Why models degrade in production in MLOps - Why It Works

Introduction

Machine learning models can perform well during testing but start making worse predictions when used in real life. This happens because the environment or data changes after the model is deployed. Understanding why models degrade helps keep them accurate and useful.

When your model's accuracy drops after deployment even though it worked well during training

When new types of data appear that the model has never seen before

When the real-world environment changes, like customer behavior or sensor conditions

When you want to monitor and maintain your model's performance over time

When you need to plan retraining or updating your model to keep it reliable

Commands

Starts the MLflow tracking server to monitor model performance and metrics over time in a web interface.

Terminal

mlflow ui

Expected Output

2024/06/01 12:00:00 Starting MLflow UI at http://127.0.0.1:5000

→

--port - Specify the port number for the UI server

Creates a new MLflow experiment to organize runs and track model performance data.

Terminal

mlflow experiments create --experiment-name model-monitoring

Expected Output

Created experiment with ID 1

Runs the MLflow project with specific parameters to train or retrain the model and log metrics for comparison.

Terminal

mlflow run . -P alpha=0.01 -P l1_ratio=0.5

Expected Output

2024/06/01 12:05:00 Run completed with status FINISHED

Metrics logged: accuracy=0.85

→

-P - Pass parameters to the MLflow project

Lists the metrics logged for a specific run to check model performance and detect degradation.

Terminal

mlflow metrics list --run-id 1234567890abcdef

Expected Output

accuracy: 0.85

loss: 0.35

Key Concept

If you remember nothing else, remember: models degrade because the data or environment changes after deployment, so continuous monitoring and retraining are essential.

Common Mistakes

Ignoring changes in input data distribution after deployment

The model sees data different from training, causing poor predictions

Set up monitoring to detect data drift and retrain the model regularly

Not logging model performance metrics during production

Without metrics, you cannot know when the model is degrading

Use tools like MLflow to track and compare model metrics continuously

Assuming a model trained once will work forever

Real-world conditions change, so the model needs updates

Plan for periodic retraining and validation with fresh data

Summary

Models degrade in production because the data or environment changes over time.

Use MLflow to track model performance and detect degradation early.

Regular retraining and monitoring keep models accurate and reliable.

Practice

1. Why do machine learning models often degrade when deployed in production?

easy

A. Because the model code is always incorrect

B. Because the data or environment changes over time

C. Because production servers are slower

D. Because models never work outside training

Solution

1. Step 1: Understand model dependency on data

   Models learn patterns from training data, so if data changes, predictions may worsen.

2. Step 2: Recognize environment changes

   Changes in user behavior or system environment can cause model performance to drop.

3. Final Answer:

   Because the data or environment changes over time -> Option B

4. Quick Check:

   Model degradation = data/environment change [OK]

Hint: Models degrade when input data changes from training data [OK]

Common Mistakes:

• Thinking model code is always wrong
• Blaming server speed for model errors
• Assuming models never work outside training

2. Which of the following is a correct way to monitor model degradation in production?

easy

A. Stop collecting new data after deployment

B. Ignore model outputs and trust initial accuracy

C. Only retrain the model once a year

D. Track model performance metrics regularly

Solution

1. Step 1: Identify monitoring best practice

   Regularly tracking metrics like accuracy or error helps detect degradation early.

2. Step 2: Eliminate poor practices

   Ignoring outputs or stopping data collection prevents noticing problems timely.

3. Final Answer:

   Track model performance metrics regularly -> Option D

4. Quick Check:

   Monitoring = track metrics regularly [OK]

Hint: Monitor metrics often to catch degradation early [OK]

Common Mistakes:

• Ignoring model outputs after deployment
• Waiting too long to retrain
• Stopping data collection

3. Consider this code snippet monitoring model accuracy over time:

accuracies = [0.95, 0.93, 0.88, 0.85, 0.80]
if accuracies[-1] < 0.85:
    alert = True
else:
    alert = False
print(alert)

What will be the output and what does it indicate?

medium

A. True; model accuracy dropped below threshold

B. False; model accuracy is stable

C. Error; syntax mistake in code

D. True; model accuracy improved

Solution

1. Step 1: Check last accuracy value

   The last accuracy is 0.80, which is less than 0.85 threshold.

2. Step 2: Evaluate condition and output

   Since 0.80 < 0.85, alert is set to True and printed.

3. Final Answer:

   True; model accuracy dropped below threshold -> Option A

4. Quick Check:

   Last accuracy < threshold = True alert [OK]

Hint: Check last accuracy value against threshold [OK]

Common Mistakes:

• Confusing less than with greater than
• Assuming code has syntax error
• Thinking True means improvement

4. You have this monitoring code snippet:

accuracy = 0.82
if accuracy <= 0.8:
    print("Retrain model")
else:
    print("Model OK")

But the model is degrading and you want retraining to trigger at 0.85 accuracy or below. What is the fix?

medium

A. Remove else block

B. Change print statements order

C. Change condition to accuracy <= 0.85

D. Change accuracy variable to 0.9

Solution

1. Step 1: Identify current threshold

   Current code triggers retrain only if accuracy is 0.8 or less.

2. Step 2: Adjust threshold to 0.85

   Change condition to accuracy <= 0.85 to retrain earlier.

3. Final Answer:

   Change condition to accuracy <= 0.85 -> Option C

4. Quick Check:

   Retrain threshold = 0.85 [OK]

Hint: Update condition threshold to desired retrain point [OK]

Common Mistakes:

• Changing print order doesn't affect logic
• Removing else block won't fix threshold
• Changing accuracy value ignores real data

5. A deployed model's accuracy drops because the input data distribution changed. Which combined approach best addresses this degradation?

hard

A. Monitor performance, retrain with new data, and update deployment

B. Switch to a simpler model without monitoring

C. Retrain only when users complain

D. Ignore changes and keep using the model

Solution

1. Step 1: Recognize need for monitoring

   Monitoring detects when model accuracy drops due to data changes.

2. Step 2: Retrain and update model

   Retraining with new data adapts model to current distribution; redeploy updated model.

3. Final Answer:

   Monitor performance, retrain with new data, and update deployment -> Option A

4. Quick Check:

   Monitor + retrain + update = best practice [OK]

Hint: Combine monitoring, retraining, and deployment updates [OK]

Common Mistakes:

• Ignoring data changes
• Waiting for complaints before retraining
• Dropping monitoring leads to unnoticed failures