DVC

DVC versions large files and datasets by storing actual content in a local cache indexed by content hash (SHA256), while tracking lightweight .dvc metadata files in Git. The system uses a two-tier architecture: Git manages .dvc files (text-based pointers with checksums), while a separate cache layer stores deduplicated file content. This enables efficient storage, deduplication, and seamless Git integration without modifying Git's core behavior.

DVC abstracts remote storage through a provider-agnostic interface supporting S3, GCS, Azure Blob Storage, HDFS, SSH, and local paths. The system uses a push/pull synchronization model where data flows between local cache and remote storage via configurable backends. Each remote is defined in .dvc/config with connection credentials, and the sync layer handles authentication, retry logic, and partial transfers without requiring manual cloud SDK management.

DVC maintains an in-memory Index of the repository state, built from dvc.yaml and dvc.lock files. The Index class provides efficient querying of stages, dependencies, and outputs without re-parsing files. This enables fast operations like 'which stages depend on this file?' or 'what are all outputs of this stage?'. The Index is rebuilt when dvc.yaml or dvc.lock changes, and caching prevents redundant rebuilds.

DVC can import data from external sources (HTTP URLs, S3, GCS, etc.) and track it as a dependency. The system downloads the data, computes its hash, and stores it in the cache. Subsequent runs use the cached version unless the source has changed. This enables pipelines to depend on external datasets without manually downloading them. The import mechanism supports versioning by URL, enabling reproducible imports of specific data versions.

DVC provides real-time progress reporting during long-running operations (data transfer, pipeline execution) through a progress reporting system that displays download/upload speed, ETA, and completion percentage. The system uses streaming output to avoid buffering large amounts of data in memory. Progress bars are rendered to the terminal with support for different output formats (plain text, colored, machine-readable).

DVC exposes a Python API through the Repo class, enabling programmatic access to all DVC operations (add, push, pull, run, reproduce, etc.). Users can import dvc.api or instantiate Repo objects to interact with DVC without using the CLI. This enables integration with Jupyter notebooks, custom scripts, and external tools. The API mirrors CLI functionality but provides Python-native interfaces and return values.

DVC pipelines are defined in dvc.yaml as directed acyclic graphs (DAGs) where each stage specifies dependencies (inputs), outputs, and the command to execute. The system builds an in-memory DAG representation (via Index and Stage classes) and uses file hash comparison to determine which stages need rerunning. Only stages with changed dependencies are re-executed, with results cached by output hash, enabling fast iteration on large ML workflows.

DVC tracks ML experiments by capturing parameters (from params.yaml or code), metrics (accuracy, loss, etc.), and outputs at experiment time. Each experiment is stored as a Git branch or commit with associated metadata, enabling side-by-side comparison of different model configurations. The system extracts metrics from files (JSON, CSV, YAML) and parameters from structured files, then computes diffs to highlight which parameter changes led to metric improvements.

DVC tracks file dependencies (inputs to stages) and outputs (results of stages) by computing and storing their content hashes in dvc.lock. When a stage is re-run, DVC compares current file hashes against stored hashes to detect changes. The Output class manages this tracking, associating files with checksums and handling different output types (regular files, metrics, plots). This enables DVC to determine which stages are out-of-date without re-executing them.

DVC extracts metrics (scalar values like accuracy, loss) and plots (visualization data like confusion matrices) from user-generated files in JSON, CSV, or YAML formats. The system parses these files at experiment time and stores the extracted values in experiment metadata. Metrics are compared across experiments to identify improvements, while plots are aggregated for visualization. This approach avoids requiring users to use a specific logging API.

DVC generates visual representations of pipeline DAGs showing stage dependencies and execution status. The system reads dvc.yaml and dvc.lock to build an in-memory graph, then outputs it in formats suitable for visualization (text-based ASCII, Mermaid, or graphviz). Status reporting shows which stages are up-to-date, which need re-running, and which have failed. This enables users to understand pipeline structure and identify bottlenecks without reading YAML files.

DVC manages configuration through a multi-level system: system-wide defaults, repository-level .dvc/config, and local .dvc/config.local (not committed to Git). Configuration includes remote storage settings, cache paths, and feature flags. The system supports environment variable substitution in config values, enabling dynamic configuration without hardcoding credentials. The ConfigError class handles validation and precedence resolution.

DVC abstracts file system operations through a unified interface supporting local paths, cloud storage (S3, GCS, Azure), HDFS, and SSH. The FileSystem class provides a consistent API for operations like exists(), open(), remove() regardless of backend. This abstraction enables DVC to work transparently with different storage systems without duplicating logic. Path handling is normalized across platforms (Windows, Linux, macOS).

DVC integrates with Git through the SCM (Source Control Management) abstraction, enabling operations like checking out branches, reading commit history, and detecting uncommitted changes. The system uses Git as the source of truth for code and metadata (.dvc files), while data lives in separate storage. Experiment execution creates Git branches automatically, and the system can reproduce experiments by checking out specific commits. This tight Git integration enables full reproducibility.