Support SQL-compliant NaN behavior on eq_dyn, neq_dyn, lt_dyn, lt_eq_…

…dyn, gt_dyn, gt_eq_dyn (#2570)

* Add sql-compliant feature for enabling sql-compliant kernel behavior

* Rename feature flag and add document

.github/workflows/arrow.yml

@@ -54,6 +54,9 @@ jobs:
       - name: Test --features=force_validate,prettyprint,ipc_compression,ffi
         run: |
           cargo test -p arrow --features=force_validate,prettyprint,ipc_compression,ffi
+      - name: Test --features=nan_ordering
+        run: |
+          cargo test -p arrow --features "nan_ordering"
       - name: Run examples
         run: |
           # Test arrow examples

arrow/Cargo.toml

@@ -88,6 +88,8 @@ pyarrow = ["pyo3", "ffi"]
 force_validate = []
 # Enable ffi support
 ffi = []
+# Enable NaN-ordering behavior on comparison kernels
+nan_ordering = []
 
 [dev-dependencies]
 rand = { version = "0.8", default-features = false, features = ["std", "std_rng"] }

arrow/src/compute/kernels/aggregate.rs

@@ -28,7 +28,7 @@ use crate::datatypes::{ArrowNativeType, ArrowNumericType, DataType};
 
 /// Generic test for NaN, the optimizer should be able to remove this for integer types.
 #[inline]
-fn is_nan<T: ArrowNativeType + PartialOrd + Copy>(a: T) -> bool {
+pub(crate) fn is_nan<T: ArrowNativeType + PartialOrd + Copy>(a: T) -> bool {
     #[allow(clippy::eq_op)]
     !(a == a)
 }