mir-backend-beamlisted

Make It Right (BEAM runtime tier). Erlang VM / BEAM reliability footguns shared across every BEAM-based backend (Phoenix, Nerves, pure Erlang) — distinct from the generic backend gates and from any one framework's mechanics. Covers: let-it-crash + supervision tree design (restart strategies, poison-message crash-loops), unbounded mailbox growth and backpressure (GenStage/Broadway/synchronous call), hot GenServer as a serial bottleneck and when to use ETS instead, blocking handle_call timing out callers and queuing all others, isolated per-process heap and how to share state (messages / ETS / persistent_term), and distributed Erlang hazards (netsplits, :global name clashes, pairwise ordering). TRIGGER when the backend runtime is Elixir or Erlang — sits between mir-backend (generic) and the framework module (e.g. mir-backend-beam-phoenix). SKIP for Python, Node, Go, JVM, Rust, .NET, Ruby, PHP runtimes (each has its own mir-backend-<runtime> tier), and for framework-library mechanics (those live in the framework
anantbhandarkar/make-it-right · ★ 12 · API & Backend · score 83

Install: claude install-skill anantbhandarkar/make-it-right

# /mir-backend-beam · Make It Right (BEAM runtime) The middle tier. `mir-backend` decides **what is correct** (any language). The framework module (e.g. `mir-backend-beam-phoenix`) knows the **library's mechanics**. This tier owns what's true for **all BEAM backends because they run on the Erlang VM** — the process model, message-passing concurrency, and fault-tolerance primitives that Phoenix, Nerves, and plain Erlang all inherit. **Runtime assumed:** Elixir 1.16+ on OTP 26+ (the notes hold for Erlang directly; syntax examples are Elixir). Load order: `mir-backend` → `mir-backend-beam` → `<framework module>`. ## The BEAM footguns AI walks into (framework-agnostic) ### 1. Let-it-crash + supervision — design the tree, don't rescue everything The BEAM's fault-tolerance story is built on supervision: a crashed worker process is restarted by its supervisor without taking down the rest of the system. AI frequently undermines this by wrapping every operation in `try/rescue` or `try/catch`, converting clean crashes into silently swallowed errors that leave the process in a corrupt-but-alive state. - **Let unexpected errors crash the process.** A supervisor restart resets state cleanly. A rescued exception that continues execution with invalid state is harder to detect and recover from. - **Design supervision trees intentionally.** Choose the right restart strategy: `one_for_one` (worker failures are independent), `one_for_rest` (failure in one cascades to later-started sibling