Consensus algorithms used in distributed systems to ensure that all processes agree on a single value.

• Paxos is fundamentally about reaching agreement in a network of unreliable components.
  • It ensures that a single value is agreed upon among the participants, even if some of them fail.
• Raft’s primary goal is to simplify the consensus process.
  • It breaks down the process into smaller sub problems.
• ZAB is about total order broadcast.
  • It ensures all actions are executed in the same order across all nodes.

How it works

Paxos (Prepare Accept XO)

• Phases: Paxos operates mainly in two phases — Prepare and Accept.
  • The Prepare phase is about proposing a value, and the Accept phase is about agreeing on it.
• Multi-Paxos: An optimization where one node remains the leader until it fails, reducing the number of messages.
  • This is especially useful in systems where leadership changes are costly.
• Advantages: Proven correctness and wide applicability.
  • It’s a foundation for many other consensus algorithms.
• Disadvantages: Its complexity makes it hard to implement and understand.
  • Many systems opt for simpler algorithms due to this.

Paxos protocol is named after a fictional legislative consensus system on the Greek island of Paxos, which is known for its complex political landscape.

Raft

• Leader Election: Raft has a unique approach to leader election.
  • A new leader is elected whenever the current one fails.
  • This ensures system availability.
• Log Replication: The leader is responsible for log replication.
  • It ensures that all logs are consistent across nodes.
• Advantages: Its main advantage is its simplicity.
  • It’s easier to understand and implement than Paxos.
• Disadvantages: Some argue it’s not as generalized as Paxos.
  • However, its simplicity often outweighs this disadvantage.

ZAB (Zookeeper Atomic Broadcast)

• Election Phase: A leader is elected.
  • This is crucial for the subsequent phases.
• Discovery Phase: The leader establishes synchronization with all followers.
  • This ensures that all nodes start from a consistent state.
• Broadcast Phase: The leader broadcasts client updates.
  • This ensures that all nodes have the same data.
• Advantages: Tailored for ZooKeeper’s needs, ensuring strong consistency.
• Disadvantages: It’s specific to ZooKeeper’s use case.
  • This might make it less suitable for other applications.

Real-life Use Cases

Paxos:

• Distributed Databases: Systems like Google’s Spanner employ Paxos for global consistency.
  • It ensures that data is consistent across global data centers.
• Configuration Management: Chubby, a lock service by Google, uses Paxos.
  • It ensures that configuration data is consistent.

Apache Cassandra uses Paxos, a consensus protocol, to achieve strong consistency. Paxos is also used in Cassandra for leader election

Raft:

• Storage Systems: CockroachDB, a distributed SQL database, uses Raft.
  • It ensures data consistency across nodes.
• Service Orchestration: Kubernetes, the container orchestration system, employs etcd which uses Raft.
  • It ensures that configuration data is consistent across nodes.
  • Other systems that use Raft include Etcd, CockroachDB, and DynamoDB.

ZAB:

• Distributed Coordination: Apache ZooKeeper, which uses ZAB, is pivotal for systems like Kafka.
  • It ensures that nodes in a distributed system coordinate effectively.

References

• https://lamport.azurewebsites.net/pubs/paxos-simple.pdf

• https://github.com/jguamie/system-design/blob/master/notes/raft-distributed-consensus.md

• https://medium.com/@remisharoon/paxos-vs-raft-vs-zab-a-comprehensive-dive-into-distributed-consensus-protocols-6243a3f6539b