Defines all the members in the cluster, including the local server.
      • id: unique id of the Raft server
      • endpoint: address - in "host:port" format - at which this server can be reached

Sample

A sample configuration file for a 5-server Raft cluster is given below. Note that self is S_00, indicating that this is the configuration file for server S_00 in the cluster. This file includes all fields, both required and optional.

{
    "minElectionTimeout": 180,
    "additionalElectionTimeoutRange": 120,
    "rpcTimeout": 30,
    "heartbeatInterval": 15,
    "connectTimeout": 5000,
    "minReconnectInterval": 10000,
    "additionalReconnectIntervalRange": 1000,

    "database": {
        "driverClass": "org.h2.Driver",
        "url": "jdbc:h2:test_db",
        "user": "test",
        "password": "test"
    },

    "cluster": {
        "self": "S_00",
        "members": [
            {
                "id": "S_00",
                "endpoint": "192.168.1.100:9990"
            },
            {
                "id": "S_01",
                "endpoint": "192.168.1.100:9991"
            },
            {
                "id": "S_02",
                "endpoint": "192.168.1.100:9992"
            },
            {
                "id": "S_03",
                "endpoint": "192.168.1.100:9993"
            },
            {
                "id": "S_04",
                "endpoint": "192.168.1.100:9994"
            }
        ]
    }
}

Using libraft

The simplest way to use libraft is to instantiate a RaftAgent within an application node (such as a server). This RaftAgent will allow the node to participate in a Raft cluster: it can become a leader or follower and will be notified of leadership changes as well as applied (i.e. committed) commands.

Creating a RaftAgent requires:

  1. A valid [configuration file or object] (#configuring-libraft "Configuring libraft").
  2. A specialization of Command for the application commands that will be committed to the Raft cluster.
  3. A specialization of RaftListener that will be notified of leadership changes and applied commands.

Sample json configuration file ("agent.config"):

{
    "database": {
        "driverClass": "org.h2.Driver",
        "url": "jdbc:h2:test_db",
        "user": "test",
        "password": "test"
    },

    "cluster": {
        "self": "agent0",
        "members": [
            {
                "id": "agent2",
                "endpoint": "192.168.1.100:9990"
            },
            {
                "id": "agent1",
                "endpoint": "192.168.1.100:9991"
            },
            {
                "id": "agent0",
                "endpoint": "192.168.1.100:9992"
            }
        ]
    }
}

Sample java code:

//
// Command to be replicated to the Raft cluster
//

public class MyCommand implements Command {

    enum CommandType {
        GET,
        SET
    }

    @JsonProperty
    CommandType commandType;

    @JsonProperty
    String key;

    @JsonProperty
    String value;

    // constructor, getters, and setters follow...
}

//
// Listener that will be notified of events in the Raft cluster
//

public class MyRaftListener implements RaftListener {

    @Override
    public void onLeadershipChange(@Nullable String leader) {
        //...
    }

    @Override
    public void applyCommitted(long index, Command command) {
        //...
    }
}

//
// Application code
//

// create the listener that will be tied to the RaftAgent
RaftListener raftListener = new MyRaftListener();

// create the RaftAgent using the configuration in "agent.config.json"
// the listener created above of leadership changes and applied commands
RaftAgent raftAgent = RaftAgent.fromConfigurationFile("agent.config", raftListener);

// indicate that the application command uses Jackson annotations
// RaftAgent has special support for this
raftAgent.setupJacksonAnnotatedCommandSerializationAndDeserialization(MyCommand.class);

// initialize the RaftAgent
// (sets up state, but does not connect to any servers, start any timers, etc.)
// this allows the application to assume that it has sole access to resources
// allowing it to perform verification and bootstrap tasks in single-threaded mode
raftAgent.initialize();

// start the RaftAgent (this starts the internal server/client, timers, etc.)
raftAgent.start();

// submit commands (if leader)
if (raftListener.isLeader()) {
    raftAgent.submitCommand(new MyCommand(MyCommand.CommandType.SET, "key0", "val0"));
    raftAgent.submitCommand(new MyCommand(MyCommand.CommandType.SET, "key1", "val1"));
    // more work...
} else {
    // more work...
}

// stop the agent (stops the internal server/clients, timers, etc.)
raftAgent.stop();

RaftAgent supports more than what is outlined here (non-Jackson-annotated command serialization, etc.) For more examples, see KayVee.java in libraft-samples/kayvee and RaftAgentTest.java in libraft-agent.

Building libraft

libraft requires java 1.6 and gradle 1.10+ to build.

To build and install into the local maven cache, run gradle from the repository root.

gradle build
gradle install

libraft also ships with a gradle wrapper, which allows the code to be built without prior installation of gradle. To use gradle wrapper run gradlew from the repository root.

./gradlew build
./gradlew install

To clean build artifacts run gradle clean or gradlew clean from the repository root.

Issues

The full list of issues can be seen at [Github Issues] (https://github.com/allengeorge/libraft/issues "libraft issues").

There are no known safety issues.

Reporting Issues

Please submit all code or documentation issues, comments and concerns or feature requests to [Github Issues] (https://github.com/allengeorge/libraft/issues "libraft issues").

If you have other libraft or Raft questions, please post to the [raft-dev] (https://groups.google.com/forum/#!forum/raft-dev "raft-dev") mailing list.

Thanks!

libraft stands on the shoulders of others. This implementation would not be possible without the very detailed (and clear!) paper published by Diego Ongaro and John Ousterhout. Moreover, this library was built on a huge host of open-source software - many thanks to the teams behind them!

Just some of the open-source software libraft uses include:

It's this - and much, much more! - that makes libraft possible.