Performance Characteristics of Consensus
You'll learn why consensus algorithms aren't speed demons and where they shine (hint: not high-throughput writes).
The Latency Cost of Agreement
Consensus protocols like Raft or Paxos require multiple round trips across the network before any write completes. A leader must send proposals to a majority of nodes, wait for acknowledgments, then confirm commitment. This typically means:
- 2-3 network round trips minimum per write operation
- Latency measured in milliseconds to tens of milliseconds (cross-datacenter scenarios push this higher)
- The slowest responding node in your quorum determines your write speed
Think of it like passing a law: you can't just decide alone—you must propose it, collect votes from a majority, then officially enact it. This democratic process is safe but inherently slower than a dictator making instant decisions.
Read Optimization Paths
Reads can be faster than writes if you're willing to trade guarantees:
- Stale reads from any follower: instant, but might be outdated
- Linearizable reads from the leader: requires a heartbeat round trip to confirm leadership
- The leader itself already knows the committed state, so local reads are temptier (but dangerous without verification)
Why Consensus Isn't for High-Throughput Writes
Consensus algorithms excel at correctness over speed. They're perfect for:
- Leader election (happens rarely)
- Configuration updates (infrequent)
- Distributed locks (low volume)
They struggle with:
- High write throughput workloads (every write is a coordination ceremony)
- Latency-sensitive applications expecting sub-millisecond responses
- Write-heavy databases needing thousands of operations per second per node
Key Takeaway: Consensus trades write performance for strong safety guarantees—use it for coordination tasks and critical metadata, not high-volume transactional workloads.