Skip to content

Clusterf**ctional — Research

Everything the talk stands on — how BEAM nodes find and talk to each other, how libcluster automates it, and what four years of running it at NETZlive taught us.

Erlang distribution

EPMD, the handshake, cookies, the wire format. What actually happens when two nodes connect. Act 1.

libcluster

Strategies for answering “who else is out there?” — Gossip, DNS, Kubernetes, Epmd. Act 1.

NETZlive in production

Service registration, versioned RPC contracts, error boundaries, two-tier messaging. Act 2.

Trade-offs

The dependency tax, the boilerplate, and when clustering is the wrong answer. Act 3.

A 35-minute talk uses a fraction of what goes into it. This site keeps the rest: the sources read, the experiments run, the claims checked against OTP source, and the things that turned out to be wrong. It is the place to come back to when a slide needs a fact, or when someone in the Q&A asks the question the slide skipped.

Not the talk itself — the outline and abstract live in the parent directory (talk-outline.md, abstract.md). Not a tutorial. Pages here exist to back claims the talk makes, so every page should be traceable to a part of the talk or deleted.