On the ordered time-series insert path we invariant that the buckets collection exists, but this is not necessarily the case if there is a race with the buckets collection being dropped.

Presumably we should implicitly create the buckets collection in this case, which is what should already be happening on the unordered insert path.