IIRC a typical scenario we've observed lead to cancelation is datastore hedging. A concurrent request is issued when the operation takes longer than usual. As soon as one of the two come back, the one still in flight will be cancelled.