How does garbage collection work, and when does it become a problem?

A garbage collector finds objects still reachable from roots such as the stack and globals, and reclaims everything else. Most modern collectors are generational, built on the observation that most objects die young, so they collect a small nursery often and the old generation rarely. It becomes a problem when pause times eat your latency budget, when allocation outruns collection, or when you leak by holding references you forgot about.

The mental model is reachability. The collector starts at the roots, the stack, registers and globals, walks everything reachable, and anything it never reached is garbage. Most collectors are generational because the vast majority of objects die almost immediately, so you scan a small young space often and touch the old space rarely. That is why allocating a short lived object is close to free, and why a long lived cache is the expensive thing. It turns into a problem in three ways. Pauses, if a stop the world collection lands inside your p99 budget. Allocation rate, where you produce garbage faster than the collector reclaims it and throughput collapses. And leaks, which in a managed runtime always mean something still holds a reference, usually a static map or a listener nobody removed. When I have chased this I start with GC logs and watch heap occupancy after each collection. If that line trends up over hours it is a leak, and then it is a heap dump to find the holder.