What is DVC?

DVC builds upon Git by introducing the concept of data versioning – large files that should not be stored in a Git repository, but still need to be tracked and versioned. It leverages Git's features to enable managing different versions of data, data pipelines, and experiments.

DVC does not replace Git! DVC metafiles such as dvc.yaml and .dvc files serve as placeholders to version data and ML pipelines. These files change along with your data, and you can use Git to place them under version control as a proxy to the actual data, which is stored in a cache (outside of Git).

DVC does, however, provide several commands similar to Git such as dvc init, dvc add, dvc checkout, or dvc push, which interact with the underlying Git repo (if one is being used, which is not required).

• DVC does not require special servers like Git-LFS demands. Any cloud storage like S3, Google Cloud Storage, or even an SSH server can be used as a remote storage. No additional databases, servers, or infrastructure are required.

• DVC does not add any hooks to the Git repo by default (although they are available).

• Git-LFS was not made with data science in mind, so it doesn't provide related features (e.g. ML pipelines, metrics, etc.).

• GitHub (common Git hosting service) has a limit of 2 GB per repository.

• DVC can use reflinks* or hardlinks (depending on the system) instead of symlinks to improve performance and the user experience.

• Git-annex is a datafile-centric system whereas DVC focuses on providing a workflow for machine learning and reproducible experiments. When a DVC or Git-annex repository is cloned via git clone, data files won't be copied to the local machine, as file contents are stored in separate remotes. With DVC however, .dvc files, which provide the reproducible workflow, are always included in the Git repository. Hence, they can be executed locally with minimal effort.

• DVC optimizes file hash calculation.

• DVC enables a new experimentation methodology that integrates easily with standard Git workflows. For example, a separate branch can be created for each experiment, with a subsequent merge of the branch if the experiment is successful.

• DVC innovates by giving users the ability to easily navigate through past experiments without recomputing them each time.

Systems to manage data pipelines and dependency graphs such as Airflow, Luigi, etc.

• DVC is focused on data science and modeling. As a result, DVC pipelines are lightweight and easy to create and modify. However, DVC lacks advanced pipeline execution features like execution monitoring, error handling, and recovering.

• dvc is purely a command line tool without a graphical user interface (GUI) and doesn't run any daemons or servers. Nevertheless, DVC can generate images with pipeline and experiment workflow visualizations.

• See also our sister project, CML, that helps fill some of these gaps.

See also the Experiment Management guide.

• DVC uses Git as the underlying version control layer for data, pipelines, and experiments. Data versions exist as metadata in Git, as opposed to using external databases or APIs, so no additional services are required.

• DVC doesn't need to run any services. There's no built-in GUI as a result, but we also have our sister project DVC Studio to fill that gap.

• DVC can generate images with experiment workflow visualizations.

• DVC has transparent design. DVC files have a human-readable format and can be easily reused by external tools.

Make and others.

• File tracking:

  • DVC tracks files based on their hash values (MD5) instead of using timestamps. This helps avoid running into heavy processes like model retraining when you checkout a previous version of the project (Make would retrain the model).

  • DVC uses file timestamps and inodes* for optimization. This allows DVC to avoid recomputing all dependency file hashes, which would be highly problematic when working with large files (multiple GB).

• DVC utilizes a directed acyclic graph (DAG):

  • The dependency graph is defined implicitly by the connections between stages, based on their dependencies and outputs.

  • Each stage defines one node in the DAG, and dvc.yaml files contain these stage definitions (think Makefiles). All stages (and corresponding processes) are implicitly combined through their inputs and outputs, simplifying conflict resolution during merges.

  • DVC stages can be written manually in an intuitive dvc.yaml file, or generated by the helper command dvc stage add, based on a terminal command, its inputs, and outputs.

* Inodes are metadata file records to locate and store permissions to the actual file contents. See Linking files in this doc for technical details (Linux).