IMHO @cassandra's heap estimate (MT * 3 * #CF + 1GB + caches) is wrong. If you have large indices, you run out of mem before the 1st flush.