What problem are we solving?
A single database eventually becomes the bottleneck. Sharding spreads the dataset and workload, but a bad partition key can still create a hot shard.
A good partition key keeps load balanced. Increasing shard count gives more capacity, but a hot key can still concentrate traffic on one partition.
Hot celebrity key
Push skew above 60%. Notice how the hottest shard becomes the real bottleneck even though total capacity increased.
How would you fix the hot key without rebuilding the entire dataset?
