Design real-time ad impression and spend tracking

Scenario

You are building a backend that helps an ad server decide whether a candidate ad can be delivered (e.g., not over budget / within pacing). The system must track:

• Impressions (ad view counts)
• Spend (money spent attributed to those impressions/clicks/etc.)

Requirements

1. Event collection : For each ad view, the system should record an event (at least: ad_id , campaign_id , user_id optional, timestamp , and fields needed to compute spend).
2. Near real-time decisioning : When a user requests an ad, the ad server evaluates up to 10,000 candidate ads and needs current counters (impressions/spend and remaining budget) to filter/score them.
3. Latency : The ad server request path should be low-latency (assume single-digit to tens of milliseconds for fetching counters).
4. Scale : High QPS for both event ingestion and read traffic from ad servers.
5. Correctness : Counters should be accurate enough for budget enforcement; clarify acceptable freshness (e.g., seconds) and consistency model.

Deliverables

Design the end-to-end system:

• APIs (event ingestion + read APIs for ad server)
• Data model
• Storage/streaming components
• How you serve counters for 10k ads per request efficiently
• Handling duplicates, retries, and late events
• Monitoring and failure recovery