[
https://issues.apache.org/jira/browse/ARROW-18334?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17634552#comment-17634552
]
Weston Pace commented on ARROW-18334:
-------------------------------------
I had an offline discussion with [~bkietz] on this issue:
* Canonicalize will only reorder types if it thinks the functions is
commutative
* There is a small list of functions which are considered commutative (and
add_checked is one of them)
* So we shouldn't need to re-bind
One way to solve this problem is to create a custom output resolver for the
add(duration, timestamp) kernels which grabs the first timestamp type and uses
it as the output.
> Expression::Canonicalize does not unbind the expression from a kernel
> ---------------------------------------------------------------------
>
> Key: ARROW-18334
> URL: https://issues.apache.org/jira/browse/ARROW-18334
> Project: Apache Arrow
> Issue Type: Bug
> Components: C++
> Reporter: Weston Pace
> Priority: Major
>
> In general, we get away with this, since most kernels are of the shape
> {{FOO(T, T)}}. However, when the input types differ, this can be a problem.
> Concretely, we have kernels:
> {noformat}
> add_checked<Timestamp,Duration>() -> types[0]
> add_checked<Duration,Timestamp>() -> types[1]
> {noformat}
> A call is made with expression {{field_ref("x") + duration_literal}}. This
> call is bound to {{add_checked<Timestamp, Duration>}}. However, the
> expression is then simplified to {{duration_literal + field_ref("x")}}.
> Oddly enough, the math in this case is correct, since it is just addition,
> but the output type is not. It assigns an output type of duration instead of
> timestamp.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
