You are viewing a plain text version of this content. The canonical link for it is here.
Posted to issues@arrow.apache.org by "Neal Richardson (Jira)" <ji...@apache.org> on 2022/08/18 18:51:00 UTC

[jira] [Created] (ARROW-17462) [R] Cast scalars to type of field in Expression building

Neal Richardson created ARROW-17462:
---------------------------------------

             Summary: [R] Cast scalars to type of field in Expression building
                 Key: ARROW-17462
                 URL: https://issues.apache.org/jira/browse/ARROW-17462
             Project: Apache Arrow
          Issue Type: Improvement
          Components: R
            Reporter: Neal Richardson
            Assignee: Neal Richardson
             Fix For: 10.0.0


After looking at the ExecPlan output of some queries, it jumped out at me how we translate {{ int_field == 5 }} in R as {{ cast(int_field, float64) == 5 }} because 5 is a double in R. 

This extra work has a noticeable performance impact. Here's a simple query on the taxi dataset, filtering down to 54 out of 1.5 billion rows and selecting a single column. My idea was to make a query that does not much other than evaluate the filter. 

{code}
> system.time(ds |> select(passenger_count) |> filter(passenger_count > 10) |> compute())
   user  system elapsed 
  0.391   0.024   0.362 

> system.time(ds |> select(passenger_count) |> filter(passenger_count > Scalar$create(10, type = int8())) |> compute())
   user  system elapsed 
  0.206   0.025   0.179 
{code}

You can see the difference in the query plans too:

{code}
> ds |> select(passenger_count) |> filter(passenger_count > 10) |> explain()
ExecPlan with 4 nodes:
3:SinkNode{}
  2:ProjectNode{projection=[passenger_count]}
    1:FilterNode{filter=(cast(passenger_count, {to_type=double, allow_int_overflow=false, allow_time_truncate=false, allow_time_overflow=false, allow_decimal_truncate=false, allow_float_truncate=false, allow_invalid_utf8=false}) > 10)}
      0:SourceNode{}

> ds |> select(passenger_count) |> filter(passenger_count > Scalar$create(10, type = int8())) |> explain()
ExecPlan with 4 nodes:
3:SinkNode{}
  2:ProjectNode{projection=[passenger_count]}
    1:FilterNode{filter=(passenger_count > 10)}
      0:SourceNode{}
{code}

Ideally Acero would do this more intelligently (cf. ARROW-11402), but we should also be able to do smarter things when assembling the Expression in R.



--
This message was sent by Atlassian Jira
(v8.20.10#820010)