Wyrd should be a globally scalable distributed file storage foundation with atomic writes. Built in Rust, Apache-2.0.
wyrd /wɪəd/ — Old English for fate. Yes, it sounds like “weird.” Anyone who has run distributed storage at 3 a.m. will agree that is the correct word.
What it is
Wyrd is the layer a mid-sized cloud provider could build a worldwide storage product on — a Drive-like service, an S3-compatible object store, or the backend for a collaborative editor. It keeps metadata and bulk data architecturally separate, yet every write commits as a single, indivisible act: from the caller’s point of view, a write either happened in full or not at all.
The model follows the Colossus-class lineage — a global control plane over per-zone storage, with bulk data flowing directly between client and storage servers so throughput scales with the fleet rather than through a bottleneck. One static binary on a laptop for development; a multi-region deployment in production. The same system, configured differently.
Why it exists
There are many of systems that do part, but no open source that combines all of it: a rigorous atomic commit point, a globally consistent namespace, erasure coding within a zone and geo-replication across zones, pluggable metadata and chunk backends behind narrow interfaces, an on-disk format specified so the data outlives the software that wrote it — and correctness treated as the headline feature, verified with deterministic simulation testing from day one.
The name
A storage system is, in the end, a keeper of wyrd: it holds what was written, weaves in what is being written, and carries the debts of what it still owes. The whole woven state is the Wyrd; the Norns are the ones who tend it. The Norns weave the Wyrd.