Data drift detection in MLOps - Step-by-Step Execution

Process Flow - Data drift detection

Collect baseline data

↓

Train model on baseline

↓

Collect new incoming data

↓

Compare new data to baseline

↓

Calculate drift metrics

↓

Is drift above threshold?

No→Continue monitoring

Yes↓

Trigger alert or retrain model

This flow shows how data drift detection compares new data to baseline data, calculates drift, and triggers alerts if drift is significant.

Execution Sample

MLOps

baseline = load_data('baseline.csv')
new_data = load_data('new.csv')
drift_score = calculate_drift(baseline, new_data)
if drift_score > 0.3:
    alert('Data drift detected')

This code loads baseline and new data, calculates a drift score, and alerts if the score exceeds 0.3.

Process Table

Step	Action	Data Compared	Drift Score	Decision	Output
1	Load baseline data	baseline.csv	-	N/A	Baseline data loaded
2	Load new data	new.csv	-	N/A	New data loaded
3	Calculate drift score	baseline vs new	0.25	0.25 <= 0.3	No alert
4	Calculate drift score	baseline vs new	0.35	0.35 > 0.3	Alert triggered: Data drift detected

💡 Execution stops after alert triggered or no alert if drift score is below threshold

Status Tracker

Variable	Start	After Step 1	After Step 2	After Step 3	After Step 4
baseline	None	Loaded baseline.csv data	Loaded baseline.csv data	Loaded baseline.csv data	Loaded baseline.csv data
new_data	None	None	Loaded new.csv data	Loaded new.csv data	Loaded new.csv data
drift_score	None	None	None	0.25 or 0.35 depending on data	0.25 or 0.35 depending on data

Key Moments - 2 Insights

Why do we compare new data to baseline data instead of just looking at new data alone?

Because drift means the new data distribution changes compared to the baseline. The execution_table shows step 3 where the comparison happens to calculate drift_score.

What does the drift_score represent and why is there a threshold?

Drift_score quantifies how much new data differs from baseline. The threshold (0.3) in step 4 decides if the difference is significant enough to alert.

Visual Quiz - 3 Questions

Look at the execution_table, what is the drift_score at step 3?

A 0.3

B 0.25

C 0.35

D No score calculated

💡 Hint

Check the 'Drift Score' column in row with Step 3 in execution_table

At which step does the system decide to trigger an alert?

A Step 2

B Step 3

C Step 4

D No alert triggered

💡 Hint

Look at the 'Decision' and 'Output' columns in execution_table rows

If the threshold was changed to 0.2, how would the alert behavior change?

A Alert triggers at drift_score 0.25

B Alert triggers only at drift_score 0.35

C No alerts would trigger

D Alerts trigger at any drift_score

💡 Hint

Compare threshold logic in code and execution_table decisions

Concept Snapshot

Data drift detection compares new data to baseline data.
Calculate a drift score to measure difference.
If drift score > threshold, trigger alert or retrain.
Common threshold example: 0.3.
Helps keep ML models accurate over time.

Full Transcript

Data drift detection is a process where we first collect baseline data and train a model on it. Then, as new data comes in, we compare it to the baseline data to see if the data distribution has changed. We calculate a drift score to quantify this difference. If the drift score is above a set threshold, for example 0.3, we trigger an alert to notify that data drift has occurred. This helps us know when to retrain or adjust our machine learning models to keep them accurate. The execution steps include loading baseline and new data, calculating the drift score, and deciding whether to alert based on the threshold.