Mercurial > lbo > hg > clusterconsensus
view README.md @ 10:c62d3bc5e8bb
A few refactorings to `Change` and `newError()`
author | Lewin Bormann <lbo@spheniscida.de> |
---|---|
date | Sat, 08 Oct 2016 10:49:22 +0200 |
parents | a8a747e3e568 |
children |
line wrap: on
line source
# clusterconsensus A library implementing a replicated state machine with generic state machine, transport and state changes. For example, a using application could use this package with protocol buffers, leveldb and gRPC; and another one might use JSON, in-memory state and HTTP. ## Protocol The protocol is a custom Paxos implementation. It works like this, basically: ### Bootstrapping * A set of participants is started. * On the participant that has been selected to become the first master, `InitMaster()` is called. * After `InitMaster()`, `AddParticipants()` adds more participants to the cluster. The list of participants could for example come from a configuration file or command line flags. * Then, the cluster should be up and running. Now, on any participant, `Submit()` can be called to submit changes to the state machine. Those changes are replicated through the master. (see next section) ### Normal operation We assume that the current instance is 12, and the sequence number in that instance is 34. * One participant wants to modify the state. It calls `Submit()` on the `Participant` with one or more changes. * If the Participant is not master, it will send the request to the master using the `Submit()` method on the stub. * If the Participant is the master, or has received a `Submit` request, it will proceed by coordinating the change. * First, all non-master participants are sent `Accept(12, 35, []Change{ *some change* })`. * The request is sent using the `ConsensusClient` stub that was returned by the `Connector` implementation. * This leads to the non-master participants *staging* that change into a special area. The change is not yet applied to the state machine. * The next time the master wants to apply another change, it sends `Accept(12, 36, []Change{ *some new change* })`. This leads to non-master participants "committing" (i.e. applying) all staged changes before sequence `36` to the state machine, including our change `35` from before. ### Master has crashed Current instance: 22; Current sequence: 44 * A participant, "B", wants to modify the state, but can't reach the master to `Submit()` a change. It proceeds by becoming a candidate. It sends `Prepare(23, Member{ addr: "B" })` to all known members. If a majority replies within the deadline with a positive vote, then `B` is the new master. * In order for all participants to know about this change, `B` submits a request `Accept(23, 1, []Change{})` which confirms the previous `Prepare()` call, and makes all non-master participants acknowledge the new master. * Now everyone knows of the new master, and the cluster continues as in *Normal operation*. ### Adding/removing members * Membership changes are relatively straight-forward; they are just special changes that don't use `Accept()`, but rather the `AddMember()` and `RemoveMember()` methods in `ParticipantStub`. * Membership changes are staged the same way as normal changes; memberships only actually change once a membership change has been committed by an `Accept()` call with a higher sequence number. * Membership adds are communicated to all old members. After successfully staging the change (i.e., getting a majority vote), the new member is started using `StartParticipation()`. * Membership removals are communicated to all members, including the one to be removed. Upon