Design streaming mention analytics with search and alerts

## Scenario You ingest a real-time external stream of social-media posts and news articles. Each item contains raw text and metadata (timestamp, source, author/site, etc.). The product tracks companies/stocks ("entities") and shows: 1. **Mention analytics**: how many times each entity was mentioned over time (similar to impressions/mentions). 2. **Charts by time window**: users can choose time spans from **30 minutes to multiple days** (tumbling or sliding windows are both acceptable). 3. **Latency**: charts may be delayed by **10–30 minutes**, but data must be aggregated before display. 4. **Subscriptions & notifications**: users can follow a set of entities, filter analytics to followed entities, and configure alerts (e.g., spike in mentions). 5. **Search**: - Users can search across **hundreds of thousands of entities** (by company/stock name). - Users can also search for the **underlying documents** (posts/articles) that mention entities. - Search supports **any number of keywords** and filtering (e.g., entity, time range, source). - Search load can be very high (e.g., **~100k RPS**). 6. **Spiky traffic**: must handle extreme bursts (breaking news, meme-stock events). 7. **Storage choices**: decide how to store both **processed/aggregated** data and **raw documents**. ## Task Design a high-level system (APIs, data flow, storage, and scaling strategy) that satisfies the above requirements. Clearly explain: - How raw streaming data is ingested, processed, and aggregated. - How time-windowed analytics are computed and served. - How document/entity search works at high QPS. - How subscriptions and alerting are implemented. - How the system remains reliable and cost-effective under spikes. State assumptions and key trade-offs (e.g., consistency, latency, storage format, retention).