A complex system can behave in ways that are hard to predict from its components alone. Interactions matter: a change that looks beneficial in one place may have an unexpected effect elsewhere. Recognizing this changes how you debug problems, measure improvements, and make decisions. section 6.14.1
Three warning signs are:
- Emergent behavior: the whole system does something that analysis of its separate components did not predict. section 6.14.1
- Delayed consequences: an issue introduced by a deployment may appear days or weeks later, making its cause harder to trace. section 6.14.1
- Local versus global optimization: improving one component can leave the overall system unchanged or make it worse. section 6.14.1
Complicated is not the same as complex
Finding the nearest available driver for a passenger is technically difficult, but established approaches such as geo-hashing can be applied. At the scale of a large machine-learning infrastructure, scheduling challenges may instead require new approaches. The distinction matters: a repeatable solution fits a complicated problem, while a complex one may demand adaptation. section 6.14
When changing a complex system:
- Prefer reversible decisions when possible, so you can experiment and adjust. Treat decisions involving security policies or compliance more deliberately. section 6.14.2
- Define both local and global success metrics before making a change. An improvement in the immediate area may conceal harm elsewhere. section 6.14.2
- Stay open to a novel solution when a conventional one does not fit the problem. section 6.14.2
