Primary-Backup Replication Explained | MIT 6.824 Lecture 4

Added:

Masterless Replication and Eventual Consistency

While primary-backup replication relies on a single primary node to order transactions and guarantee strong consistency (linearizability), it introduces write bottlenecks, single points of failure, and complex split-brain mitigation. A major alternative paradigm is Masterless (or Dynamo-style) Replication. Popularized by systems like Amazon's Dynamo and Apache Cassandra, masterless replication allows any node in the system to accept read and write operations. Instead of enforcing strict synchronous replication through a primary, it prioritizes high availability and low latency—aligning with the CAP theorem—by allowing replica divergence and resolving conflicts asynchronously (using techniques like vector clocks or Last-Write-Wins). This counterpoint argues that for global-scale systems, the strict synchronization and failover overhead of primary-backup models are impractical, making eventual consistency and decentralized writes a more resilient and scalable design choice.

Basic distributed system architectures, including client-server communication models, network latency, and independent machine failures.
The concept of State Machine Replication (SMR) and how deterministic state machines transition based on ordered input sequences.
The distinction between deterministic and non-deterministic operations (e.g., system clocks, interrupts, and random number generators) in program execution.
The 'split-brain' problem in distributed networks and why simple dual-node setups struggle with network partitions.
Consensus protocols such as Raft or Paxos, which dynamically elect leaders and safely resolve split-brain scenarios in a replicated system.
Consistency models in distributed systems, specifically Linearizability and Sequential Consistency, to formalize correctness guarantees.
The trade-offs between State Transfer (sending raw memory/disk state) and Replicated State Machine (sending deterministic instruction logs) methodologies.
Byzantine Fault Tolerance (BFT), which extends fault tolerance to handle malicious or arbitrary node behavior rather than just fail-stop crashes.
Real-world distributed storage and coordination engines that leverage advanced replication, such as Google Spanner, Apache ZooKeeper, or etcd.
86.7K views0likes1:19:20@6.824Original Release: 2020-02-18

Primary-backup replication is a fault tolerance technique where a primary server and backup server execute in lockstep to ensure service availability during hardware failures; VMware FT implements this at the machine level by replicating the entire virtual machine state (memory and registers) between two physical machines, using a logging channel to synchronize execution and handle non-deterministic events like timer interrupts and external inputs, while the output rule ensures consistent client responses by delaying output until the backup acknowledges receipt of all preceding log records.