Universal Scalability Law

Amdahl's law explains why a serial fraction limits speedup. The universal scalability law adds the term that explains something more alarming: past a certain point, adding capacity makes a system slower, not merely no faster.

Two penalties act on scaling. Contention is the serial portion — a lock, a single writer, a shared queue — and it flattens the curve towards an asymptote. Coherency is the cost of keeping participants consistent with each other, and it grows quadratically with the number of participants, because every node must reconcile with every other. That term eventually dominates and bends the curve downward.

The practical recognition is a system that gets worse when you add nodes: a cluster that performs better at eight members than sixteen, a connection pool that degrades when enlarged, a cache cluster that slows as it grows. That is the coherency term becoming visible, and the instinctive response of adding more capacity makes it worse.

The design implication is that reducing coordination beats adding hardware. Sharding so nodes do not coordinate, partitioning so writes are single-owner, and eventual consistency where the domain allows it all attack the coherency term directly — which is the only thing that changes the shape of the curve rather than moving along it.