Distributed time-series query engine

Design the distributed query engine for a time-series database holding 5 years of data across hundreds of storage nodes. Typical queries: 'p95 latency per service over the last 7 days, 1-minute resolution' and 'top 20 hosts by CPU this month'. A single such query can touch billions of points across many shards. You need interactive (<2s) responses for recent-data queries and bounded resource use so one heavy query doesn't take down the cluster. Design query planning, distribution, aggregation push-down, and cost control.