Stratified Sampling from Class-Bucketed JSON

ML-coding sub-round (~20 min slot inside a triple-task phone screen). Input is a JSON object keyed by class label; the values are example arrays. Implement a uniform-over-classes sampler — pick a class uniformly at random first, then sample one example from that class. Requirements def stratified_sample(training: dict[str, list]) -> tuple[str, object]: # training = {"cat": [...], "dog": [...], "fox": [...]} # Return (class_label, sampled_example) such that the marginal distribution # over class_label is uniform across keys (regardless of bucket sizes). The expected implementation: Sample a class label uniformly from the keys. Sample an example uniformly from that class's list. The trap is using random.choice over the flattened training set, which biases toward majority classes — that's random sampling, not stratified. The round tests whether the candidate articulates why uniform-over-classes ≠ uniform-over-examples. Notes Stratification semantics differ across teams. Asking "uniform over classes or proportional to class frequency?" is the high-signal opening. Reported as part of a tightly packed 60-minute phone screen (BQ + ML coding + system design); manage time aggressively and avoid optimizing for sub-millisecond performance. The MLE / data-engineering follow-up sometimes asks how to do this efficiently when classes live in shards / on disk — reservoir sampling per shard then a final uniform draw across shards. Preparation Verbalize the difference between stratified and proportional sampling in < 30 seconds. Implement both variants — uniform-over-classes and uniform-over-examples — so you can switch if the spec changes mid-round. Brush up on reservoir sampling for the streaming follow-up.