As a direct answer to your question: there is the :array module from erlang.

That module uses a set of nested tuples and reading is O(1). I’m not sure about its writing runtime though, and you’d have to implement filling logic on delete by yourself (either take the last in the hole, or let every following item go one step forward, thats up to you, depending on if original order matters).

There is also the array package which is a wrapper around the erlang module, providing some protocols as well to make it easier to handle them from elixir site. But it hasn’t been updated in the last 3 years, therefore it might spit some warnings in a current elixir or not work at all.

To try to find another solution to your problem. It sounds a bit as if you want to create some kind of load balancer which is giving away its work randomly on the workers.

By picking randomly from a fixed set, you might get non-uniform distribution due to how PRNGs work.

Therefore I’d suggest one of these solutions:

1. Only shuffle once, use that as a base for roundrobbin. You can use :queue module for a datastructure which can pull from the head and push to the tail in amortized o(1).
2. Shuffle the original list and keep a copy of it around. Each time you need to send a job, use the head of the shuffled list until it is empty, then reshuffle the original list and keep going.

I have two fine packages for you:

• Arrays exposes a common array interface, with multiple different implementations that you can try if the default performance characteristics are not to your liking. (The two built-in implementations are a map-based implementation and an :arrays-based implementation.)

This is probably perfect for your needs, although the last time I had time to work on it I was unable to write proper documentation and tests, so these are still lacking.

If you need to do more mathematical stuff with them:

• Tensor gives you one-dimensional vectors, two-dimensional matrices and n-dimensional tensors, and defines many common operations on them. Tensor stores these in a sparse format, which may or may not be suited well to your problem domain.

Tensor is very stable and greatly documented and tested.

Does order matter? Do you need to access items by key or just at random? Map has been mentioned, but there is also MapSet if you don’t need keys. Both implement the enumerable protocol, so to get a random value you can pass it to Enum.random. What I’ve done in the past when I want constant time lookups by key and also maintain order is to create two structures: a map for the constant lookups, and a list of keys for the order.

You may also consider looking into using a Registry if you want a container for processes, though for your use case it might be easier to just keep simple collection of pids.