Currently not setting the table flags HA_NOT_EXACT_COUNT means the handler::records attribute is assumed to always be exact. This attribute is also used frequently for calculating statistics needed for query optimization. Since calculating the number of records in ndbcluster requires network communication it is not entirely free. For smaller queries this cost can be 50% of the total query execution cost. There should be some way to avoid re-calculating the exact number of records for every small query and instead use an earlier "good enough" count.