Ah interesting, it looks like you are right. If the any partition grows to > 10GB on the main table or a GSI, that partition splits into sub-partitions using the sort key as part of the hashing function. If there is no sort key, the partitioning scheme equally distributes items across partitions, so all partitions sub-divide at the same time.
It's pretty opaque. My impression is that each shard (storage backend quorum) gets a roughly equal share of capacity you pay for, and items in the same partition tend to live on the same shard to keep range queries small (and local indices require one-shard partitions). They've made improvements in loaning cold shards' unused capacity to hot shards, but they still recommend avoiding hot partitions and keeping load roughly even.
