UPI's transaction growth has outpaced almost every internal capacity model built five years ago. Switches designed for the volumes of 2021 have needed multiple emergency scaling exercises since — and each one is more disruptive than the last if the underlying architecture wasn't built to scale horizontally from day one.
Stateless routing beats clever routing
The switches that scaled smoothly share one trait: routing logic that doesn't depend on in-memory state tied to a specific node. That sounds obvious, but a surprising number of switches accumulated per-node caching or session affinity over time as quick fixes for performance problems, and those shortcuts are exactly what makes horizontal scaling painful later.
Retry logic is a capacity problem too
Naive retry-on-timeout logic quietly multiplies load during exactly the moments a switch is already under stress — a downstream rail slows down, timeouts trigger, retries pile on top of an already-slow system, and a localized slowdown turns into a cascading failure.
Exponential backoff with jitter, and circuit breakers that fail fast to a queued-retry path instead of hammering a struggling rail, are the difference between a rail-level hiccup and a system-wide incident.