Currently we create update operations using $set to perform timeseries inserts into existing buckets. However, because $set inserts new fields into a sorted position for idempotency reasons, this means that the bucket orders elements in order like:
0, 1, 10, 11, 2, 3, 4, 5, 6, 7, 8, 9
This is a problem for unwinding and will also make it impossible to implement future optimizations to efficiency: eventually, we'd like to be able to blindly append new fields to the end of existing objects without parsing all existing fields.