I did finish the Fedor Pikus talk about lock-free programming, and while I'm not qualified to do the most fine-toothed analysis, I have a few observations.
The first, of course, is that there is no one size that fits all when it comes to concurrency, and much like allocation, specificity is better than generality. I'll say more about this in a minute, but don't reach for an MPMC queue if an SPSC will do. The talk is mainly about a different axis - different underlying primitives work well under low-contention vs high contention, locks inherently affect worst-case latency, etc.
Next, on a hardware level, there's no atomic writes bigger than a word. Everything else is physically subject to tearing. Both push and pop are abstractions over actually two states: "I am using this" and "I'm done and releasing it back". All the "empty" space from the write head onward to the read head (remember, this is circular) is, from a reader perspective, in a quantum superposition of being written to vs not.
@MaddieM4 I haven't seen the talk in question so maybe I'm missing something, but "no atomic writes bigger than a word" is over-simplified, right? Outside the CPU core, the modern atomic unit is the cache line; and from the perspective of instruction sets, some provide double compare-and-swap or even transactional memory as primitives. I think transactional memory is particularly interesting to study from this perspective because it makes the existing cache coherency protocol provide the synchronization you need—though I've never heard of anyone seriously using it and I'm not sure anyone should, I just think it's a really interesting point in the design space.