Fault-Tolerant Replication with Pull-Based Consensus in MongoDB

In this paper, we present the design and implementation of strongly consistent replication in MongoDB. MongoDB provides linearizability and tolerates any minority of failures through a novel consensus protocol that derives from Raft. A major difference between our protocol and vanilla Raft is that MongoDB deploys a unique pull-based data synchronization model: a replica pulls new data from another replica. This pull-based data synchronization in MongoDB can be initiated by any replica and can happen between any two replicas, as opposed to vanilla Raft, where new data can only be pushed from the primary to other replicas. This flexible data transmission topology enabled by the pull-based model is strongly desired by our users since it has an edge on performance and monetary cost. This paper describes how this consensus protocol works, how MongoDB integrates it with the rest of the replication system, and the extensions of the replication protocol that support our rich feature set. Our evaluation shows that MongoDB effectively achieved the design goals and can replicate data efficiently and reliably.