Why data versioning is harder than code versioning in MLOps - The Real Reasons

The Big Idea

Discover why managing data versions feels so tricky compared to code--and how to fix it!

The Scenario

Imagine you are working on a project where you need to keep track of changes in your code and data separately. You use a simple folder to save your data files and a Git repository for your code. Every time you update your data, you manually copy new files into the folder and try to remember which version you used for each experiment.

The Problem

This manual way is slow and confusing. Data files are often large, so copying them wastes time and space. It's easy to lose track of which data version matches which code version. Mistakes happen, like using the wrong data for training, leading to wrong results and frustration.

The Solution

Data versioning tools automatically track changes in data files, just like Git does for code. They store only differences to save space and link data versions to code versions. This makes it easy to reproduce experiments and share exact data states without confusion or extra copying.

Before vs After

✗ Before

Copy data files manually to new folders for each version

✓ After

Use a data versioning tool to track and switch data versions automatically

What It Enables

It enables reliable and efficient tracking of data changes, making experiments reproducible and collaboration smooth.

Real Life Example

In a machine learning project, you can switch between different cleaned datasets easily to compare model results without mixing up files or wasting storage.

Key Takeaways

Manual data tracking is slow, error-prone, and wastes space.

Data versioning tools automate tracking and save storage by storing differences.

This leads to reproducible experiments and easier collaboration.

Practice

1.

Why is data versioning generally harder than code versioning?

easy

A. Because code does not need to be tracked for changes.

B. Because code is written in many different programming languages.

C. Because data files are usually much larger and change more frequently than code files.

D. Because data is always stored in databases, unlike code.

Solution

1. Step 1: Understand size and frequency differences

   Data files tend to be much larger and updated more often than code files, making tracking harder.

2. Step 2: Compare code and data versioning challenges

   Code changes are usually smaller and easier to manage with tools like Git, unlike large, frequently changing data.

3. Final Answer:

   Because data files are usually much larger and change more frequently than code files. -> Option C

4. Quick Check:

   Data size and change frequency = D [OK]

Hint: Remember: bigger and frequent changes make data versioning tough [OK]

Common Mistakes:

• Thinking code is harder because of multiple languages
• Assuming data is always in databases
• Believing code doesn't need versioning

2.

Which of the following is a correct statement about data versioning tools?

Choose the correct syntax to initialize a data versioning repository using dvc command line.

easy

A. git dvc init

B. dvc init

C. init dvc

D. dvc start

Solution

1. Step 1: Recall dvc initialization command

   The correct command to start a data versioning repo with DVC is dvc init.

2. Step 2: Eliminate incorrect syntax

   Commands like git dvc init, init dvc, and dvc start are invalid or do not exist.

3. Final Answer:

   dvc init -> Option B

4. Quick Check:

   DVC init command = A [OK]

Hint: Use simple dvc init to start data versioning [OK]

Common Mistakes:

• Adding git before dvc command
• Reversing command words
• Using non-existent commands like dvc start

3.

Consider this simplified code snippet using DVC commands:

dvc add data.csv
git add data.csv.dvc
git commit -m "Add data version"
dvc push

What is the main purpose of the dvc add data.csv command here?

medium

A. It tracks the data file data.csv in DVC and creates a pointer file.

B. It uploads data.csv to the remote storage immediately.

C. It deletes the local data.csv file after tracking.

D. It commits the data file directly to Git.

Solution

1. Step 1: Understand dvc add function

   The dvc add command tracks the data file and creates a small pointer file (like data.csv.dvc) to represent it.

2. Step 2: Clarify what dvc add does not do

   It does not upload data to remote storage (that's dvc push), nor delete the local file or commit to Git directly.

3. Final Answer:

   It tracks the data file data.csv in DVC and creates a pointer file. -> Option A

4. Quick Check:

   dvc add tracks data locally = A [OK]

Hint: dvc add tracks data locally, dvc push uploads [OK]

Common Mistakes:

• Confusing dvc add with dvc push
• Thinking it deletes local data
• Assuming it commits data to Git

4.

Given this error when trying to push data versions:

Error: failed to push data to remote storage: permission denied

What is the most likely cause and fix?

medium

A. Git repository is not initialized; fix by running git init.

B. The local data file is missing; fix by adding the file again.

C. DVC is not installed; fix by reinstalling DVC.

D. The remote storage credentials are missing or incorrect; fix by configuring access keys.

Solution

1. Step 1: Analyze the permission denied error

   This error usually means the remote storage (like S3, GCS) credentials are missing or wrong.

2. Step 2: Identify the correct fix

   Configuring or updating access keys or permissions for the remote storage resolves this issue.

3. Final Answer:

   The remote storage credentials are missing or incorrect; fix by configuring access keys. -> Option D

4. Quick Check:

   Permission denied = fix credentials [OK]

Hint: Permission denied usually means remote access keys need fixing [OK]

Common Mistakes:

• Assuming local file is missing
• Thinking Git init fixes remote errors
• Believing DVC installation causes permission errors

5.

In a team working on machine learning, why is good data versioning critical compared to just versioning code?

Choose the best explanation.

hard

A. Because data changes impact model training results, and tracking data versions ensures reproducibility and reliable improvements.

B. Because code versioning tools cannot handle any files larger than 1MB.

C. Because data versioning replaces the need for code versioning entirely.

D. Because data versioning automatically fixes bugs in the code.

Solution

1. Step 1: Understand the role of data in ML models

   Data directly affects how models learn and perform, so knowing exactly which data version was used is essential.

2. Step 2: Explain why data versioning matters for teams

   Good data versioning helps teams reproduce results and improve models reliably by tracking data changes alongside code.

3. Final Answer:

   Because data changes impact model training results, and tracking data versions ensures reproducibility and reliable improvements. -> Option A

4. Quick Check:

   Data affects models; versioning ensures reproducibility = B [OK]

Hint: Data versioning ensures model results can be repeated and improved [OK]

Common Mistakes:

• Thinking data versioning replaces code versioning
• Believing code tools can't handle files over 1MB
• Assuming data versioning fixes code bugs