November 9, 2019

Notes on raft

Below are my notes from watching Martin Thompson’s talk on Cluster consensus in Aeron.

Raft - a consensus algorithm

Most fitting definition of consensus

the judgement agreed on by the majority

Consensus can apply to a single value or shared state made of different primitives.

Prior work - Paxos

Consensus algorithms were an active research topic in the 1970s and 1980s and Paxos was one of the early ones.

However, the implementation complexity has let Paxos down and made it hard to use.

What problem is Raft solving

Raft is a consensus algorithm that optimises for human understanding and ease of implementation.

In a nutshell

Raft makes the implementation easy to follow by minimising the state space as much as possible and sticking to a few key principles.

Principles

  • Monotonically increasing time divided into terms: every term starts with an election. Elections can either results in a new leader for this term or undecided, in which case a new term starts with another election.
  • Randomisation settles conflicts. Followers have a random timeout. If they don’t receive an AppendEntries RPC from their leader after this time out, they assume that the king/leader is dead (long live the leader), they randomly decide to throw in the ring and become candidates for this election.

Roles

There are only 3 roles and transitions between them can be represented by this finite state machine.

  • Follower
Every node starts as a follower. Followers are passive and issue no requests,
they can only serve the requests of others: candidates or the current leader.
  • Candidate
A node that wants to become leader in this term/election cycle. Sends the
RequestVote RPC to followers with its current state.
  • Leader
The only node that can send AppendEntries requests to other nodes. Is in
charge of dealing with requests from the client and passing them to
followers.

RPC

  • RequestVote
Candidates get on the campaign trail by asking followers to back them.
Instead of hosting debates, this election is settled by checking if the
candidate "knows more" i.e. is further ahead in its state than the
follower that receives the request.

When a follower receives a RequestVote it compares its state with that of
the candidate and sends a positive vote if the candidate is further ahead
in its state.

This ensures that no candidate that is behind in terms is elected to be leader.

> This is one of the crucial differences between Paxos and Raft. Paxos allows
electing a leader that is behind in its state. After electing such a leader,
they need to catch up/replay the state from their followers. This removes the
monotonic principle and increases implementation complexity by needing more RPC
types.
  • AppendEntries
After successfully winning an election, the leader uses the same RPC to
replicate his state and heartbeat with all other nodes.

If there are no entries to append, this RPC serves as a heartbeat. In
case, there are new entries, the leader sends it out to all its followers.

Applications

  • Using Raft enables aeron to focus on “Guaranteed Processing” of messages instead of “Guaranteed Delivery”. If processing is idempotent, delivering the same message twice doesn’t change system state.
  • Dynamically changing the state of the system - adding new nodes or hot upgrades to currently running nodes.
  • State machine replication for event-sourced systems. Once the leader sequences messages from the outside world and its followers commit them to their logs, we can do a Gaddafi without losing system state.

Cool visualition with more detail

http://thesecretlivesofdata.com/raft/