Scaling Candidate Generation

Your platform has 100M+ videos and must return a homepage in p99 200 ms at 55k QPS peak. A junior engineer suggests: "just run the ranking model over every video for each user and take the top 30."

1. Explain quantitatively why this does not work.
2. Design the candidate generation stage that replaces it, including how it stays within latency and how you would combine multiple retrieval sources.
3. How does this stage need to change to serve a video uploaded 5 minutes ago?