Majority Quorum

The **Majority Quorum** pattern is the mathematical foundation for strong consistency and high availability in distributed clusters. It ensures that a system can tolerate node failures without losing data or allowing conflicting updates, provided a majority of the cluster remains operational.

1. The Quorum Inequality

A distributed system consists of $N$total nodes. To maintain strong consistency (Linearizability), the number of nodes required for a successful write ($W$) and the number of nodes polled for a read ($R$) must satisfy the following inequality:$$R + W > N$$### The Pigeonhole Principle

The logic relies on the **Pigeonhole Principle**: if the sets of nodes used for reading and writing overlap by at least one node, that overlapping node acts as the "witness" that carries the most recent state.

2. Mathematical Proof

We can prove that$R + W > N$guarantees that every read will see the latest write.

1. Let$S$be the set of all$N$nodes.

2. Let$V_w$be the set of$W$nodes that acknowledge the latest write.

3. Let$V_r$be the set of$R$nodes polled for a read.

4. The number of nodes that **did not** participate in the write is$N - W$.

5. If the read set$V_r$were to contain *no* nodes from$V_w$, then$V_r$must be a subset of the non-write set:$V_r \subseteq (S \setminus V_w)$.

6. This is only possible if the size of the read set is less than or equal to the size of the non-write set:$R \le N - W$.

7. Rearranging gives$R + W \le N$.

8. By contradiction, if **$R + W > N$**, there must be at least one node$n$that is in both sets:$n \in (V_w \cap V_r)$.

3. Common Quorum Configurations

The selection of$R$and$W$values allows architects to tune the system for specific workload profiles:

| :--- | :--- | :--- | :--- |

4. Fault Tolerance Calculation

For a cluster of size$N$using strict majority ($W = R = \lfloor N/2 \rfloor + 1$), the number of nodes that can fail ($f$) while maintaining availability is:$$f = \lfloor \frac{N-1}{2} \rfloor$$| Cluster Size ($N$) | Max Failures ($f$) | Majority Required |

| :--- | :--- | :--- |

| 3 | 1 | 2 |

| 5 | 2 | 3 |

| 7 | 3 | 4 |

**Architectural Note:** Distributed clusters almost always use **odd numbers** of nodes. Increasing from 3 to 4 nodes does not increase the fault tolerance (both can only survive 1 failure), but it increases the number of nodes that must be coordinated for a majority (from 2 to 3), actually decreasing performance.

5. Usage in Industry

* **Cassandra / Dynamo:** Allows per-request configuration of$R$and$W$ (e.g., `QUORUM`, `LOCAL_QUORUM`, `ONE`).

* **Raft / Paxos:** Mandates a strict majority for every log entry to ensure safe leader transitions and durability.