/atomix

A reactive framework for building fault-tolerant distributed systems for the JVM

Primary LanguageJavaApache License 2.0Apache-2.0

Atomix

Build Status Coverage Status Maven Central Gitter Open Source Helpers

A framework for building fault-tolerant distributed systems for the JVM

Atomix 2.1 documentation is currently under development. Check the website for updates!

  1. Overview
  2. Background
  3. Architecture
  4. Java API
  5. REST API
  6. Docker container

Overview

Atomix 2.1 is a fully featured framework for building fault-tolerant distributed systems. Combining ZooKeeper's consistency with Hazelcast's usability and performance, Atomix uses a set of custom communication APIs, a sharded Raft cluster, and a multi-primary protocol to provide a series of high-level primitives for building distributed systems. These primitives include:

Background

Atomix was originally conceived in 2014 along with its sister project Copycat (deprecated) as a hobby project. Over time, Copycat grew into a mature implementation of the Raft consensus protocol, and both Copycat and Atomix were put into use in various projects. In 2017, development of a new version was begun, and Copycat and Atomix were combined in Atomix 2.x. Additionally, significant extensions to the projects originally developed for use in ONOS were migrated into Atomix 2.x. Atomix is now maintained as a core component of ONOS by the Open Networking Foundation.

Architecture

At its core, Atomix is built on a series of replicated state machines backed by consensus and primary-backup protocols. The most critical component of Atomix is one of the most complete implementations of the Raft consensus protocol. A single Raft cluster is used for internal coordination, and multiple additional Raft clusters are used to store state for consistent primitives. The primary-backup protocol uses the core Raft cluster to elect and balance primaries and backups for all data partitions. Primitives can be configured for consistency, persistence, and replication, modifying the underlying protocol according to desired semantics.

The Atomix cluster

Atomix clusters consist of two types of nodes:

  • DATA nodes store persistent and ephemeral primitive state
  • CLIENT nodes do not store any state but must connect to DATA nodes to store state remotely

Primitive partitions (both Raft and primary-backup) are evenly distributed among the DATA nodes in a cluster. Initially, an Atomix cluster is formed by bootstrapping a set of DATA nodes. Thereafter, additional DATA or CLIENT nodes may join and leave the cluster at will by simply starting and stopping Atomix instances. Atomix provides a ClusterService that can be used to learn about new CLIENT and DATA nodes joining and leaving the cluster.

Java API

Bootstrapping the cluster

Atomix relies heavily upon builder APIs to build high-level objects used to communicate and coordinate distributed systems.

To create a new Atomix instance, create an Atomix builder:

Atomix.Builder builder = Atomix.builder();

The builder should be configured with the local node configuration:

builder.withLocalNode(Node.builder("server1")
  .withType(Node.Type.DATA)
  .withEndpoint(Endpoint.from("localhost", 5000))
  .build());

In addition to configuring the local node information, each instance must be configured with a set of bootstrap nodes from which to form a cluster. When first starting a cluster, all instances should provide the same set of bootstrap nodes. Bootstrap nodes must be DATA nodes:

builder.withBootstrapNodes(
  Node.builder("server1")
    .withType(Node.Type.DATA)
    .withEndpoint(Endpoint.from("localhost", 5000))
    .build(),
  Node.builder("server2")
    .withType(Node.Type.DATA)
    .withEndpoint(Endpoint.from("localhost", 5001))
    .build(),
  Node.builder("server3")
    .withType(Node.Type.DATA)
    .withEndpoint(Endpoint.from("localhost", 5002))
    .build());

Additionally, the Atomix instance can be configured with the data directory, the number of Raft partitions, the number of primary-backup partitions, and other options. Once the instance has been configured, build the instance by calling build():

Atomix atomix = builder.build();

Finally, call start() on the instance to start the node:

atomix.start().join();

Note that in order to form a cluster, a majority of instances must be started concurrently to allow Raft partitions to form a quorum. The future returned by the start() method will not be completed until all partitions are able to form. If your Atomix instance is blocking indefinitely at startup, ensure you enable DEBUG logging to debug the issue.

