The first revision
A tenant or key range suddenly receives far more writes than its neighbors. The storage fleet has spare average capacity, but one partition is saturated.
Why it gets worse
Adding a machine does not help when the partition key keeps routing the same hot key to the same place. Queue age and tail latency rise for one slice of traffic.
Repair options
Split the hot tenant, add a salt to write keys, or isolate large tenants. These options spread writes but can make reads more complicated because the result must be reassembled.
Replay
Change traffic distribution from even to skewed. Compare per shard utilization and p99 latency. A cluster average can hide the revision that matters.