Why models degrade in production in MLOps - Test Your Understanding

Practice - 5 Tasks

1fill in blank

easy

Complete the code to identify the main cause of model degradation in production.

MLOps

if data_distribution_changes == [1]:
    alert('Model performance may degrade')

Drag options to blanks, or click blank then click option'

ATrue

BFalse

CNone

D0

Attempts:

3 left

💡 Hint

Common Mistakes

Using False instead of True

Using None which is not a boolean

Model degradation often happens when the data distribution changes in production, so the condition should be True.

2fill in blank

medium

Complete the code to log model accuracy drop in production.

MLOps

if current_accuracy [1] baseline_accuracy:
    log('Accuracy dropped')

Drag options to blanks, or click blank then click option'

A>=

B!=

C<

D==

Attempts:

3 left

💡 Hint

Common Mistakes

Using >= which means accuracy improved or stayed same

Using == which only detects equality

We want to log when current accuracy is less than baseline, indicating degradation.

3fill in blank

hard

Fix the error in the code that checks for data drift.

MLOps

if detect_data_drift() [1]:
    retrain_model()

Drag options to blanks, or click blank then click option'

Ais False

B!=

C== None

Dis True

Attempts:

3 left

💡 Hint

Common Mistakes

Using 'is False' which means no drift

Using '== None' which is incorrect for booleans

The function detect_data_drift() returns True if drift is detected, so we check if it is True.

4fill in blank

hard

Fill both blanks to create a dictionary comprehension that tracks feature importance changes above a threshold.

MLOps

importance_changes = {feature: importance[1]old_importance for feature, importance in new_importances.items() if importance[1]old_importance [2] threshold}

Drag options to blanks, or click blank then click option'

A-

B>

C<

D+

Attempts:

3 left

💡 Hint

Common Mistakes

Using '+' instead of '-' for difference

Using '<' which filters smaller changes

The difference is calculated by subtracting old importance from new, and we filter features where this difference is greater than the threshold.

5fill in blank

hard

Fill all three blanks to create a filtered dictionary of features with significant importance drop.

MLOps

significant_drops = { [1]: [2] for [3], importance in old_importances.items() if importance - new_importances.get([1], 0) > 0.1 }

Drag options to blanks, or click blank then click option'

Afeature

Bimportance

Dfeat

Attempts:

3 left

💡 Hint

Common Mistakes

Using different variable names causing errors

Mixing up keys and values

The dictionary keys and values use 'feature' and 'importance' variables consistently to track drops.

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