Connecting a client node

Atomix also supports client nodes which can connect to an Atomix cluster and operate on distributed primitives but does not itself store primitive state. To create a client node, simply indicate in the node builder that the node is a CLIENT:

Atomix atomix = Atomix.builder()
  .withLocalNode(Node.builder("client")
    .withType(Node.Type.CLIENT)
    .withEndpoint(Endpoint.from("localhost", 5003))
    .build())
  .withBootstrapNodes(
      Node.builder("server1")
        .withType(Node.Type.DATA)
        .withEndpoint(Endpoint.from("localhost", 5000))
        .build(),
      Node.builder("server2")
        .withType(Node.Type.DATA)
        .withEndpoint(Endpoint.from("localhost", 5001))
        .build(),
      Node.builder("server3")
        .withType(Node.Type.DATA)
        .withEndpoint(Endpoint.from("localhost", 5002))
        .build())
  .build();

atomix.start().join();

This example connects a client node aptly named client to a set of data nodes. Once the instance is started, the client node will be visible to all data nodes and vice versa, and primitives created by the client node will be managed by the data nodes.

Cluster management

Atomix provides a set of APIs for discovering the nodes in the cluster. The ClusterMetadataService provides information about the set of data nodes in the cluster. The ClusterService provides information about all nodes in the cluster, including the local node, data nodes, and client nodes.

To get the set of nodes in the cluster, use the ClusterService:

Set<Node> nodes = atomix.clusterService().getNodes();

Failure detection

The Node objects provided by the ClusterService provide a Node.State that can indicates the liveliness of the given node. The cluster service uses a phi accrual failure detection algorithm internally to detect failures, and nodes' states are updated as failures are detected:

Node.State state = atomix.clusterService().getNode(NodeId.from("foo")).state();

Additionally, listeners can be added to the ClusterService to react to changes to both the set of nodes in the cluster and the states of individual nodes:

atomix.clusterService().addListener(event -> {
  if (event.type() == ClusterEvent.Type.NODE_ACTIVATED) {
    // A node's state was changed to ACTIVE
  } else if (event.type() == ClusterEvent.Type.NODE_DEACTIVATED) {
    // A node's state was changed to INACTIVE
  }
});

Cluster communication

Atomix provides three services that can be used for general cluster communication:

  • MessagingService is a low-level IP address based communication API
  • ClusterCommunicationService is a high-level point-to-point/unicast/multicast/broadcast messaging API
  • ClusterEventService is a high-level messaging API modelled on the Vert.x event bus that abstracts producers from consumers

The default implementation of communication abstractions uses Netty for all inter-node communication.

Direct messaging

Atomix provides the ClusterCommunicationService for point-to-point messaging between Atomix nodes. It provides support for unicast, multicast, broadcast, and request-reply messaging patterns.

Registering message subscribers

To register a message subscriber, use the subscribe methods:

atomix.messagingService().subscribe("test", message -> {
  return CompletableFuture.completedFuture(message);
});

Three types of subscribers can be registered:

  • A synchronous subscriber that returns a result and must provide an Executor on which to consume messages
  • An asynchronous subscriber that must return CompletableFuture
  • A consumer that must provide an Executor on which to consume messages

Sending messages

As noted above, messages can be sent using a variety of different communication patterns:

  • unicast sends a message to a single node without awaiting a response
  • multicast sends a message to a set of nodes without awaiting any responses
  • broadcast sends a message to all nodes known to the local ClusterService without awaiting any responses
  • sendAndReceive sends a message to a single node and awaits a response
// Send a request-reply message to node "foo"
atomix.messagingService().send("test", "Hello world!", NodeId.from("foo")).thenAccept(response -> {
  System.out.println("Received " + response);
});

