The Fundamental Mechanism of Scaling Is Partitioning
The fundamental tool that makes distributed systems outperform a single machine is not consensus or replication it is coordination avoidance. Scaling comes from finding ways to make progress without requiring machines to talk to each other.
"The fundamental approach used to scale distributed systems is avoiding coordination. Finding ways to make progress on work that doesn't require messages to pass between machines." Marc Brooker
A common misconception is that Paxos, Raft, and other consensus protocols are the tools that build the largest systems. These protocols are essential building blocks for availability, durability, and integrity, but they do not make systems scale. In fact, they do the opposite they require coordination between nodes, which inherently limits throughput to the speed of inter-node communication. Scaling comes from partitioning work so that independent pieces can proceed without waiting for each other.
This principle manifests everywhere. Amazon recognized it in 1998 when the Distributed Computing Manifesto proposed moving from a monolithic architecture to services with workflow-based processing, where "instead of processes coming to the data, the data would travel to the process." DynamoDB scales because each partition can handle reads and writes independently. Availability Zones work because infrastructure is partitioned into independent failure domains. Even within a single machine, CPU caches work because they allow each core to operate on its own copy of data without coordinating with other cores on every access.
The optimization framework follows from this principle. Prefer partial order over total order not every event needs a globally defined position. Embrace monotonicity immutable data and idempotent operations are "pre-replicated throughout the universe" and establish consensus without communication. Use optimistic assumptions that avoid or delay coordination, and reserve pessimistic (coordinating) approaches for the cases where you truly need global agreement. Every lock, every consensus round, every cross-partition transaction is a tax on scalability. The art is minimizing how often you must pay it.
Takeaway: Scale by finding ways to avoid coordination, not by making coordination faster the work that does not require a message between machines is the work that scales.
See also: Choose Boring Technology | Static Stability Over Dynamic Failover | Tail Latency Dominates User Experience