Product Context

AdNova is a self-serve ads platform used by mid-market and enterprise advertisers to target users for email, push, and in-app campaigns. You are designing the ML system that predicts which users should be included in a campaign audience, with feature drift treated as a first-class design concern.

Scale

Campaigns vary from broad re-engagement blasts to narrow high-value segments. User behavior, campaign mix, seasonality, privacy constraints, and upstream schema changes frequently shift feature distributions.

Task

1. Define the functional and non-functional requirements for a campaign targeting system that supports both batch audience generation and low-latency online scoring.
2. Propose an end-to-end architecture, including data ingestion, feature computation, training, model serving, and a monitoring loop for feature drift, label drift, and training-serving skew.
3. Design a multi-stage decision system where appropriate (for example: candidate retrieval/filtering → scoring/ranking → policy re-ranking or thresholding), and justify where batch vs online inference should be used.
4. Explain how you would build the feature store and feature pipelines so that the same features are used consistently in training and serving.
5. Define offline and online evaluation, including how you would detect when drift is hurting campaign performance before advertisers notice.
6. Identify major failure modes and mitigations, especially around stale features, delayed labels, schema changes, privacy deletions, and cost/latency regressions.

Constraints

• Advertisers expect fresh audiences within 15 minutes of major user behavior changes.
• Some features are only available in batch; others arrive via streaming events.
• Privacy rules require user deletion requests to propagate within 24 hours.
• Cost matters: average scoring cost must stay below $0.001 per 1,000 user-campaign evaluations.
• The system must remain available during partial feature pipeline outages, even if it must degrade gracefully.