What is stored on each of the index servers when you shard based on tweet object? Is it still a big hash table with key = word, value = set of tweet ids containing that word? Or something else? How does this differ from sharding by word?