Not totally sure what you're looking for, but what I did when I made a package for arena allocators in Julia, I made the arena task-local rather than global. T

Not totally sure what you're looking for, but what I did when I made a package for arena allocators in Julia, I made the arena task-local rather than global. That means that if you spawn a concurrent task and then request the arena to allocate something to, you'll be writing to a completely separate chunk of memory than the memory used by the arena in the parent task. Users can still cause race conditions by explicitly passing an existing arena across task boundaries, but if they use the function to fetch the currently-active task-local allocator, it's naturally threadsafe without any slow locks or atomics required. Ref: https://github.com/MasonProtter/Bumper.jl https://github.com/MasonProtter/Bumper.jl