I think the fact that it is complicated to reason about is precisely why systems developers don’t trust GC’s.
It’s far easier to write a threadsafe bump (slab) allocator than to locate and diagnose the code that someone wrote two years ago and, as of last week, started causing the GC to blow up the cache, contend on a lock, fragment the heap, add unbounded pauses, burn an extra cpu, etc, etc.
(Though, at this point, most mallocs are so good that the slab allocator loses anyway and there’s no need to bother.)
It’s far easier to write a threadsafe bump (slab) allocator than to locate and diagnose the code that someone wrote two years ago and, as of last week, started causing the GC to blow up the cache, contend on a lock, fragment the heap, add unbounded pauses, burn an extra cpu, etc, etc.
(Though, at this point, most mallocs are so good that the slab allocator loses anyway and there’s no need to bother.)