Designing the Message Storage Schema

You need to support the query "fetch the most recent 50 messages of conversation C, and let a client page further back in time."

1. Design a wide-column schema (partition key, clustering key, columns) that serves this query efficiently.
2. Explain why a naive relational design with a single global messages table and a created_at index performs poorly at hundreds of thousands of writes/s.
3. A very active conversation (a huge public channel) can produce far more messages than a typical DM. How does this affect your partitioning choice, and what is the standard mitigation for a partition that grows too hot or too large?