Skip to main content

​
What is DVC?

Data Version Control (DVC) is a command-line tool and VS Code extension designed to help you develop reproducible machine learning projects. Think of it as “Git for data” — it brings version control capabilities to large files, datasets, and ML models while leveraging Git for code and pipeline tracking.
DVC is an open-source tool distributed under the Apache 2.0 license, maintained by Iterative and used by thousands of ML practitioners worldwide.

​
Why Use DVC?

Machine learning projects face unique challenges that traditional version control systems weren’t designed to handle:

Data is Too Large for Git

Datasets often exceed gigabytes or terabytes, making them impractical to store in Git repositories.

Models Need Versioning

ML models are binary artifacts that evolve through experimentation and need proper version tracking.

Reproducibility is Critical

You need to reproduce any experiment, model, or result from any point in your project’s history.

Collaboration is Complex

Teams need to share large datasets and models without bloating repositories or manual file transfers.
DVC solves these problems by providing Git-like workflows for data and models while keeping your repository lightweight.

​
Core Features

​
1. Version Your Data and Models

DVC stores version information about your data in Git, while the actual files live in a cache or remote storage. 
# Track a dataset with DVC
dvc add data/train.csv

# Git tracks only the metadata
git add data/train.csv.dvc data/.gitignore
git commit -m "Add training data v1"
When you run dvc add, DVC creates a .dvc file containing metadata (like a hash of your data) and adds the actual data file to .gitignore. This keeps your Git repository clean while maintaining full version control.

​
2. Build Reproducible Pipelines

DVC pipelines connect your code and data into a computational graph, tracking dependencies and outputs. 
# Create a preprocessing stage
dvc stage add -n preprocess \
  -d data/raw.csv \
  -o data/processed.csv \
  python preprocess.py

# Create a training stage
dvc stage add -n train \
  -d data/processed.csv \
  -d train.py \
  -o models/model.pkl \
  -M metrics.json \
  python train.py
DVC only reruns stages when their dependencies change, making your workflows efficient and reproducible.

​
3. Track Experiments Locally

Run and compare multiple experiments without servers or external tools. 
# Run an experiment
dvc exp run -n baseline

# Modify hyperparameters and run again
dvc exp run -n experiment-1

# Compare all experiments
dvc exp show
DVC stores experiment information in your local Git repository, so you don’t need any external experiment tracking server. Share experiments by sharing Git commits.

​
4. Share Data Seamlessly

Configure remote storage once, then push and pull data like you do with Git. 
# Add remote storage (S3, Azure, GCS, SSH, etc.)
dvc remote add -d storage s3://mybucket/dvcstore

# Push data to remote
dvc push

# Teammates can pull your data
dvc pull

​
How DVC Works

DVC uses three key concepts to enable data version control:
1

Git for Metadata

DVC stores lightweight .dvc files in Git that contain metadata about your data (file hashes, locations, dependencies).
2

Cache for Data

Actual data files are stored in a local .dvc/cache directory using content-addressable storage (files are named by their hash).
3

Remote for Sharing

To collaborate, DVC can push/pull data files to/from remote storage (S3, Azure Blob, Google Cloud Storage, SSH servers, etc.).
DVC workflow diagram
This architecture means:
  • Your Git repository stays lightweight (only metadata)
  • Data is deduplicated automatically
  • You can work offline with cached data
  • Remote storage is flexible and cost-effective

​
Key Analogies

If you’re familiar with other tools, here’s how DVC compares:
Like Git-LFS but without requiring a special server. DVC works with any cloud storage or SSH server, and integrates seamlessly with your Git workflow.
DVC pipelines are like Makefiles for machine learning. They describe how data and model artifacts are built from code and other data, with automatic dependency tracking.
Instead of MLflow or Weights & Biases servers, DVC turns your machine into an experiment management platform using Git as the backend.

​
What DVC Doesn’t Do

DVC is not:
  • A database or data warehouse
  • A data processing framework (like Spark or pandas)
  • A model serving platform
  • A full MLOps orchestration system (though it integrates with them)
DVC focuses on version control, reproducibility, and collaboration for ML projects.

​
Use Cases

DVC excels in these scenarios:
  • Solo ML practitioners who want to version experiments and maintain reproducibility
  • ML teams collaborating on shared datasets and models
  • Research projects requiring reproducible results for publications
  • Production ML systems where model provenance and data lineage are critical
  • Model registries needing to track which data and code produced each model

​
Integration Ecosystem

DVC works alongside your existing tools:

Git Platforms

GitHub, GitLab, Bitbucket — DVC leverages your existing Git workflow

Cloud Storage

AWS S3, Azure Blob, Google Cloud Storage, and 10+ other storage types

VS Code Extension

Visual interface for experiments, plots, and data management

CI/CD Systems

Jenkins, GitHub Actions, GitLab CI — automate model training and testing

​
Getting Started

Ready to try DVC? Here’s what to do next:

Install DVC

Set up DVC on your system using pip, conda, brew, or other package managers.

Quick Start Tutorial

Follow our hands-on tutorial to track data, build pipelines, and push to remote storage.

Command Reference

Explore all available DVC commands and their options.

Python API

Learn how to use DVC programmatically in your Python scripts.

​
Community and Support

Join thousands of ML practitioners using DVC:
DVC is actively developed with frequent releases. Star the GitHub repository to stay informed about updates.

Build docs developers (and LLMs) love

Get started for freeTalk to us