It’s why you need “enough” data. Generally the distribution of teens is a power-law distribution. There’s a set of common terms and a very long tail of other terms. This will be the case across every shard if you do in fact have enough data per shard. This becomes a factor in certain distributed aggregations as well. YMMV of course.

In general you trade accuracy for speed. You tend to get a lot of speed for just a small amount of accuracy sacrifice.