Message serialization

The ClusterCommunicationService uses a default serializer to serialize a variety of core data structures, but often custom objects need to be communicated across the wire. The ClusterCommunicationService provides overloaded methods for providing arbitrary message encoders/decoders for requests/replies:

Serializer serializer = Serializer.using(KryoNamespace.builder()
  .register(KryoNamespaces.BASIC)
  .register(NodeId.class)
  .register(ClusterHeartbeat.class)
  .build());

ClusterHeartbeat heartbeat = new ClusterHeartbeat(atomix.clusterService().getLocalNode().id());
atomix.messagingService().broadcast("test", heartbeat, serializer::encode);

Publish-subscribe messaging

Publish-subscribe messaging is done using the ClusterEventService API, which is closely modelled on the ClusterCommunicationService API. Indeed, while the two appear to be almost the exact same, their semantics differ significantly. Rather than sending messages to specific nodes using NodeIds, the ClusterEventService actually replicates subscriber information and routes messages internally. Point-to-point messages sent via the ClusterEventService are delivered in a round-robin fashion, and multicast messages do not require any specific node information. This separates specific nodes from senders.

// Add an event service subscriber
atomix.eventingService().subscribe("test", message -> {
  return CompletableFuture.completedFuture(message);
});

// Send a request-reply message via the event service
atomix.eventingService().send("test", "Hello world!").thenAccept(response -> {
  System.out.println("Received " + response);
});

// Broadcast a message to all event subscribers
atomix.eventingService().broadcast("test", "Hello world!");

Coordination primitives

Coordination primitives can be used to elect leaders and synchronize access to shared resources in the cluster. These primitives typically require the use of consensus but allow certain consistency models to be relaxed, e.g. on reads. Coordination primitives include:

  • DistributedLock - coarse grained fair distributed lock
  • LeaderElection - single leader election
  • LeaderElector - multiple leader election with automatic leader balancing
  • WorkQueue - distributed persistent work queue with ack/fail mechanisms

DistributedLock

As with many other objects in Atomix, primitives are constructed using a builder pattern. All core primitive builders are exposed via the Atomix API:

DistributedLock lock = atomix.lockBuilder("test-lock")
  .withLockTimeout(Duration.ofSeconds(2))
  .build();

The lock timeout is the approximate time it takes to determine whether a lock has failed. If a lock's underlying Raft session is expired, the lock will be released and granted to the next waiting process.

lock.lock();
try {
  // Do something
} finally {
  lock.unlock();
}

As with all Atomix primitives, a fully asynchronous version of the lock can be constructed by simply calling the async() method on the lock:

AsyncDistributedLock asyncLock = lock.async();
lock.lock().thenRun(() -> {
  // Do something
  lock.unlock().thenRun(() -> {
    // Unlocked
  });
});

LeaderElector

The LeaderElector primitive is used to elect multiple leaders within a cluster, automatically balancing leaders among all instances of the primitive across the cluster.

LeaderElector<NodeId> elector = atomix.leaderElectorBuilder("test-elector")
  .withElectionTimeout(Duration.ofSeconds(2))
  .build();

As with DistributedLock, the LeaderElector (and LeaderElection) primitive supports a timeout after which inactive leaders will be demoted and a new leader will be elected, e.g. in the case of a crash or network partition. Similarly, as with all primitives, an AsyncLeaderElector can be created via the async() getter:

AsyncLeaderElector<NodeId> asyncElector = elector.async();

To enter into an election, use the run method:

asyncElector.run("foo", atomix.clusterService().getLocalNode().id()).thenAccept(leadership -> {
  if (leadership.leader().id().equals(atomix.clusterService().getLocalNode().id())) {
    // Local node elected leader!
  }
});

The topic passed as the first argument to the run method is an identifier for a named election in which to participate. Multiple elections can be managed by a single leader elector, and leaders' locations will be automatically balanced among all elections.

