A complex system is safer to grow through working intermediate states than to introduce in one big bang: on the evolutionary path every step yields a working system you can measure and roll back, whereas with a big bang the first contact with reality coincides with the launch date.
The source is John Gall, General Systemantics (1975; later editions as Systemantics and The Systems Bible): “A complex system that works is invariably found to have evolved from a simple system that worked. A complex system designed from scratch never works and cannot be patched up to make it work.” Gall was a pediatrician writing a deliberately humorous book about how systems fail; the “law” is an aphorism drawn from cases, not a measured regularity, and “never” is rhetoric. What makes it useful is the mechanism behind it: a design contains assumptions about load, data and users that can only be tested by running it, and a big-bang launch tests all of them at once, with no working state to fall back to.
In practice: a walking skeleton (Alistair Cockburn’s term) first — the thinnest end-to-end path that runs in production — then increments. For a new event pipeline that means one event type flowing from producer through the broker to one table, with deploy, alerting and backfill working, before the second event type or the clever partitioning. For a replacement system it means routing a slice of traffic to it while the old one still runs (73) and releasing behind a flag (72).
Limits. Some decisions cannot be grown into: a storage format, a partition key, a public API shape (20). Those need thought before the first step, because the simple system that works will fix them in place. Incrementalism without direction also fails — a sequence of working states can wander into a shape nobody chose. Gall says how to get there, not where to go.