Comment-Likelihood Model: Feature and Model Serving

Reported in interviews at Reddit

Problem statement

How would you predict whether someone will comment on a post they view? A busy discussion feed needs fast probabilities for every candidate post, blended with other ranking signals. Comments are rare, popular-post counters change quickly, and the current ordering shapes what people see.

Design the scoring service, test exposure effects and correct sampled probabilities.

Show hint

What event and exposure condition must be fixed before a probability can be compared across posts?

Relevant roles: ML Engineer, MLOps Engineer and Applied Scientist.

Related free problems