Federated Learning with a C++ Backbone and a 9GB Docker Habit

What it does MetisFL is a federated learning framework with a C++ backend and Python frontend. It orchestrates model training across distributed clients without pooling raw data in one place. The README is heavy on Docker setup and light on architecture specifics, so exactly how it secures gradients or handles stragglers remains unclear from the docs.

The interesting bit Most open-source federated learning stacks are pure Python; MetisFL claims to be the first to mix a C++ implementation with Python bindings. That could mean lower-level control over networking and encryption, though the README never quite spells out the payoff.

Key highlights

• C++ core with Python 3.8–3.10 bindings
• Distributed via Docker images clocking in at roughly 9 GB without CUDA and 12 GB with it
• Built with Bazel; targets Ubuntu 22.04 and RockyLinux 8
• BSD-3-Clause license
• Academic backing via an arXiv citation

Caveats

• The README is almost entirely installation scaffolding; actual API details, supported aggregation schemes, and communication protocols are missing
• Library inconsistencies across operating systems forced a Docker-only recommendation, which suggests cross-platform builds are still brittle
• Project planning lives on a public Trello board rather than GitHub issues

Verdict Worth a look if you want a compiled federated learning backend with lower-level control. Skip it if you need turnkey Python ergonomics or exhaustive documentation.