Type system improves message safety in elixir actor processes

You Have Mail: Set-Theoretic Types for Actors

Programming Languages

Summary

Elixir programs use many independent parts called processes that talk by sending and receiving messages. The authors designed a way to check what types of messages a process should expect and send, even when messages contain functions using the process’s own mailbox in tricky ways. This system figures out limits for what mailboxes the functions can access, making it easier to find mistakes without extra hassle for programmers. It works by automatically guessing most information and only needs programmers to specify the interface when starting a process.

What this means in practice

Authors

Giuseppe Castagna, Luca Padovani, Daniele Varacca, Amine Nasri

Abstract

We present a type system for the processes of Elixir, a dynamic language based on the actor model whose static type system covers only its functional core. Each process is typed by $\mathtt{pid}(t)$, where the ``interface'' $t$ is the type of the messages the process expects in its mailbox: sends are checked against the recipient's interface, and receives rely on the interface of the process that performs them. One of the difficulties of typing Elixir is that mailboxes are implicit: a function shipped in a message may evaluate $\mathtt{self()}$ and execute receive instructions, and both refer to whichever process eventually runs the function---$\mathtt{self()}$ returns that process's pid, receive reads that process's mailbox---so the mailbox they act on is not known where the function is written. We solve this problem by inferring, for every function, bounds on the type of the mailboxes it can be used with; as a result, the system types higher-order functions that travel in messages even when they contain reference to $\mathtt{self()}$ and receive instructions. The only new annotation the programmer writes is the interface of each spawned process; everything else is inferred by reusing the existing subtyping and constraint-resolution machinery of Elixir, and the system is proved sound. For legacy code, the interface can be reconstructed from the receive clauses: what the branches do with their messages bounds it from above, and the most permissive solution is sharp enough to recover received pids and closures.