To get the current leader for a topic, use getLeadership:

Leadership<NodeId> leadership = elector.getLeadership("foo");

The Leadership provided by an elector contains both a Leader and a list of candidates in the election. For each unique Leader for a topic, a monotonically increasing unique term and wall clock time will be provided.

Finally, to listen for changes to leaderships managed by a LeaderElector, register a LeadershipEventListener:

elector.addListener(event -> {
  // Either a leader or candidates changed
  Leadership leadership = event.newLeadership();
});

Data primitives

Data primitives store state and can be replicated either using the partitioned Raft cluster or a multi-primary replication protocol. These primitives are typically data structures, including:

  • AtomicCounter - simple distributed atomic counter
  • AtomicValue - simple distributed atomic value
  • ConsistentMap - partitioned Map with change events and support for optimistic locking
  • ConsistentTreeMap - distributed TreeMap with change events and support for optimistic locking
  • ConsistentMultimap - partitioned multimap with change events
  • DistributedSet - partitioned Set with change events
  • DocumentTree - partitioned tree-like structure with change events and support for optimistic locking

ConsistentMap

The ConsistentMap primitive is modelled on Java's Map and provides support for listening for changes to the map and optimistic locking. To create a map, use the ConsistentMapBuilder:

ConsistentMap<String, String> map = atomix.consistentMapBuilder("test-map")
  .withPersistence(Persistence.EPHEMERAL)
  .withBackups(2)
  .build();

map.put("foo", "Hello world!");

Map values returned by the ConsistentMap are wrapped in a Versioned object which can be used for optimistic locking:

Versioned<String> value = map.get("foo");
map.replace("foo", "Hello world again!", value.version());

As with all other primitives, an AsyncConsistentMap can be created via the async() method:

AsyncConsistentMap<String, String> asyncMap = map.async();
asyncMap.put("bar", "baz").thenRun(() -> {
  // put complete
});

To listen for changes to the map, add a MapEventListener to the map:

map.addListener(event -> {
  switch (event.type()) {
    case INSERT:
      ...
    case UPDATE:
      ...
    case REMOVE:
      ...
  }
});

Transactions

Atomix supports transactional operations over multiple primitives. Transactions are committed using a two-phase commit protocol. To create a transaction, use the TransactionBuilder:

Transaction transaction = atomix.transactionBuilder()
  .withIsolation(Isolation.REPEATABLE_READS)
  .build();

To begin the transaction, call begin:

transaction.begin();

Once a transaction has been started, primitives can be created via the Transaction primitive builders. Once all operations have been performed, call commit() to commit the transaction:

CommitStatus status = transaction.commit();

The CommitStatus returned by the commit() call indicates whether the transaction was committed successfully. Transactions may not commit successfully if a concurrent transaction is already modifying the referenced resources.

To abort a transaction, call abort():

transaction.abort();

TransactionalMap

To create a transactional map, use the mapBuilder() method:

Transaction transaction = atomix.transactionBuilder()
  .withIsolation(Isolation.REPEATABLE_READS)
  .build();

transaction.begin();

TransactionalMap<String, String> transactionalMap = transaction.mapBuilder("my-map")
  .withSerializer(Serializers.using(KryoNamespaces.BASIC))
  .build();
transactionalMap.put("foo", "bar");
transactionalMap.put("bar", "baz");

if (transaction.commit() == CommitStatus.SUCCESS) {
  ...
}

The map will inherit the isolation level configured for the transaction. Changes to the named map will be reflected in non-transactional maps of the same name upon commit.

TransactionalSet

To create a transactional set, use the setBuilder() method:

Transaction transaction = atomix.transactionBuilder()
  .withIsolation(Isolation.REPEATABLE_READS)
  .build();

transaction.begin();

TransactionalSet<String> transactionalSet = transaction.setBuilder("my-set")
  .withSerializer(Serializers.using(KryoNamespaces.BASIC))
  .build();
transactionalSet.add("foo");
transactionalSet.add("bar");

if (transaction.commit() == CommitStatus.SUCCESS) {
  ...
}

The set will inherit the isolation level configured for the transaction. Changes to the named set will be reflected in non-transactional set of the same name upon commit.

