Created by: gongweibao

1. Should a trainer exit when it meets errors? For example, read data error?And
2. Will The whole job exit when a specific task exit abnormally continuously many times which is large than our limit? When this happens, it means that these errors can't be fixed by starting new trainers. And
3. Do we need a TaskFail interface for trainer report it's error task before exit? A trainer reports task's error state by only timeout now, and it's very slow. And
4. We should change the life cycle of a single task picture to add failed task queue and so on.