This Raft library is stable and feature complete. As of 2016, it is **the most widely used** Raft library in production, serving tens of thousands clusters each day. It powers distributed systems such as etcd, Kubernetes, Docker Swarm, Cloud Foundry Diego, CockroachDB, TiDB, Project Calico, Flannel, and more.
Most Raft implementations have a monolithic design, including storage handling, messaging serialization, and network transport. This library instead follows a minimalistic design philosophy by only implementing the core raft algorithm. This minimalism buys flexibility, determinism, and performance.
To keep the codebase small as well as provide flexibility, the library only implements the Raft algorithm; both network and disk IO are left to the user. Library users must implement their own transportation layer for message passing between Raft peers over the wire. Similarly, users must implement their own storage layer to persist the Raft log and state.
In order to easily test the Raft library, its behavior should be deterministic. To achieve this determinism, the library models Raft as a state machine. The state machine takes a `Message` as input. A message can either be a local timer update or a network message sent from a remote peer. The state machine's output is a 3-tuple `{[]Messages, []LogEntries, NextState}` consisting of an array of `Messages`, `log entries`, and `Raft state changes`. For state machines with the same state, the same state machine input should always generate the same state machine output.
A simple example application, _raftexample_, is also available to help illustrate how to use this package in practice: https://github.com/coreos/etcd/tree/master/contrib/raftexample
The primary object in raft is a Node. Either start a Node from scratch using raft.StartNode or start a Node from some initial state using raft.RestartNode.
To allow a new node to join this cluster, do not pass in any peers. First, add the node to the existing cluster by calling `ProposeConfChange` on any existing node inside the cluster. Then, start the node with an empty peer list, like so:
1. Write Entries, HardState and Snapshot to persistent storage in order, i.e. Entries first, then HardState and Snapshot if they are not empty. If persistent storage supports atomic writes then all of them can be written together. Note that when writing an Entry with Index i, any previously-persisted entries with Index >= i must be discarded.
2. Send all Messages to the nodes named in the To field. It is important that no messages be sent until the latest HardState has been persisted to disk, and all Entries written by any previous Ready batch (Messages may be sent while entries from the same batch are being persisted). To reduce the I/O latency, an optimization can be applied to make leader write to disk in parallel with its followers (as explained at section 10.2.1 in Raft thesis). If any Message has type MsgSnap, call Node.ReportSnapshot() after it has been sent (these messages may be large). Note: Marshalling messages is not thread-safe; it is important to make sure that no new entries are persisted while marshalling. The easiest way to achieve this is to serialise the messages directly inside the main raft loop.
3. Apply Snapshot (if any) and CommittedEntries to the state machine. If any committed Entry has Type EntryConfChange, call Node.ApplyConfChange() to apply it to the node. The configuration change may be cancelled at this point by setting the NodeID field to zero before calling ApplyConfChange (but ApplyConfChange must be called one way or the other, and the decision to cancel must be based solely on the state machine and not external information such as the observed health of the node).
4. Call Node.Advance() to signal readiness for the next batch of updates. This may be done at any time after step 1, although all updates must be processed in the order they were returned by Ready.
Second, all persisted log entries must be made available via an implementation of the Storage interface. The provided MemoryStorage type can be used for this (if repopulating its state upon a restart), or a custom disk-backed implementation can be supplied.
Third, after receiving a message from another node, pass it to Node.Step:
Finally, call `Node.Tick()` at regular intervals (probably via a `time.Ticker`). Raft has two important timeouts: heartbeat and the election timeout. However, internally to the raft package time is represented by an abstract "tick".
If the proposal is committed, data will appear in committed entries with type raftpb.EntryNormal. There is no guarantee that a proposed command will be committed; the command may have to be reproposed after a timeout.
This implementation is up to date with the final Raft thesis (https://ramcloud.stanford.edu/~ongaro/thesis.pdf), although this implementation of the membership change protocol differs somewhat from that described in chapter 4. The key invariant that membership changes happen one node at a time is preserved, but in our implementation the membership change takes effect when its entry is applied, not when it is added to the log (so the entry is committed under the old membership instead of the new). This is equivalent in terms of safety, since the old and new configurations are guaranteed to overlap.
To ensure there is no attempt to commit two membership changes at once by matching log positions (which would be unsafe since they should have different quorum requirements), any proposed membership change is simply disallowed while any uncommitted change appears in the leader's log.
This approach introduces a problem when removing a member from a two-member cluster: If one of the members dies before the other one receives the commit of the confchange entry, then the member cannot be removed any more since the cluster cannot make progress. For this reason it is highly recommended to use three or more nodes in every cluster.