{
  "lesson": "21-feature-stores-and-training-serving-skew",
  "title": "Feature Stores and Training-Serving Skew",
  "questions": [
    {
      "stage": "pre",
      "question": "A team joins churn labels to a user feature table on user_id only. The table holds each user's newest values. What goes wrong?",
      "options": [
        "The join drops every user who has more than one label in the set",
        "Training rows get feature values from after the label time",
        "Nothing, because the newest values are the most accurate ones",
        "The model trains slower because the table has too many columns"
      ],
      "correct": 1,
      "explanation": "Newest values come from after most label times. For churn, they already show the quiet period that the label measures, so the input contains the answer."
    },
    {
      "stage": "check",
      "question": "A label is at day 20. Feature rows exist at days 6, 17, and 24, and the TTL is 7 days. Which row does a point-in-time join use?",
      "options": [
        "No row, so the default value is used",
        "The day 6 row",
        "The day 17 row",
        "The day 24 row"
      ],
      "correct": 2,
      "explanation": "The join keeps rows at or before day 20, takes the latest one, and checks its age. Day 17 is 3 days old, inside the 7-day TTL."
    },
    {
      "stage": "check",
      "question": "Training fills a missing purchase history with 365 days. The serving service fills it with -1. What kind of skew is this?",
      "options": [
        "Different defaults between the two paths",
        "Stale features from a failed materialization job",
        "Concept drift in the relation between input and label",
        "Time travel from a naive join on the user key"
      ],
      "correct": 0,
      "explanation": "The two paths disagree only when a value is missing. New users then get an input that the model never saw in training."
    },
    {
      "stage": "check",
      "question": "Why does a feature registry create a new version when a definition changes, instead of editing the old one?",
      "options": [
        "Editing a definition deletes the backfilled history in the offline store",
        "Models trained on the old definition still need the old values",
        "The online store can hold only one feature per entity at a time",
        "Versions make the point-in-time join run faster on large tables"
      ],
      "correct": 1,
      "explanation": "A deployed model expects the feature exactly as it was during training. A silent edit changes its inputs without a retrain."
    },
    {
      "stage": "post",
      "question": "Logged serving values match the training join exactly, but values computed fresh on request differ on 83 percent of requests. What does this show?",
      "options": [
        "The online store is broken, because fresh values are always the correct values",
        "The point-in-time join leaks the future, because snapshots are newer than requests",
        "The model trained on snapshot values, so a fresh compute path would add skew",
        "The registry has two versions of each feature, and serving reads the wrong one"
      ],
      "correct": 2,
      "explanation": "Freshness is part of the feature. A path that is fresher than the training data gives the model inputs from a different distribution."
    },
    {
      "stage": "post",
      "question": "A backfill that runs today writes rows for last year's dates, using data that was corrected last month. How does a strict point-in-time join avoid the leak?",
      "options": [
        "It joins on the entity key only and takes the newest row for each label",
        "It also requires the created timestamp to be at or before the label time",
        "It shortens the TTL so that last year's rows expire before training starts",
        "It runs the backfill again with the online store as its data source"
      ],
      "correct": 1,
      "explanation": "The event timestamp says when a value was true. The created timestamp says when the system knew it. A strict join checks both."
    }
  ]
}