A topic in the Open Knowledge Graph — a free, open map of 15,290 topics and the order to learn them in.
Replication strategies (primary-backup, multi-leader, leaderless, active-passive) represent different points on tradeoffs between consistency, availability, and latency. The choice depends on whether writes must be linearizable, whether replicas can drift temporarily, and how failures should be handled.
Create a matrix: each strategy as a row, each property (write latency, consistency, failure recovery) as a column. Mark which are strong, weak, or medium. Understand why stronger consistency often means lower availability.
From your study of consistency models, you know that distributed systems face a fundamental tension: keeping copies of data in sync requires coordination, and coordination costs time. Replication strategies are the concrete architectural choices that navigate this tension. Each strategy makes a different bet about what matters most — low latency, strong consistency, high availability, or simple failure recovery — and no strategy wins on all fronts simultaneously.
Primary-backup replication (also called leader-follower) is the most intuitive approach: one node handles all writes and forwards updates to replicas. This gives you a single source of truth, making consistency straightforward — reads from the primary are always up to date. The cost is that the primary is a bottleneck and a single point of failure. If the primary crashes, you need failover, and during failover you must choose: do you promote a replica that might be slightly behind (risking lost writes), or do you wait until you are certain it is fully caught up (risking downtime)? This is the consistency-availability tradeoff made concrete.
Multi-leader replication allows writes at multiple nodes, which is appealing for geographically distributed systems where routing all writes to one datacenter adds unacceptable latency. The price is write conflicts — two leaders might accept conflicting updates to the same record simultaneously. You need a conflict resolution strategy: last-writer-wins (simple but lossy), application-level merging (correct but complex), or conflict-free replicated data types (CRDTs, which sidestep the problem for specific data structures). Multi-leader replication trades the simplicity of a single write path for lower write latency across regions.
Leaderless replication (as used in Dynamo-style systems) pushes the tradeoff further: any node can accept reads and writes, and the system uses quorum rules to determine success. A write succeeds if W out of N replicas acknowledge it; a read succeeds if R replicas respond, and the system requires W + R > N to guarantee overlap between read and write sets. This maximizes availability — no single node's failure blocks the system — but consistency becomes probabilistic and depends on tuning W, R, and N. The right replication strategy depends entirely on your workload: a banking ledger demands primary-backup with synchronous replication; a social media timeline can tolerate leaderless eventual consistency. The matrix of strategies is not a ranking — it is a map of tradeoffs that your application requirements navigate.