Replication

Copying data across nodes for durability, availability, and read scaling. Can be synchronous (strong, slow) or asynchronous (fast, lossy on failover). Single-leader, multi-leader, and leaderless are the main topologies.

When to use

HA, durability, read scaling, and geo-local reads.
Any data you can't afford to lose.

When not to

When a single copy with backups suffices.
Write-heavy workloads where replicas only add cost.

Definition

Replication keeps copies of data on multiple nodes. A primary accepts writes and ships them to replicas (single-leader); multi-leader allows writes at several; leaderless (Dynamo-style) writes to a quorum. Synchronous vs async sets the durability/latency tradeoff.

Why it exists

One copy means one failure loses everything. Replication adds redundancy (durability/availability) and read capacity, and geographic replicas cut latency. It's the foundation of HA.

How it works

Writes go to the leader, which replicates via a log (WAL/binlog). Sync replication waits for replicas before ack (strong, slower); async acks immediately (fast, may lose recent writes on failover). Replicas serve reads (eventually consistent) and can promote on leader failure.

Scaling limits

Sync replication adds commit latency; async risks data loss on failover. Replication lag causes stale reads. Multi-leader needs conflict resolution. More replicas = more write amplification.

Cost shape

Storage + network per replica. Blow-ups: over-replication, cross-region replication egress, and lag causing read staleness.

Alternatives

Erasure coding — Less storage overhead for durability, but no read scaling or failover.
Backups only — Cheaper but slow recovery and no live redundancy.

Failure modes

Replication lag causing stale reads.
Split-brain on failover.
Data loss with async replication.

WAL (write-ahead log), Quorum, Raft, Eventual consistency, Snapshot