If done exceptionally well, the only visible evidence of the leak will be substantial performance disparities between seemingly symmetric features (well done! this is unusual).
If done with the normal level of software design quality, the evidence will show up as quirky behavior, imperfect symmetry, "well this one always works but the other one is not reentrant", etc.
When a reference count hits zero, that's when refcounting begins to be concerned with a dead object; and that part of its operation corresponds to the sweep activity in garbage collection, not to the tracing of live objects.
It is not a dual to garbage collection concerned with its negative spaces; it's simply a less general (we could legitimately say lesser) garbage collection that doesn't deal with cycles on its own.
There are different implementation and performance trade-offs associated with both. I’ll focus on the two that are most meaningful to me.
Reference counting can be added as a library to languages that don’t want or can’t have a precise garbage collector. If you work in C++ (or Rust), it’s a very viable way to assure that you have some measure of non-manual clean up while maintaining precise resource control.
Similarly, when performance matters reference counting is essentially deterministic much easier to understand and model.
In a lot of situations, garbage collection is an overall better strategy, but it’s not a strict superset, and not always the right choice.
In any case, if one is doing GC in such a language, a full tracing collector (whether copying or mark & sweep) is madness, as to find live references means walking nearly the entire heap including the portions living in secondary storage, and now you're in a world of pain.
In this case, an intelligent cycle collecting garbage collector in the Bacon style was the answer. You keep in in-memory table of reference counts, and you only trace when you hit cycles. [and hopefully design your language semantics to discourage cycles]
How do you tell when you've hit a cycle?
> hopefully design your language semantics to discourage cycles
Why? Cyclical structures can be very useful. For example, it can be very handy in many situations for a contained object to have a back-pointer to its container.
[UPDATE] Two responses have now pointed out that this particular case can be handled with weak pointers. But then I can just point to a general graph as an example where that won't work.
That's not a true cycle, it's just a back link for which "weak" reference counts suffice. The containment relation implies that the container "owns" the object, so we don't need to worry about the case where the container might just go away without dropping its contents first.
(Edit: I agree that when dealing with a truly general graph some form of tracing is the best approach. These problem domains are where tracing GC really helps.)
Does it frequently need an owning reference though or would a weak reference suffice? Usually the latter situation suffices.
But then I'll just choose a different example, like a general graph.
I can see a periodic compacting phase could be useful in a system like that.
In the DB world there's good research around similar topics. e.g. LeanStore and Umbra -- Umbra in particular does some nice things with variable sized buffers that I believe are expected to help with fragmentation https://db.in.tum.de/~freitag/papers/p29-neumann-cidr20.pdf