Feature Skew Detection Between Training and Servingmedium

Feature Skew Detection Between Training and Serving

medium ML Systems & MLOps 15 min

NumPy allowed · plain Python earns 10 bonus points

Implement real training/serving skew detection: comparing each feature's mean between the training set and live traffic to catch a genuinely common, silent production bug.

Task

Implement detect_feature_skew(train_stats, serving_stats, threshold_pct). Both dicts map feature name -> its mean value. For each feature present in BOTH dicts, compute relative change abs(serving - train) / abs(train) * 100 (skip a feature if train's value is exactly 0, to avoid division by zero). Return a sorted list of feature names where that relative change EXCEEDS threshold_pct.

Function Signature

detect_feature_skew(train_stats: dict[str, float], serving_stats: dict[str, float], threshold_pct: float) -> list[str]

Examples

Example 1: no significant skew
Input: train_stats = {'age': 30}
serving_stats = {'age': 31}
threshold_pct = 10
Output: []

Example 2: significant skew flagged
Input: train_stats = {'age': 30}
serving_stats = {'age': 60}
threshold_pct = 10
Output: ['age']

Example 3: multiple features, some skewed
Input: train_stats = {'a': 10, 'b': 100}
serving_stats = {'a': 10.5, 'b': 200}
threshold_pct = 20
Output: ['b']

Constraints

threshold_pct >= 0

Python3Saved ✓

def detect_feature_skew(train_stats, serving_stats, threshold_pct):

# Your implementation here

pass

Case 1: no significant skew

Input: train_stats = {'age': 30}
serving_stats = {'age': 31}
threshold_pct = 10

Expected: []

Case 2: significant skew flagged

Input: train_stats = {'age': 30}
serving_stats = {'age': 60}
threshold_pct = 10

Expected: ['age']

Case 3: multiple features, some skewed

Input: train_stats = {'a': 10, 'b': 100}
serving_stats = {'a': 10.5, 'b': 200}
threshold_pct = 20

Expected: ['b']