When inserting into a time-series bucket, we keep a consistent schema/BSON type per column in each bucket. However, there appear to be some edge cases where we're not properly maintaining this property.