Trending-detection with acceleration

Design a trending-detection system that surfaces topics whose interest is accelerating, not just topics that are popular. A perennially-high term (e.g. 'weather') should NOT trend; a term going from 200 to 20,000 mentions in 10 minutes SHOULD. You ingest 800k events/sec across millions of distinct terms, want updates every ~30s, and must resist a coordinated bot push of a hashtag. Define the trending score, the windowing, and the abuse defenses.