agg enable corr covar_pop and covar_samp

[liujiayi771]

enable corr, covar_pop and covar_samp.

[zhejiangxiaomai]

Would you create a pr to upstream this change to velox community when you are free? @liujiayi771

[liujiayi771]

Would you create a pr to upstream this change to velox community when you are free? @liujiayi771

OK. I will discuss this order of calculation with meta. Because the current modification is to meet the calculation results consistent with spark, but it may lead to inconsistent results with presto, we need to see how to solve it better.

[zhejiangxiaomai]

Would you create a pr to upstream this change to velox community when you are free? @liujiayi771

OK. I will discuss this order of calculation with meta. Because the current modification is to meet the calculation results consistent with spark, but it may lead to inconsistent results with presto, we need to see how to solve it better.

Thank you very much to explain the root cause.

[liujiayi771]

Would you create a pr to upstream this change to velox community when you are free? @liujiayi771

OK. I will discuss this order of calculation with meta. Because the current modification is to meet the calculation results consistent with spark, but it may lead to inconsistent results with presto, we need to see how to solve it better.

Thank you very much to explain the root cause.

I will try to fix this on our side without change Velox codes.

[liujiayi771]

We should import Epsilon Compare for floating-point value.

[rui-mo]

We should import Epsilon Compare for floating-point value.

Hi @liujiayi771, do you suggest importing it in Gluten CI?

[liujiayi771]

Hi, @rui-mo , maybe we can allow a certain error for floating-point numbers when comparing gluten results with vanilla spark. This is what velox does when comparing results with duckdb.

[rui-mo]

Hi, @rui-mo , maybe we can allow a certain error for floating-point numbers when comparing gluten results with vanilla spark. This is what velox does when comparing results with duckdb.

@liujiayi771 That makes sense to me. We can add some precision tolerance when comparing floats.