DVC Documentation

What DVC does

DVC is a command-line tool, with a VS Code extension, for making machine learning projects reproducible. Git is built for text, not for gigabytes of data and model files, so DVC keeps those files in a cache and a remote storage location, and puts small pointer files in Git instead. Each commit then points at the exact data it used. On top of that, DVC describes how outputs are produced as a pipeline of stages, and it can track experiments locally without any server.

Who maintains it now

DVC was created by Iterative. In November 2025 lakeFS announced that it had acquired the DVC open source project, and said DVC would continue as an independent open source tool for data scientists, with DVC, DVCLive and the VS Code extension kept maintained. The documentation now lives at doc.dvc.org.

The parts worth learning first

• Data versioning: dvc add tracks a file or folder, dvc push uploads it to a remote such as Amazon S3, Google Cloud Storage, Azure or an SSH server, and dvc pull brings back the version that matches your current Git commit.
• Pipelines: stages in dvc.yaml declare a command, its dependencies, its parameters and its outputs. dvc repro reruns only what changed, and dvc.lock records the exact hashes used.
• Metrics and experiments: dvc metrics diff compares results across commits, and dvc exp run runs experiments you can compare with dvc exp show.

How long it takes

The Get Started section is short and hands-on: data versioning, pipelines and experiments each take an evening. After that, the command reference and user guide are for looking things up.

Prerequisites

Comfort with Git and the terminal, and a Python project that trains something. For a remote, a cloud bucket is typical, but a local folder or network drive works for learning.

Free or paid?

DVC is free and open source under the Apache 2.0 licence. You pay only for whatever storage you use as a remote.

How it fits an MLOps workflow

DVC covers the versioning and lineage part of the lifecycle. It pairs naturally with MLflow for run tracking and the model registry, and with an orchestrator such as Prefect or Airflow when pipelines need schedules and retries rather than a single machine.

Common questions

Is DVC the same as Git LFS? Both keep large files out of Git, but DVC needs no special server, supports many storage back ends and adds pipelines and experiments.

What can I do after it? Rebuild any past model from a commit, prove which data trained the live model, and rerun only the pipeline stages a change affects.

Best for: students and engineers whose models depend on datasets that change, and anyone preparing a reproducible MLOps portfolio project.