Delayed Labels in a Streaming Trainer

Your pCTR model is fine-tuned continuously from a stream of impressions. Impressions are logged instantly; clicks arrive with a median delay of 40 seconds and a 99th percentile of 25 minutes. Conversions (for the pCVR head) arrive with a median delay of 6 hours and a 7-day attribution window.

1. Describe the naive streaming approach and the bias it introduces for each head.
2. Propose a label-handling scheme for the click head and a different one for the conversion head, and justify the difference.
3. How would you detect in production that delayed labels are biasing the model?