Data validation in CI pipeline in MLOps - Commands & Configuration

Introduction

Data validation in a CI pipeline helps catch errors in data before it is used in machine learning models. It ensures the data meets quality standards automatically every time new data or code is added.

When you want to check if new data files have missing or unexpected values before training a model

When you want to automatically stop a pipeline if data quality is poor

When you want to track data quality metrics over time to detect data drift

When you want to enforce data schema rules in your automated tests

When you want to integrate data checks as part of your code review process

Commands

Install Great Expectations, a popular Python library for data validation used in CI pipelines.

Terminal

pip install great_expectations

Expected Output

Collecting great_expectations

  Downloading great_expectations-0.16.18-py3-none-any.whl (1.2 MB)

Installing collected packages: great_expectations

Successfully installed great_expectations-0.16.18

Initialize a Great Expectations project in the current directory to set up configuration and folders.

Terminal

great_expectations init

Expected Output

Great Expectations has been successfully initialized!



Your new Great Expectations project is ready to use.

Run a Python script that loads data and validates it using Great Expectations in the CI pipeline.

Terminal

python validate_data.py

Expected Output

Validation succeeded: All data checks passed.

Key Concept

If you remember nothing else, remember: automate data quality checks in your CI pipeline to catch data issues early and prevent bad data from breaking your ML models.

Code Example

MLOps

import great_expectations as ge

def validate_data():
    # Load data as a Great Expectations dataset
    df = ge.read_csv('data/sample_data.csv')

    # Expect column 'age' to have no nulls
    result1 = df.expect_column_values_to_not_be_null('age')

    # Expect column 'salary' to be greater than 0
    result2 = df.expect_column_values_to_be_between('salary', min_value=1)

    # Check if all expectations passed
    if result1.success and result2.success:
        print('Validation succeeded: All data checks passed.')
    else:
        print('Validation failed: Data checks did not pass.')

if __name__ == '__main__':
    validate_data()

Common Mistakes

Not installing the data validation library before running validation scripts

The validation commands will fail because the required tools are missing.

Always install dependencies like Great Expectations before running validation commands.

Running validation without initializing the project configuration

Validation will fail because the project lacks necessary config files and folders.

Run 'great_expectations init' once to set up the project before validating data.

Ignoring validation failures in the CI pipeline

Bad data can silently pass through and cause model errors later.

Fail the CI build if data validation reports errors to enforce data quality.

Summary

Install Great Expectations to add data validation capabilities.

Initialize the Great Expectations project to create config files.

Run validation scripts in the CI pipeline to automatically check data quality.

Fail the pipeline if data validation fails to prevent bad data usage.

Practice

1. What is the main purpose of adding data validation in a CI pipeline for machine learning projects?

easy

A. To speed up the model training process

B. To catch data problems early before training models

C. To reduce the size of the dataset

D. To automatically deploy models to production

Solution

1. Step 1: Understand the role of CI pipelines

   CI pipelines automate checks and tests to ensure quality before further steps.

2. Step 2: Identify the purpose of data validation

   Data validation ensures data quality and format correctness to avoid errors in training.

3. Final Answer:

   To catch data problems early before training models -> Option B

4. Quick Check:

   Data validation = catch problems early [OK]

Hint: Data validation stops bad data early in pipeline [OK]

Common Mistakes:

• Thinking validation speeds training
• Confusing validation with deployment
• Assuming validation reduces data size

2. Which of the following is the correct way to fail a CI pipeline step if a data validation script returns a non-zero exit code in a bash script?

easy

A. python validate_data.py || exit 1

B. python validate_data.py && exit 1

C. python validate_data.py; exit 0

D. python validate_data.py | exit 1

Solution

1. Step 1: Understand bash exit codes and operators

   The '||' operator runs the command after it if the first command fails (non-zero exit).

2. Step 2: Apply to data validation script

   If 'validate_data.py' fails, 'exit 1' stops the pipeline with error.

3. Final Answer:

   python validate_data.py || exit 1 -> Option A

4. Quick Check:

   Fail on error = '|| exit 1' [OK]

Hint: Use '|| exit 1' to fail on script error [OK]

Common Mistakes:

• Using '&&' instead of '||' to fail
• Using pipe '|' incorrectly
• Exiting with 0 always

3. Given this Python snippet in a CI pipeline step:

import sys

def validate(data):
    if not data or len(data) < 5:
        return False
    return True

if __name__ == '__main__':
    data = sys.argv[1] if len(sys.argv) > 1 else ''
    if validate(data):
        print('Validation passed')
        sys.exit(0)
    else:
        print('Validation failed')
        sys.exit(1)

What will be the output and exit code if the pipeline runs python validate.py "abc"?

medium

A. Validation failed and exit code 0

B. Validation passed and exit code 0

C. Validation passed and exit code 1

D. Validation failed and exit code 1

Solution

1. Step 1: Check input data length

   Input is 'abc' which length is 3, less than 5, so validate returns False.

2. Step 2: Determine output and exit code

   Since validate returns False, it prints 'Validation failed' and exits with code 1.

3. Final Answer:

   Validation failed and exit code 1 -> Option D

4. Quick Check:

   Short data fails validation = A [OK]

Hint: Check input length to predict validation result [OK]

Common Mistakes:

• Assuming any input passes
• Confusing exit codes 0 and 1
• Ignoring input length check

4. You have this YAML snippet in a CI pipeline to run data validation:

steps:
  - name: Validate Data
    run: |
      python validate.py data.csv
      echo "Data validation complete"

The pipeline does not fail even when validation.py returns exit code 1. What is the likely problem?

medium

A. The shell does not stop on errors by default; need 'set -e'

B. The 'if' condition is incorrect and never triggers

C. The 'exit 1' is inside the if but the script continues after

D. The validate.py script always returns 0

Solution

1. Step 1: Understand shell error handling

   By default, shell scripts continue even if a command fails unless 'set -e' is used.

2. Step 2: Apply to CI step

   Without 'set -e', the script continues after python fails, runs the echo which succeeds, so step exit code is 0.

3. Final Answer:

   The shell does not stop on errors by default; need 'set -e' -> Option A

4. Quick Check:

   Use 'set -e' to fail pipeline on errors [OK]

Hint: Add 'set -e' to stop on errors in shell scripts [OK]

Common Mistakes:

• Assuming exit 1 always stops pipeline
• Misreading if condition syntax
• Ignoring shell default behavior

5. You want to add a data validation step in your CI pipeline that checks if a CSV file has no missing values and all numeric columns are within a specific range. Which approach best fits this requirement?

hard

A. Use a shell script with grep to search for empty fields and numeric ranges

B. Manually inspect the CSV file before running the pipeline

C. Write a Python script using pandas to check missing values and ranges, then fail with exit code 1 if invalid

D. Skip validation and rely on model training to catch errors

Solution

1. Step 1: Identify tools for data validation

   Pandas in Python is ideal for checking missing values and numeric ranges programmatically.

2. Step 2: Implement validation and fail pipeline

   Script should exit with code 1 if validation fails to stop the pipeline safely.

3. Final Answer:

   Write a Python script using pandas to check missing values and ranges, then fail with exit code 1 if invalid -> Option C

4. Quick Check:

   Use pandas script + exit 1 for robust validation [OK]

Hint: Use pandas for detailed CSV validation and fail on error [OK]

Common Mistakes:

• Using grep which can't handle numeric ranges well
• Relying on manual checks
• Skipping validation entirely