You balanced data distribution and request distribution to avoid hot partitions. What specific technique or strategy did you use, and how did you monitor CPU, memory, and request rates to ensure optimal performance?

We use a hash‑based partitioning scheme combined with dynamic rebalancing. Each partition is assigned a unique key range, and we monitor CPU, memory, and request rates at the partition level using Prometheus metrics. When a partition’s CPU or request rate exceeds a threshold, we trigger a split: the partition is divided into two, and the new key range is assigned to a less‑busy node. Conversely, if a partition is underutilized, we merge it with a neighbor. We also use a global load balancer that routes incoming requests to the least‑loaded node for a given key. Alerts are configured to notify the ops team when memory usage spikes or latency exceeds SLA. By continuously monitoring these metrics and automatically rebalancing partitions, we keep each node’s load within acceptable bounds and prevent any single partition from becoming a performance bottleneck.