Distributed systems do not become reliable by adding services. They become reliable when each service has a boundary that can be understood, tested, and operated on its own.

Boundaries are contracts

A boundary is not a network hop. It is a promise about what crosses it: a schema, a latency expectation, an ownership rule. If the promise is implicit, the boundary is decorative.

Prefer fewer, stronger seams

Two well-understood services with a clear contract are usually better than six small services with a shared database and a shared on-call rotation. The goal is not maximum decomposition; it is minimum ambiguity.

Design the failure

For every boundary, write down what happens when the other side is slow, absent, or wrong. If the answer is “it depends,” the boundary needs work. A system that fails in a predictable shape is easier to operate than one that fails in a novel shape every time.