Skip to content

Improve performance of the SUM accumulator #6878

@alamb

Description

@alamb

Is your feature request related to a problem or challenge?

While working on #6800 / #4973 I noticed that the SUM accumulators intermediate state contains two fields:

  • sum: Native
  • count: usize

Maintaining this state is on the hit path of the aggregates (e.g. TPCH Q1 has 4 of sums) thus reducing the work is performance critical

The counts are only needed for the retractable version of the accumulator (aka sliding accumulator used in window functions) but they are actually carried through all implementations.

Sum accumulator is here https://github.com/apache/arrow-datafusion/blob/main/datafusion/physical-expr/src/aggregate/sum.rs

Describe the solution you'd like

I would like to remove the counts field in the sum accumulator (or only use it when we have sliding windows)

Describe alternatives you've considered

No response

Additional context

No response

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementNew feature or request

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions