I’m incredibly excited to finally introduce confingy to the world.

• GitHub: https://github.com/runwayml/confingy
• Docs: https://runwayml.github.io/confingy

This is a library that solves a problem that I’ve had for over a decade at every company at which I’ve worked. The library is also the answer to my previous blog post, Why can’t we separate YAML from ML?.

Major shoutout to my employer, Runway, for letting me open source the library. Internally, all research configuration code was migrated off of YAML and onto confingy months ago.

So what actually is confingy?

I’m calling confingy an implicit configuration system for Python. By adding a simple @track decorator to any class:

from confingy import track


@track
class Foo:
    def __init__(self, bar: str):
        ...

you get 3 features:

1. All arguments to the class’ constructor get tracked for later.
2. Lazily-instantiate any tracked class.
3. Serialize tracked classes to JSON and deserialize back into Python.

Who needs this?

confingy is designed for anybody who builds iterative or flexible workflows, such as machine learning model training, data pipelines, DAGs, prompt orchestration, and so on. These workflows often involve YAML configs, OmegaConf, dataclasses, or Pydantic models that define different steps in the workflow. These configs become interfaces to your code and inevitably require duplication of logic for each new step.

People often build out these complicated configuration systems because they need reproducibility reusability. Their systems inevitably grow into a mess of dynamic code instantiation and other separations between configuration and code that actively harm the software architecture, such as by incentizing inheritance over composition.

With confingy, your code is the config.

That’s it?

There is a lot more information in the repo and docs. With the 3 features above as core principles, many benefits fall out of this library. We get type validation for free. There’s a transpiler to go from a serialized JSON config back into actual python code. There’s a visualization server for diffing configs, and a CLI for quick serialization/deseriliazition/transpilation. Please take a look and provide feedback!