Replication Topologies

Why Replicate?

Replication copies data across multiple nodes for three reasons: availability (survive node failures), read scalability (serve reads from replicas), and latency (serve reads from geographically close replicas). The topology determines how data flows between nodes.

Leader-Follower

Leader-follower (primary-replica) is the most common replication topology. The leader handles all writes. Followers replicate from the leader and serve reads. This provides strong consistency for reads from the leader and eventual consistency from followers.

Leader (Primary)
  │
  ├── async replication ──► Follower 1 (us-east-1)
  │                          └── serves read queries
  │
  ├── async replication ──► Follower 2 (us-west-2)
  │                          └── serves read queries
  │
  └── async replication ──► Follower 3 (eu-west-1)
                             └── serves read queries

Writes ──► Leader only
Reads ──► Followers (eventual consistency)

Multi-Leader Replication

Multi-leader replication allows writes at multiple nodes. Each node is a leader that can accept writes and replicate to other leaders. This is useful for multi-datacenter deployments where each datacenter needs local write performance.

Leader A (us-east-1) ◄──► Leader B (eu-west-1)
       │                         │
       └── replication ──►┘      └── replication ──►┘

Writes at A replicate to B (and vice versa)
Conflict resolution required when same key written at both

Quorum Replication

Quorum replication requires a minimum number of nodes to acknowledge a write (write quorum) or serve a read (read quorum). The formula W + R > N guarantees reading the latest write. This balances consistency with availability.

N = 5  # Total replicas

# Strong consistency: W=5, R=1
# All replicas must acknowledge writes, any replica can serve reads
# High write latency, low read latency

# Balanced: W=3, R=3
# Quorum writes, quorum reads
# Moderate latency for both

# High availability: W=1, R=5
# Any replica can accept writes, all replicas must agree on reads
# Low write latency, high read latency

# Eventual consistency: W=1, R=1
# Any replica for both — fastest, least consistent

Replication Lag

Replication lag is the delay between a write at the leader and its appearance at a follower. Lag ranges from sub-milliseconds (synchronous) to seconds (async). During high load, lag can spike. Clients reading from followers may see stale data.

• Read-after-write consistency — Route reads to the leader after a write.
• Monotonic reads — Ensure a client always reads from the same or newer replica.
• Bounded staleness — Read from a follower only if lag is below a threshold.
• Causal consistency — Order reads based on causal relationships.

Choosing a Topology