Design comment-likelihood prediction platform

Scenario

You’re building an ML platform component that serves a model to predict the likelihood that a user will comment on a given post.

The interviewer says you can treat the model as a black box (you don’t need to pick a specific architecture); focus on ML infrastructure: feature pipelines, feature store, training/serving, inference, and production concerns.

Goals

Design an end-to-end system that supports:

• Offline training data generation and model training
• Online inference (real-time scoring) for product surfaces (e.g., feed ranking, notification candidate scoring)
• A feature store strategy (offline + online)
• Monitoring, logging, and iteration

Requirements & constraints (you may make reasonable assumptions)

• High QPS online scoring (potentially tens of thousands/sec)
• P95 latency budget for scoring: e.g., 50–150 ms end-to-end (state your assumption)
• Freshness: some features need near-real-time updates (seconds to minutes)
• Avoid training/serving feature skew
• Handle cold start (new users/posts)
• Support A/B testing and safe rollout

Deliverables

Explain:

1. Data sources and event logging
2. Feature engineering and feature store design
3. Training pipeline and dataset versioning
4. Online inference architecture (batch vs real-time, caching, fallbacks)
5. Monitoring (data + model), retraining triggers, and reliability considerations