SparkDriverLess

RDD

For every wide dependency transformation, create another repartition RDD between its parent and itself. So that all transformations become narrow dependent, except repartition.

Client

Background server:

1. Worker discoverer
2. Partition discoverer
3. Job server:
   • After a job is taken by a worker, deactivate the job and set a timer.
   • If timeout and no result returned, reactivate the job for others.

Algorithm:

1. Create RDD lineage by user's input.
2. When collect results:
   1. Create partitions lineage, passing in function closure. There is no pointer in partition object to RDD.
   2. For every partition in target RDD, find it in partition_discoverer. Fetch it if it exists.
   3. For every missing partition, create a job in job_server.
   4. keep discover those missing partitions. Fetch finished partitions from workers and remove that job from job_server.
   5. After all partitions are done, merge all results and return.
3. Repeat

Worker

Background servers:

1. Worker server: broadcast itself as a worker service.
2. Job discoverer
3. Job server
4. Partition discoverer
5. Partition server

Algorithm:

1. Keep trying to get a job from job_discoverer.
2. Fetch the partition associated to the job from remote.
3. Check if it's already in partition_server. Skip to next job if so.
4. Check all partitions lineage in partition_discoverer, get the results of the exists.
5. Run the function of the partition:
   • If it's a narrow dependent partition, or all dependencies are done:
     • Do it right away
   • Otherwise, for every missing dependency:
     1. Broadcast a job in job_server.
     2. Append current job back to jobs queue.
     3. Skip to next job.
6. When finish the partition successfully, add the result to partition_server
7. Repeat