True, there's always another way. But you can say same about almost any subject. In my experience about lock-free (and wait-free) programming, safe and high performance deallocation of the objects is often a problem. (Especially in kernel mode when context switching is not an option.)
Whenever high performance is required and it's possible to avoid atomics (CAS, FAA, LL/SC, etc.), (false) cache line sharing or even memory operations altogether, I'll take it. Not to mention CAS (or equivalent) loops and mutexes...