The site is slow. What is your first step?

First, measure: which resource is saturated and where the time goes. App CPU at 95% and DB idle is a different problem from DB CPU at 95% with app instances waiting on connections. A trace of a slow request tells me whether it is one slow query, many small ones (N+1), or an external call. Half the time the answer is a missing index or a query doing a sequential scan, which is a one-line fix with no architecture change.

Second, the cheapest capacity: vertical scaling. Doubling the instance size takes ten minutes and buys time to do the rest properly. Then horizontal for the stateless tier — move sessions out, add a load balancer, run three instances — which introduces the state problem and the balancer as a component to run. Third, the database, which is where growth actually bites: connection pooling if connections are the limit, a cache for hot repeated reads if the profile is read-heavy, read replicas if the read set is broad and lag is tolerable. Each of those introduces staleness in some form.

Sharding, queues and services come only when a specific measured problem demands them: write throughput beyond one primary, background work distorting request latency, or teams blocking each other. Every rung buys capacity by adding a problem, so I climb only as far as the evidence requires.