Repartition Mapping

Shuffles data to produce a specified amount of Spark partitions by using the specified columns to determine the partitioning. This operation is useful for caching carefully repartitioned data before multiple join operations.

Example

mappings:
  repartitioned_customer_basket:
    kind: repartition
    input: customer_basket
    partitions: 200
    columns:
     - customer_id
     - product_id
   sort: true

Fields

  • kind (mandatory) (string): repartition
  • broadcast (optional) (type: boolean) (default: false): Hint for broadcasting the result of this mapping for map-side joins.
  • cache (optional) (type: string) (default: NONE): Cache mode for the results of this mapping. Supported values are
    • NONE - Disables caching of teh results of this mapping
    • DISK_ONLY - Caches the results on disk
    • MEMORY_ONLY - Caches the results in memory. If not enough memory is available, records will be uncached.
    • MEMORY_ONLY_SER - Caches the results in memory in a serialized format. If not enough memory is available, records will be uncached.
    • MEMORY_AND_DISK - Caches the results first in memory and then spills to disk.
    • MEMORY_AND_DISK_SER - Caches the results first in memory in a serialized format and then spills to disk.
  • input (mandatory) (string): The name of the input mapping
  • columns (mandatory) (list:string): The list of column names used for partitioning the data
  • partitions (optional) (integer): The number of output partitions
  • sort (optional) (boolean): Specifies if the records within each partition should also be sorted
  • filter (optional) (type: string) (default: empty): An optional SQL filter expression that is applied before the repartition operation itself.

Outputs

  • main - the only output of the mapping

Description

This transformation can be used as part of a processing optimization where you want to repartition the result of some mapping because you perform multiple join operations on the same columns afterwards.