Inter-Operator Parallelism

Inter-Operator Parallelism#

Types of Partitioning#

Shared data can be partitioned in three ways by:

  1. Range

    • Great for equijoins, ranged queries, and group by

  2. Hash

    • Great for equijoins and group by

  3. Random (Round-Robin)

    • Good for spreading load

Queries#

  • Scan

    • Scan in parallel and merge before output

    • \(\sigma_p\) skip entire machines that has no tuples satisfying \(p\)

  • Search by Key

    • If data partitioned by range of key then use only that machine

    • Otherwise have every machine do lookup.

  • Insert

    • If partitioned by key, then insert there

    • Otherwise insert anywhere

    • If the key field is unique then check everything.

Contents