westonpace commented on PR #14434: URL: https://github.com/apache/arrow/pull/14434#issuecomment-1281604803
> I don't think DecodeOptionlessOverflowableArithmetic is suitable for these functions, nor for division or for any of the floating-point versions of the functions it's already used for. At the moment we should focus on exposing the capability that we have, and describing it accurately. For example, in Arrow we have "_checked" but "what exactly is checked?" For example, in Arrow we have `divide` and `divide_checked`. They seem to behave as follows (through trial and error): | Method | Data Type | Problem | Behavior | | --- | --- | --- | --- | | divide | integral | overflow | silent | | divide | integral | divide by zero | error | | divide | integral | domain error | error | | divide | floating | divide by zero | limit | | divide | floating | domain error | nan | | divide_checked | integral | overflow | error | | divide_checked | integral | divide by zero | error | | divide_checked | integral | domain error | error | | divide_checked | floating | divide by zero | error | | divide_checked | floating | domain error | error | Thus we should have the following behavior for substrait calls | Call | Data Type | Options | Arrow Result | | --- | --- | --- | --- | | divide | integral | overflow=SILENT | divide | | divide | integral | overflow=SATURATE | reject plan | | divide | integral | overflow=ERROR | divide_checked | | divide | floating | rounding=* | reject plan[1] | | divide | floating | on_domain_error=NAN | divide | | divide | floating | on_domain_error=ERROR | divide_checked | | divide | floating | on_division_by_zero=LIMIT | divide | | divide | floating | on_division_by_zero=NAN | reject plan | | divide | floating | on_division_by_zero=ERROR | divide_checked | | divide | floating | on_domain_error=NAN on_division_by_zero=ERROR | reject plan[2] | | divide | floating | on_domain_error=ERROR on_division_by_zero=LIMIT | reject plan[2] | 1. I'm not sure what rounding behavior we have implemented in Acero and therefore the safest thing to do would be to reject all plans that are asking for a specific rounding behavior. 2. We cannot satisfy these specific combinations between divide and divide_checked. Caveat / Needs clarification: IMHO, It is not clear, in Substrait, if division by zero for integral arguments counts as overflow or not. Also, it is possible to have 0/0 in integer arithmetic, which I have been interpreting as "domain error" (technically division by zero requires non-zero/0) but have no idea what the integral behavior should be. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: github-unsubscr...@arrow.apache.org For queries about this service, please contact Infrastructure at: us...@infra.apache.org
