Skip to main content

​
What is Data Versioning?

Data versioning in DVC is a lightweight mechanism to track large files and directories without storing them in Git. Instead of bloating your Git repository with massive datasets or model files, DVC stores metadata pointers (.dvc files) in Git while the actual data lives in a cache.
Key Concept: DVC uses content-addressable storage, meaning files are identified by their hash (checksum), not their name. If the content changes, the hash changes, creating a new version.

​
Why Data Versioning Matters

Traditional version control systems like Git aren’t designed for large binary files. DVC solves this by:
  • Keeping Git lightweight: Only small metadata files are tracked in Git
  • Enabling true reproducibility: Every code version links to specific data versions
  • Supporting massive datasets: Handle gigabytes or terabytes of data efficiently
  • Providing data lineage: Track how data evolves over time

​
How Data Versioning Works

When you add a file to DVC using dvc add, several things happen under the hood:

​
1. Content Hashing

DVC computes a hash (checksum) of your file or directory. The implementation can be found in dvc/output.py: 
# From dvc/output.py:534-565
def _get_hash_meta(self):
    if self.use_cache:
        odb = self.cache
    else:
        odb = self.local_cache
    _, meta, obj = self._build(
        odb,
        self.fs_path,
        self.fs,
        self.hash_name,
        ignore=self.dvcignore,
        dry_run=not self.use_cache,
    )
    return meta, obj.hash_info
By default, DVC uses md5 hashing for local files and adapts to cloud-specific checksums (etag for S3, checksum for HDFS) for remote storage.
DVC 3.x uses md5 as the default hash algorithm. For directories, DVC creates a .dir file containing hashes of all files inside.

​
2. Cache Storage

The actual file content is moved (or linked) to the DVC cache directory at .dvc/cache/files/md5/. The cache uses a content-addressable structure: 
.dvc/cache/files/md5/
  ab/
    cd1234567890abcdef1234567890abcd  # Your actual file
The first two characters of the hash become a subdirectory for efficient file system performance.

​
3. DVC File Creation

A .dvc file is created with metadata about your tracked file: 
outs:
- md5: abcd1234567890abcdef1234567890ab
  size: 1048576
  path: data/train.csv
  cache: true
This .dvc file goes into Git, serving as a pointer to the actual data in the cache.

​
4. Workspace Linking

DVC creates a link (reflink, hardlink, symlink, or copy) from your workspace to the cached file. The implementation in dvc/output.py:752-800 handles this: 
def commit(self, filter_info=None, relink=True) -> None:
    if not self.exists:
        raise self.DoesNotExistError(self)
    
    if self.use_cache:
        # Transfer to cache
        otransfer(
            staging,
            self.cache,
            {obj.hash_info},
            shallow=False,
            hardlink=hardlink,
            callback=cb,
        )
        if relink:
            # Link from cache back to workspace
            self._checkout(
                filter_info or self.fs_path,
                self.fs,
                obj,
                self.cache,
                relink=True,
                state=self.repo.state,
                prompt=prompt.confirm,
                progress_callback=cb,
                old=obj,
            )

​
Directory Versioning

Directories are handled specially in DVC. From dvc/output.py:220-232: 
def _serialize_tree_obj_to_files(obj: Tree) -> list[dict[str, Any]]:
    key = obj.PARAM_RELPATH
    return sorted(
        (
            {
                key: posixpath.sep.join(parts),
                **_serialize_hi_to_dict(hi),
                **meta.to_dict(),
            }
            for parts, meta, hi in obj
        ),
        key=itemgetter(key),
    )
For directories:
  1. Each file inside is hashed individually
  2. A .dir file is created in cache containing JSON with all file hashes
  3. The directory hash is the hash of this .dir file
  4. The .dvc file stores just the directory hash and metadata
outs:
- md5: a1b2c3d4e5f6.dir
  size: 5242880
  nfiles: 42
  path: data/images/
The .dir file in cache contains:
[
  {"relpath": "cat.jpg", "md5": "abc123...", "size": 12345},
  {"relpath": "dog.jpg", "md5": "def456...", "size": 23456}
]

​
Version Switching

To switch between data versions: 
# Switch to a different Git branch/commit
git checkout experiment-branch

# Update workspace to match DVC files in that version
dvc checkout
The dvc checkout command reads the .dvc files in your workspace and restores files from cache.

​
Data Lineage

Track Changes

Use Git history to see how .dvc files changed
git log data/train.csv.dvc

Compare Versions

View differences between data versions
git diff HEAD~1 data/train.csv.dvc

​
File Status and Changes

DVC tracks whether files have changed from what’s recorded in .dvc files. The implementation in dvc/output.py:599-642: 
def changed_checksum(self):
    return self.hash_info != self.get_hash()

def changed_cache(self, filter_info=None):
    if not self.use_cache or not self.hash_info:
        return True
    obj = self.get_obj(filter_info=filter_info)
    if not obj:
        return True
    try:
        ocheck(self.cache, obj)
        return False
    except (FileNotFoundError, ObjectFormatError):
        return True

def status(self) -> dict[str, str]:
    if self.hash_info and self.use_cache and self.changed_cache():
        return {str(self): "not in cache"}
    return self.workspace_status()
This enables commands like:
  • dvc status - Shows which tracked files have changed
  • dvc diff - Shows differences between versions
Important: If you modify a DVC-tracked file directly, you must run dvc add again to update its version. DVC doesn’t automatically detect changes.

​
Output Types

DVC supports different output types with special handling:
TypePurposeSchema Field
Regular OutputStandard tracked filescache: true
MetricNumerical results for comparisonmetric: true
PlotData for visualizationplot: true
No-cacheTracked but not cachedcache: false
PersistNot removed on dvc repropersist: true
From dvc/output.py:281-303: 
PARAM_PATH = "path"
PARAM_CACHE = "cache"
PARAM_METRIC = "metric"
PARAM_PLOT = "plot"
PARAM_PERSIST = "persist"
PARAM_REMOTE = "remote"
PARAM_PUSH = "push"
  • dvc add - Start tracking a file or directory
  • dvc checkout - Restore tracked files from cache
  • dvc status - Check for changes in tracked files
  • dvc push - Upload tracked data to remote storage
  • dvc pull - Download tracked data from remote storage

​
Next Steps

Remote Storage

Learn how to share versioned data with your team

Pipelines

Connect versioned data with processing steps

Build docs developers (and LLMs) love

Get started for freeTalk to us