Back to catalog
DVC
Apache-2.0DVC replaces proprietary MLOps platforms like Domino Data Lab with a Git-based CLI for versioning datasets, tracking pipelines, and managing experiments.
Open-source alternative to — compare it against:
Overview
Former Microsoft data scientist Dmitry Petrov created DVC in 2017 to bring Git workflows to machine learning without storing large files in repositories. By tracking datasets and models through lightweight text pointers while offloading raw binaries to cloud object storage, Petrov avoided the need for a proprietary central server. That origin makes DVC feel like native tooling for software developers rather than an isolated data platform.
Why DVC instead of Domino Data Lab
Domino Data Lab approaches machine learning as an all-in-one enterprise platform, requiring specialized infrastructure, managed compute clusters, and proprietary web workspaces. That design creates an operational silo separate from standard developer tools. DVC instead integrates directly into existing software workflows: it versions datasets, model weights, and multi-stage DAGs using plain metadata files tracked in Git. DevOps engineers can plug dvc repro directly into GitHub Actions or GitLab CI runners without deploying extra servers, routing raw artifact storage to Amazon S3, Google Cloud Storage, or Azure Blob buckets under existing IAM policies.
Where Domino excels is unified compute orchestration, hosted Jupyter notebooks, and turn-key model endpoints for non-technical data scientists. DVC leaves compute provisioning and model serving to your infrastructure stack, focusing exclusively on artifact versioning and pipeline reproduction. Teams replacing Domino with DVC gain open-source flexibility, eliminate steep per-seat licensing, and preserve standard Git reviews, but must manage compute resources independently.
What it does well
- Plain-text tracking files version large files in Git without increasing repository size or hitting file limits.
- Deterministic pipeline execution with
dvc repro, caching stage outputs and skipping execution for stages whose dependencies and code have not changed. - Multi-cloud storage support synchronizing cached artifacts with Amazon S3, Google Cloud Storage, Azure Blob, and SSH remotes through push and pull commands.
- Zero-copy checkout using reflinks and hard links on supported filesystems, avoiding duplicated disk usage across local project workspaces.
- Git-native experiment tracking via experiment run and diff subcommands to compare parameters, metrics, and plots across branches without an external database.
