Design distributed word count without MapReduce

Design a distributed system to compute word frequencies over terabytes of text without using MapReduce. Specify how you will ingest data (e.g., log or stream service), partition tokens across shards (e.g., consistent hashing with salting to mitigate hot keys), aggregate partial counts, and produce global results and top‑K across shards. Describe fault tolerance, idempotency, exactly‑once vs. at‑least‑once semantics, backpressure, recovery from node failures, and how results are stored and served.