When NOT to Use Consensus
What you'll learn: You'll discover scenarios where consensus algorithms add unnecessary complexity and slower alternatives serve you better.
The Core Idea
Consensus algorithms like Raft and Paxos provide strong guarantees — linearizability, total ordering, and fault tolerance — but at a steep cost: latency, complexity, and reduced availability during partitions. Not every problem needs these guarantees. Sometimes simpler models like eventual consistency, leaderless replication, or even manual processes are smarter choices.
When to Skip Consensus
High-Throughput, Low-Latency Systems
If you're building a shopping cart, user preferences, or session data where temporary inconsistency is acceptable, eventual consistency is faster and more scalable. You don't need all replicas to agree immediately.
Geographically Distributed Writes
Consensus requires majority agreement, which means cross-datacenter latency on every write. For global applications where users expect local-speed writes (social media posts, document edits), leaderless replication with conflict resolution wins.
Read-Heavy Workloads
If 99% of operations are reads and writes are rare, consensus overhead on every write may be wasteful. Use simpler replication with occasional manual reconciliation instead.
Tolerable Manual Intervention
Configuration changes that happen monthly? Bank reconciliations done overnight? Human-in-the-loop processes can replace automated consensus for infrequent coordination.
Think of It Like This
Consensus is like召集 (convening) a board meeting for every decision — thorough but slow. If you're just collecting suggestions (eventual consistency) or letting anyone post sticky notes (leaderless writes), you move faster at the cost of immediate agreement.
Key Takeaway: Use consensus only when you truly need strong ordering, linearizability, or fault-tolerant coordination. For most use cases involving user-facing data, caching, or analytics, eventual consistency or leaderless models deliver better performance without the complexity tax.