»Consensus Protocol

Nomad uses a consensus protocol to provide Consistency (as defined by CAP). The consensus protocol is based on "Raft: In search of an Understandable Consensus Algorithm". For a visual explanation of Raft, see The Secret Lives of Data.

»Raft Protocol Overview

Raft is a consensus algorithm that is based on Paxos. Compared to Paxos, Raft is designed to have fewer states and a simpler, more understandable algorithm.

There are a few key terms to know when discussing Raft:

  • Log - The primary unit of work in a Raft system is a log entry. The problem of consistency can be decomposed into a replicated log. A log is an ordered sequence of entries. We consider the log consistent if all members agree on the entries and their order.

  • FSM - Finite State Machine. An FSM is a collection of finite states with transitions between them. As new logs are applied, the FSM is allowed to transition between states. Application of the same sequence of logs must result in the same state, meaning behavior must be deterministic.

  • Peer set - The peer set is the set of all members participating in log replication. For Nomad's purposes, all server nodes are in the peer set of the local region.

  • Quorum - A quorum is a majority of members from a peer set: for a set of size n, quorum requires at least ⌊(n/2)+1⌋ members. For example, if there are 5 members in the peer set, we would need 3 nodes to form a quorum. If a quorum of nodes is unavailable for any reason, the cluster becomes unavailable and no new logs can be committed.

  • Committed Entry - An entry is considered committed when it is durably stored on a quorum of nodes. Once an entry is committed it can be applied.

  • Leader - At any given time, the peer set elects a single node to be the leader. The leader is responsible for ingesting new log entries, replicating to followers, and managing when an entry is considered committed.

Raft is a complex protocol and will not be covered here in detail (for those who desire a more comprehensive treatment, the full specification is available in this paper). We will, however, attempt to provide a high level description which may be useful for building a mental model.

Raft nodes are always in one of three states: follower, candidate, or leader. All nodes initially start out as a follower. In this state, nodes can accept log entries from a leader and cast votes. If no entries are received for some time, nodes self-promote to the candidate state. In the candidate state, nodes request votes from their peers. If a candidate receives a quorum of votes, then it is promoted to a leader. The leader must accept new log entries and replicate to all the other followers. In addition, if stale reads are not acceptable, all queries must also be performed on the leader.

Once a cluster has a leader, it is able to accept new log entries. A client can request that a leader append a new log entry (from Raft's perspective, a log entry is an opaque binary blob). The leader then writes the entry to durable storage and attempts to replicate to a quorum of followers. Once the log entry is considered committed, it can be applied to a finite state machine. The finite state machine is application specific; in Nomad's case, we use MemDB to maintain cluster state.

Obviously, it would be undesirable to allow a replicated log to grow in an unbounded fashion. Raft provides a mechanism by which the current state is snapshotted and the log is compacted. Because of the FSM abstraction, restoring the state of the FSM must result in the same state as a replay of old logs. This allows Raft to capture the FSM state at a point in time and then remove all the logs that were used to reach that state. This is performed automatically without user intervention and prevents unbounded disk usage while also minimizing time spent replaying logs. One of the advantages of using MemDB is that it allows Nomad to continue accepting new transactions even while old state is being snapshotted, preventing any availability issues.

Consensus is fault-tolerant up to the point where quorum is available. If a quorum of nodes is unavailable, it is impossible to process log entries or reason about peer membership. For example, suppose there are only 2 peers: A and B. The quorum size is also 2, meaning both nodes must agree to commit a log entry. If either A or B fails, it is now impossible to reach quorum. This means the cluster is unable to add or remove a node or to commit any additional log entries. This results in unavailability. At this point, manual intervention would be required to remove either A or B and to restart the remaining node in bootstrap mode.

A Raft cluster of 3 nodes can tolerate a single node failure while a cluster of 5 can tolerate 2 node failures. The recommended configuration is to either run 3 or 5 Nomad servers per region. This maximizes availability without greatly sacrificing performance. The deployment table below summarizes the potential cluster size options and the fault tolerance of each.

In terms of performance, Raft is comparable to Paxos. Assuming stable leadership, committing a log entry requires a single round trip to half of the cluster. Thus, performance is bound by disk I/O and network latency.

»Raft in Nomad

Only Nomad server nodes participate in Raft and are part of the peer set. All client nodes forward requests to servers. The clients in Nomad only need to know about their allocations and query that information from the servers, while the servers need to maintain the global state of the cluster.

Since all servers participate as part of the peer set, they all know the current leader. When an RPC request arrives at a non-leader server, the request is forwarded to the leader. If the RPC is a query type, meaning it is read-only, the leader generates the result based on the current state of the FSM. If the RPC is a transaction type, meaning it modifies state, the leader generates a new log entry and applies it using Raft. Once the log entry is committed and applied to the FSM, the transaction is complete.

Because of the nature of Raft's replication, performance is sensitive to network latency. For this reason, each region elects an independent leader and maintains a disjoint peer set. Data is partitioned by region, so each leader is responsible only for data in their region. When a request is received for a remote region, the request is forwarded to the correct leader. This design allows for lower latency transactions and higher availability without sacrificing consistency.

»Consistency Modes

Although all writes to the replicated log go through Raft, reads are more flexible. To support various trade-offs that developers may want, Nomad supports 2 different consistency modes for reads.

The two read modes are:

  • default - Raft makes use of leader leasing, providing a time window in which the leader assumes its role is stable. However, if a leader is partitioned from the remaining peers, a new leader may be elected while the old leader is holding the lease. This means there are 2 leader nodes. There is no risk of a split-brain since the old leader will be unable to commit new logs. However, if the old leader services any reads, the values are potentially stale. The default consistency mode relies only on leader leasing, exposing clients to potentially stale values. We make this trade-off because reads are fast, usually strongly consistent, and only stale in a hard-to-trigger situation. The time window of stale reads is also bounded since the leader will step down due to the partition.

  • stale - This mode allows any server to service the read regardless of if it is the leader. This means reads can be arbitrarily stale but are generally within 50 milliseconds of the leader. The trade-off is very fast and scalable reads but with stale values. This mode allows reads without a leader meaning a cluster that is unavailable will still be able to respond.

»Deployment Table

Below is a table that shows quorum size and failure tolerance for various cluster sizes. The recommended deployment is either 3 or 5 servers. A single server deployment is highly discouraged as data loss is inevitable in a failure scenario.

ServersQuorum SizeFailure Tolerance