Getting Started with roar and GLaaS
On this page
roar automatically tracks the reproducible lineage of your work by observing your commands as they run without requiring you to define a pipeline or write extra configuration.
When paired with GLaaS (Global Lineage-as-a-Service), this observation approach scales to a global registry, allowing teams to search for any artifact by its hash, navigate the connections between jobs, and easily reproduce results on other machines.
Here's a quick breakdown of the documentation:
1. Get Started
Start here to understand the philosophy behind roar and GLaaS, and why we think it's going to change how ML models are developed. This document explains the core concepts of "implicit observation" and "lineage without orchestration".
2. End-to-End Example
Walk through this 6-step example to get a feel for how to use these tools.
- Generate and combine small binary artifacts using
roar run. - Inspect the inferred DAG (Directed Acyclic Graph) of your local pipeline.
- Authenticate and register the final artifact with GLaaS.
- Use a hash to perfectly reproduce the artifact in a completely fresh directory.
3. Core Concepts
Introduces the structural building blocks of the system. Learn how roar defines Jobs (recorded command executions) and Artifacts (files connected by jobs and identified by content hashes). It also explains how these elements form a reproducible lineage (the DAG) that acts as a recipe for recreating artifacts.
4. roar Guide
A practical manual for the roar CLI. It covers everything you need to manage your daily workflow, including:
- Installation: Setting up roar (e.g.,
pip install roar-cli). - Configuration: Initializing your project database with
roar init. - Commands: Using
roar run,roar build, and viewing your current session withroar dag. - Authentication: Step-by-step instructions for linking your GitHub SSH key to GLaaS via
roar authso you can publish your lineages.
5. Use Cases
How does roar and GLaaS solve real-world problems? This page provides workflows for answering critical questions like "Where did this model come from?" and "What data was this trained on?", as well as instructions for perfectly recreating a teammate's artifact using roar reproduce.
6. FAQ
Check the Frequently Asked Questions for deep dives into roar's technical implementation. This page covers how the Rust-based ptrace mechanism works under the hood, explains how roar filters out system noise to recognize true artifacts, discusses the astronomical unlikelihood of hash collisions, and details why roar tracks git commit states.