And @tustvold notes For people looking to emulate spark which only supports precision up to 38, casting to Decimal256 and then truncating down to Decimal128 will be equivalent, and is what a precision loss arithmetic kernel would do
Describe the solution you'd like
If anyone needs spark compatible decimal division rules, I suggest:
Add a new config option
Apply the rewrite suggested by @tustvold (cast to Decimal256, divide, and then cast to Decimal128) as an AnalyzerRule
Describe alternatives you've considered
See ticket -- we discussed at length changing the semantics of division in arrow-rs and concluded there was no one agreed upon ideal behavior
Is your feature request related to a problem or challenge?
As described in detail by @liukun4515 and @tustvold and @viirya on https://github.com/apache/arrow-datafusion/pull/6832, DataFusion's decimal devision semantics.
@liukun4515 notes https://github.com/apache/arrow-datafusion/pull/6832#issuecomment-1680098056 that spark has the config to control the precision loss : https://github.com/apache/spark/blob/2be20e54a2222f6cdf64e8486d1910133b43665f/sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/expressions/arithmetic.scala#L246
And @tustvold notes For people looking to emulate spark which only supports precision up to 38, casting to Decimal256 and then truncating down to Decimal128 will be equivalent, and is what a precision loss arithmetic kernel would do
Describe the solution you'd like
If anyone needs spark compatible decimal division rules, I suggest:
Describe alternatives you've considered
See ticket -- we discussed at length changing the semantics of division in arrow-rs and concluded there was no one agreed upon ideal behavior
Additional context
No response