Member-only story
Why 3 Replicas Are Better Than 4: The “ODD” Math Behind Distributed Systems
If one copy of your data is good, why do distributed systems keep multiple copies?
And if 3 replicas are good, why not use 4? Wouldn’t an extra replica make the system even more reliable?
To understand why systems like Kafka, etcd, ZooKeeper, and many distributed databases commonly use 3 or 5 replicas, we first need to understand two simple concepts:
Replica → A replica is simply another copy of the same data or service state maintained on a different machine or location.
If multiple replicas exist, how do we decide which replicas should participate in an operation?
For example, if a client writes new data, do we need to wait for all 3 replicas?What happens if one replica is down? What happens if the network is temporarily partitioned and replicas cannot communicate with each other?
This is where quorum comes in.
Quorum → A quorum is the minimum number of replicas that must agree or participate for a distributed operation to be considered valid.
Suppose we have 3 replicas:
Replica 1 → YES
Replica 2 → YES
Replica 3 → NO
We have two votes for YES and one for NO.
The majority is therefore: 2 out of 3 . That majority is our quorum.
