Basically, the key to this is a combination of Erlang's use of tail-calls as the only means of looping, and the VM's use of what I might call (I'm not actually sure of the name for it) "hybrid budget-based cooperative/pre-emptive scheduling."

Since all loops are implemented as tail-calls, every Erlang function effectively is a straight O(N)-time-for-N-instructions shot that ends in either another call (which either adds, or reuses a stack frame) or a return. This gives the VM the opportunity to act like a pre-emptive machine, while gaining the advantages of cooperative multitasking.

In cooperatively multitasked VMs, where coroutines have to explicitly "yield" to pass the baton, you get a huge advantage: since the instruction-set architecture can be designed to ensure that memory is in a well-defined state whenever you yield, you don't have to do the expensive context-switch thing that pre-emptive architectures do: stashing and unstashing registers, switching out memory descriptors, etc. It can literally just be a jmp instruction.

BEAM is basically a cooperatively-multitasked VM, except that every "call" instruction is also an implicit "yield". (More specifically, it's a "yield if this process has executed >= 2000 call instructions since it received control.") So you get everything nice about cooperative multitasking (you never decode only 0.8 of an audio frame before being interrupted), and everything nice about pre-emptive multitasking (nothing can hog the processor forever[1]), together; thus, soft real-time.

Another place where this design comes into play is in hot code loading. The design of the loader itself is pretty simple: modules are kept in a hash-table, keyed by name. A module is referenced by address on the stack, and in loaded (threaded) bytecode; and by key in unloaded (abstract) bytecode. When you upgrade a module, you just replace the value of the key in the module dictionary. Functions that call other functions in their own module as they're running stay in the previous version of a module; other functions that try to jump into the module--or functions from the module that make a "remote" (fully-qualified) call back into the module--jump into the new version.

This would be a lot harder and more complex, if we didn't have that guarantee that every looping construct in Erlang is implemented in terms of tail-calls. Since they are, we don't have to worry about "interrupting" a process to upgrade it when it still has some dirty state; we can just wait around, and every process will yield after a few microseconds, and we can upgrade it then, when its state is well-defined.

---

[1] --unless you call into C and that C code does something that takes a million years. This is why people tell you to not do CPU-bound things in Erlang. The semantics of BEAM's instruction-set architecture (and therefore its non-optimal speed) is essential to how it multitasks; you have to insert explicit process-yielding checks in your C code if you want it to be "non-blocking." Or you can write your C code as a "port", which means letting the OS manage it as a separate process--which, if you're writing one global matrix-transformer process or one per-MMO-zone physics-engine process, isn't that bad; but if you're writing a per-user speech-to-text analyzer, you've handed your OS the job of managing 100k real native processes. Better to just write it in plain Erlang, let the VM do its concurrency thing, take the speed hit, and scale horizontally a bit sooner than you would have had to with C. (Erlang is great at scaling horizontally.)