Why data versioning is harder than code versioning in MLOps - Why It Works

Introduction

Tracking changes in data is more difficult than tracking changes in code because data files are often large, binary, and frequently updated. Unlike code, data can be messy, have many versions, and require special tools to manage efficiently.

When you want to keep track of different versions of datasets used in machine learning experiments

When you need to reproduce a model training exactly with the same data snapshot

When multiple team members update or add data and you want to avoid conflicts or data loss

When you want to audit or compare changes in data over time to understand model performance shifts

When you want to store large datasets efficiently without duplicating entire files for every change

Commands

Initialize a Git repository to track code changes. This works well for small text files like code but not for large data files.

Terminal

git init

Expected Output

Initialized empty Git repository in /home/user/project/.git/

Add a data file to Git tracking. This is possible but inefficient for large or frequently changing data files.

Terminal

git add data.csv

Expected Output

No output (command runs silently)

Commit the data file to the Git repository. Git stores the entire file each time it changes, which can quickly use a lot of space.

Terminal

git commit -m "Add initial data file"

Expected Output

[master (root-commit) abc1234] Add initial data file

 1 file changed, 0 insertions(+), 0 deletions(-)

 create mode 100644 data.csv

Initialize DVC (Data Version Control) in the project. DVC is designed to handle large data files efficiently by storing metadata and linking to external storage.

Terminal

dvc init

Expected Output

Initialized DVC repository.

You can now track data files with 'dvc add <filename>'.

Tell DVC to track the data file. DVC creates a small metafile to track the data version without storing the whole file in Git.

Terminal

dvc add data.csv

Expected Output

Adding 'data.csv' to DVC tracking.

Computing checksum...

Saving information to 'data.csv.dvc'.

Add the DVC metafile and updated .gitignore to Git. This keeps Git tracking small and efficient.

Terminal

git add data.csv.dvc .gitignore

Expected Output

No output (command runs silently)

Commit the DVC metafile to Git. The actual data is stored separately, making versioning large files practical.

Terminal

git commit -m "Track data.csv with DVC"

Expected Output

[master abc5678] Track data.csv with DVC

 2 files changed, 10 insertions(+), 2 deletions(-)

 create mode 100644 data.csv.dvc

Key Concept

If you remember nothing else from this pattern, remember: data files are large and change differently than code, so special tools like DVC are needed to version them efficiently.

Common Mistakes

Trying to version large data files directly with Git

Git stores full copies of files on each change, causing slow performance and huge repositories.

Use data versioning tools like DVC that track data changes efficiently without storing full copies in Git.

Ignoring .gitignore updates when adding data files

Git may try to track large data files directly, causing repository bloat and slow operations.

Update .gitignore to exclude large data files and track only metadata files with Git.

Summary

Git works well for code but is inefficient for large or frequently changing data files.

Data versioning tools like DVC track data changes by storing metadata and linking to external storage.

Combining Git for code and DVC for data keeps repositories small and reproducible.