Currently, we always use an upsert to create/update a time-series bucket. However, we can instead use an insert when we are creating a new bucket (as opposed to extending an existing one), which should improve performance in this case.