Serialization

All data primitives support custom serialization via a Serializer provided to the builder:

Serializer dataSerializer = new Serializer() {
  @Override
  public <T> byte[] encode(T object) {
    return ...;
  }

  @Override
  public <T> T decode(byte[] bytes) {
    return ...;
  }
};

ConsistentMap<String, Data> map = atomix.consistentMapBuilder("data")
  .withSerializer(dataSerializer)
  .build();

Atomix also provides a core Kryo based serialization abstraction:

Serializer dataSerializer = Serializer.using(KryoNamespace.builder()
  .register(KryoNamespaces.BASIC)
  .register(Data.class)
  .build());

ConsistentMap<String, Data> map = atomix.consistentMapBuilder("data")
  .withSerializer(dataSerializer)
  .build();

By default, the KryoNamespace uses Kryo's FieldSerializer, but custom Kryo serializers may be provided via the builder. Additionally, a set of default serializers can be registered using the KryoNamespaces.BASIC constant.

REST API

The Atomix agent is a standalone server or client node with a REST API through which Java or HTTP clients can manage Atomix primitives remotely.

Running a standalone Atomix agent

The atomix script is used to run the Atomix agent and CLI:

mvn clean package
bin/atomix agent

Bootstrapping a standalone cluster

To start a cluster, the atomix agent command must be provided with a local node name, host, and port and a set of nodes with which to bootstrap the cluster.

The format of addresses passed to the agent command is:

{name}:{host}:{port}

Node names must be unique among all nodes in the cluster and will default to the local host name if not explicitly specified.

bin/atomix agent a:localhost:5000 --bootstrap a:localhost:5000 b:localhost:5001 c:localhost:5002 --http-port 6000 --data-dir data/a

bin/atomix agent b:localhost:5001 --bootstrap a:localhost:5000 b:localhost:5001 c:localhost:5002 --http-port 6001 --data-dir data/b

bin/atomix agent c:localhost:5002 --bootstrap a:localhost:5000 b:localhost:5001 c:localhost:5002 --http-port 6002 --data-dir data/c

Starting a client node

Client nodes provide the same HTTP API as server nodes but do not themselves store state. Client nodes can be useful for deploying an Atomix node with which to communicate locally and which can use more efficient binary communication when transporting requests across the network.

bin/atomix agent --client --bootstrap a:localhost:5000 b:localhost:5001 c:localhost:5002

REST API examples

This section provides a series of example usages of primitives via the HTTP API:

Acquire a lock

curl -XPOST http://localhost:5678/v1/primitives/locks/my-lock

Release a lock

curl -XDELETE http://localhost:5678/v1/primitives/locks/my-lock

Set a value in a map

curl -XPUT http://localhost:5678/v1/primitives/maps/my-map/foo -d value="Hello world!" -H "Content-Type: text/plain"

Get a value in a map

curl -XGET http://localhost:5678/v1/primitives/maps/my-map/foo

Send an event

curl -XPOST http://localhost:5678/v1/events/something-happened -d "Something happened!" -H "Content-Type: text/plain"

Receive events

curl -XGET http://localhost:5678/v1/events/something-happened

Docker

mvn clean package
docker build -t atomix .
docker run -p 5678:5678 -p 5679:5679 atomix docker:0.0.0.0

Acknowledgements

Thank you to the Open Networking Foundation and ONOS for continued support of Atomix!

YourKit supports open source projects with its full-featured Java Profiler. YourKit, LLC is the creator of YourKit Java Profiler and YourKit .NET Profiler, innovative and intelligent tools for profiling Java and .NET applications.

YourKit