Preparing your design workspace

Loading the problem and your latest saved attempt…

Design a Comment-Likelihood Prediction Platform

Train and serve versioned probabilities that a viewer will comment on a candidate post, with point-in-time features and delayed-label evaluation.

Problem statement

For this exercise, assume a platform that predicts the probability that an authenticated viewer will create a comment on a candidate post within 24 hours of an eligible impression.

A ranking consumer submits bounded viewer-post pairs and receives calibrated probabilities plus model, feature, and score timestamps; the score is advisory and the consumer owns final ranking.

Assume online features use point-in-time definitions and freshness metadata, while training examples join eligible impressions to deduplicated comments after the 24-hour label window.

The platform prevents future leakage, preserves deletion and consent policy, compares candidate models on calibrated aggregate and slice metrics, and rolls out through deterministic cohorts with fallback.

Design event collection, label construction, offline and online feature serving, training datasets, model registry, low-latency inference, caching, monitoring, rollout, privacy, and recovery. Designing the post recommender, content policy, or model algorithm is outside scope.

What to build

  1. Construct one point-in-time 24-hour comment label per eligible viewer-post impression.
  2. Serve versioned online features with freshness and offline parity evidence.
  3. Return calibrated version-attributed comment-likelihood scores for bounded viewer-post batches.
  4. Compare candidate and baseline models, monitor delayed outcomes, and activate or roll back one reviewed release.

At scale

  • For this exercise, assume 200,000 viewer-post scores per second in batches of at most 100 pairs.
  • For this exercise, target 75 milliseconds p99 for a healthy scoring batch and make feature values older than five minutes visibly stale or unavailable.
  • For this exercise, publish one daily training snapshot from 30 days of eligible impressions and retain versioned aggregate evaluation metadata for one year.
  • For this exercise, a process or zone failure may degrade to a pinned baseline score, but must not mix feature or model versions silently or duplicate labels.