MLOps DSL

A domain-specific language (DSL), implemented in Rascal, for describing ML pipelines (loading, splitting, feature selection, transformation, model training, evaluation, deployment, and monitoring) in a declarative way.

Repository structure

This repository is a monorepo with two tightly coupled sub-projects:

.
├── mlopsdsl/            The actual language implementation (Rascal, Maven)
└── mlopsdsl-extension/  VS Code extension that packages mlopsdsl as a standalone JAR

Both folders belong together and are intentionally versioned together: mlopsdsl-extension contains no logic of its own — only packaging and activation code for exactly the JAR built from mlopsdsl.

mlopsdsl — Language implementation

Contains:

• Grammar (Syntax.rsc) — parser for .mlops files
• Type checking (TypeSystem.rsc, Checker.rsc) — including schema inference from CSV files via an external Python script
• Python bridge (PythonBridge.rsc) — starts/communicates with Python helper scripts (src/main/python/) at runtime
• Language services (Plugin.rsc) — registers parser, type checking, code lenses, and execution with the Rascal LSP

Requirements

• Java 11+ and Maven
• Rascal VS Code extension 0.13.5 (for the main() test mode during development)

Building

cd mlopsdsl
mvn clean package

Registering manually for testing

In the Rascal terminal inside VS Code:

import Plugin;
main();

This loads the language and its Python scripts directly from the local source folders (src/main/rascal, src/main/python) — including a virtual Python environment created locally under src/main/.mlopsenv.

mlopsdsl-extension — VS Code Extension

Packages the JAR built from mlopsdsl together with the Python scripts (excluding .mlopsenv) into an installable .vsix file, so users can use the language without their own Rascal installation.

Requirements

• Node.js + npm

Building

cd mlopsdsl-extension
npm install
npx vsce package     # produces the .vsix after running npm run prepackage (producing JAR) and npm run compile (compiling TypeScript) automatically

Runtime behavior

• Java: Automatically detected by the underlying library (@usethesource/rascal-vscode-dsl-lsp-server); if missing, VS Code automatically offers a download/installation dialog.
• Python: Managed by the extension itself. On first opening a .mlops file, it checks whether a virtual environment already exists at VS Code's own persistent storage location (globalStorage); if not, it creates one automatically.

Language example

pipeline example {
  load(path = "data/raw.csv", target = "label")
  split(train_size = 0.8, random_state = 42)
  select(features = ["age", "income", "score"])
  transformation(
    fillna("age", mean),
    encode("income", onehot),
    scale("score", std)
  )
  model LinReg(dir = "models/linreg", learning_rate = 0.01)
  evaluation(accuracy, f1)
  deployment(port = 8080)
  monitoring(
    driftKS(feature = "age", window = 1000) <= 0.05,
    latency <= 200
  )
}

License

TODO: add license.