A Java implementation of Lamport's Paxos algorithm.
This is an implementation of totally ordered atomic broadcast protocol. The algorithm used is a variant of Lamport's Paxos. This library is a lightweight alternative to Apache Zookeeper. Zookeeper uses Paxos to keep consistency among the replicas, but the clients are remote. This poses several limits to how Zookeeper can be used. Our library is used from within the VM and clients can use it to build replication with strong consistency guarantees regardless of how state is represented.
The original algorithm requires state to be persisted. This significantly reduces performance, but it is required for enabling members to recover. Our solution does not support recovery of members, but instead supports changing the members of the group. This allows us to avoid persistence and have a much higher throughput.
This is the basic group implementation
WARNING: The BasicGroup has several limitations, you should use the Dynamic Group instead.
// this is the list of members
Members members = new Members(
new Member(), // this is a reference to a member on the localhost on default port (2440)
new Member(2441), // this one is on localhost with the specified port
new Member(InetAddress.getLocalHost(), 2442)); // you can specify the address and port manually
// we need to define a receiver
class MyReceiver implements Receiver {
// we follow a reactive pattern here
public void receive(Serializable message) {
System.out.println("received " + message.toString());
}
};
// this actually creates the members
BasicGroup group1 = new BasicGroup(members.get(0), new MyReceiver());
BasicGroup group2 = new BasicGroup(members.get(1), new MyReceiver());
BasicGroup group3 = new BasicGroup(members.get(2), new MyReceiver());
// this will cause all receivers to print "received Hello"
group2.broadcast("Hello");
Thread.sleep(1); // allow the members to receive the message
group1.close(); group2.close(); group3.close();
The BasicGroup has one big limitation: it doesn't support messages larger than a UDP packet. Even if you never send large messages, the broadcast protocol may use large messages internally to synchronize state. Large messages are handled by the FragmentingGroup implementation. This implementation has a small overhead (about 10% lower throughput), but supports messages of any size.
As we said before, our Paxos implementation does not support recovery of members. Instead we support adding new members to the group. In order to take advantage of this you must use the DynamicGroup implementation. State transfer upon joining is left to the user, but we guarantee that every new member receives a continguous subsequence of messages.