[jira] [Commented] (BEAM-11048) Add alternate Sorting transform as an implementation of CombineFn

["Claire McGinty (Jira)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/BEAM-11048?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17211087#comment-17211087 ] 

Claire McGinty commented on BEAM-11048:
---------------------------------------

Implementation here: https://github.com/apache/beam/pull/13059

> Add alternate Sorting transform as an implementation of CombineFn
> -----------------------------------------------------------------
>
>                 Key: BEAM-11048
>                 URL: https://issues.apache.org/jira/browse/BEAM-11048
>             Project: Beam
>          Issue Type: Improvement
>          Components: extensions-java-sorter
>            Reporter: Claire McGinty
>            Priority: P2
>
> My team has been using the [SortValues|https://github.com/apache/beam/blob/master/sdks/java/extensions/sorter/src/main/java/org/apache/beam/sdk/extensions/sorter/SortValues.java] transform in `extensions-java-sorter` to sort pre-grouped values by a secondary sorter key. However, for large key groups, we've run into many OOM issues and have to increase disk size quite a bit to accommodate the larger key groups spilling to disk, even if there are only a few large key groups and most fit in memory.
> I drafted a new iteration of a Sorter that's a distributed merge-sort implemented as a `CombineFn`: each Accumulator maintains an always-sorted list of elements, and those Accumulators can be merged simply by zipping their lists together. This has the extra advantage that `extractOutput` can be lazily evaluated as a merging Iterator rather than as a fully materialized list. I also observed that this implementation is able to scale more effectively than the old SortValues, and for several use cases where `SortValues` ran OOM, the CombineFn-based implementation was able to complete using only the default Dataflow disk specs.
> Finally, from an API perspective, I think it's a little easier to use, because the user doesn't have to extract the sortKey out into the PCollection itself, but instead provide a function mapping each element type T to its sort key K, which will be evaluated inside the combiner. So I think in that sense it's more intuitive and similar to a Comparator-style sort.



--
This message was sent by Atlassian Jira
(v8.3.4#803005)