Tracking datasets with DVC in MLOps - Step-by-Step Execution

Process Flow - Tracking datasets with DVC

Initialize DVC in project

↓

Add dataset to DVC tracking

↓

DVC creates .dvc file and stores data hash

↓

Push dataset to remote storage (optional)

↓

Modify dataset locally

↓

DVC detects changes and updates tracking

↓

Pull dataset from remote when needed

This flow shows how DVC tracks datasets by initializing, adding data, storing metadata, pushing to remote, detecting changes, and pulling data.

Execution Sample

MLOps

dvc init

dvc add data/dataset.csv

git add data/dataset.csv.dvc .gitignore

git commit -m "Track dataset with DVC"

This code initializes DVC, adds a dataset file to DVC tracking, stages the DVC metadata files, and commits them to Git.

Process Table

Step	Command	Action	Result
1	dvc init	Initialize DVC in project folder	Creates .dvc folder and config files
2	dvc add data/dataset.csv	Add dataset file to DVC tracking	Generates data/dataset.csv.dvc file with hash info
3	git add data/dataset.csv.dvc .gitignore	Stage DVC metadata files for Git	Files ready to commit
4	git commit -m "Track dataset with DVC"	Commit changes to Git	Dataset tracking metadata saved in Git
5	Modify data/dataset.csv	Change dataset content	Local file changed, DVC not updated yet
6	dvc status	Check dataset status	Shows dataset.csv is modified and needs update
7	dvc add data/dataset.csv	Update DVC tracking for changed dataset	Updates .dvc file with new hash
8	git add data/dataset.csv.dvc	Stage updated metadata	Ready to commit updated dataset info
9	git commit -m "Update dataset version"	Commit updated tracking	New dataset version tracked in Git
10	dvc push	Push dataset to remote storage	Dataset files uploaded to remote storage
11	dvc pull	Retrieve dataset from remote	Dataset files downloaded locally if missing

💡 Process ends after dataset is tracked, updated, and optionally pushed or pulled from remote storage.

Status Tracker

Variable	Start	After Step 2	After Step 7	After Step 10
data/dataset.csv.dvc	Not present	Created with initial hash	Updated with new hash after dataset change	Same updated file pushed to remote
data/dataset.csv	Original file	Original file tracked	Modified file locally	File synced with remote after pull

Key Moments - 3 Insights

Why do we need to run 'dvc add' again after modifying the dataset?

Because DVC tracks dataset versions by file hash, after modification the hash changes. Running 'dvc add' updates the .dvc file with the new hash to track the latest version, as shown in steps 5 to 7 in the execution table.

What is the role of the .dvc file created when adding a dataset?

The .dvc file stores metadata including the hash of the dataset file. It tells DVC which version of the data is tracked. This is why we commit the .dvc file to Git (step 4), so dataset versions are linked with code versions.

Why do we push datasets separately with 'dvc push' after committing changes?

Because datasets can be large, DVC stores actual data in remote storage separately from Git. 'dvc push' uploads the data files to remote storage, while Git only tracks metadata. This separation is shown in step 10.

Visual Quiz - 3 Questions

Look at the execution table, what happens at step 2?

A DVC initializes the project

B Dataset file is added to DVC tracking and .dvc file is created

C Dataset is pushed to remote storage

D Git commits the dataset metadata

💡 Hint

Refer to row 2 in the execution_table where 'dvc add' command is run.

At which step does DVC detect that the dataset file has changed?

A Step 6

B Step 5

C Step 4

D Step 9

💡 Hint

Check the 'dvc status' command in step 6 that shows dataset modification.

If you skip 'git add' after 'dvc add', what will happen?

A Dataset file will not be tracked by DVC

B DVC will fail to add the dataset

C Dataset changes won't be saved in Git history

D Remote storage will not accept the dataset

💡 Hint

Look at steps 3 and 4 where git add and commit save the .dvc metadata.

Concept Snapshot

Tracking datasets with DVC:
- Run 'dvc init' once per project
- Use 'dvc add <file>' to track datasets
- DVC creates a .dvc file with dataset hash
- Commit .dvc files to Git to version data metadata
- Use 'dvc push' to upload data to remote storage
- Use 'dvc pull' to download data when needed
- Re-run 'dvc add' after dataset changes to update tracking

Full Transcript

This lesson shows how to track datasets using DVC step-by-step. First, initialize DVC in your project folder with 'dvc init'. Then add your dataset file using 'dvc add', which creates a .dvc file storing the dataset's hash. Stage and commit this .dvc file with Git to version your data metadata alongside code. When you modify the dataset, run 'dvc status' to check changes and 'dvc add' again to update tracking. Commit the updated .dvc file to Git. To share or backup data, use 'dvc push' to upload dataset files to remote storage. Others can retrieve data with 'dvc pull'. This process helps keep data versions organized and linked to your code changes.