alamb commented on code in PR #18449:
URL: https://github.com/apache/datafusion/pull/18449#discussion_r2535597293
##########
datafusion/physical-expr/src/expressions/in_list.rs:
##########
@@ -80,66 +75,56 @@ struct ArrayHashSet {
map: HashMap<usize, (), ()>,
}
-struct ArraySet<T> {
- array: T,
- hash_set: ArrayHashSet,
-}
-
-impl<T> ArraySet<T>
-where
- T: Array + From<ArrayData>,
-{
- fn new(array: &T, hash_set: ArrayHashSet) -> Self {
- Self {
- array: downcast_array(array),
- hash_set,
+impl ArrayHashSet {
+ /// Checks if values in `v` are contained in the `in_array` using this
hash set for lookup.
+ fn contains(
+ &self,
+ v: &dyn Array,
+ in_array: &dyn Array,
+ negated: bool,
+ ) -> Result<BooleanArray> {
+ // Null type comparisons always return null (SQL three-valued logic)
+ if v.data_type() == &DataType::Null || in_array.data_type() ==
&DataType::Null {
+ return Ok(BooleanArray::from(vec![None; v.len()]));
}
- }
-}
-impl<T> Set for ArraySet<T>
-where
- T: Array + 'static,
- for<'a> &'a T: ArrayAccessor,
- for<'a> <&'a T as ArrayAccessor>::Item: IsEqual,
-{
- fn contains(&self, v: &dyn Array, negated: bool) -> Result<BooleanArray> {
downcast_dictionary_array! {
v => {
- let values_contains = self.contains(v.values().as_ref(),
negated)?;
+ let values_contains = self.contains(v.values().as_ref(),
in_array, negated)?;
let result = take(&values_contains, v.keys(), None)?;
return Ok(downcast_array(result.as_ref()))
}
_ => {}
}
- let v = v.as_any().downcast_ref::<T>().unwrap();
- let in_array = &self.array;
- let has_nulls = in_array.null_count() != 0;
+ let needle_nulls = v.logical_nulls();
+ let needle_nulls = needle_nulls.as_ref();
+ let haystack_has_nulls = in_array.null_count() != 0;
+
+ with_hashes([v], &self.state, |hashes| {
+ let cmp = make_comparator(v, in_array, SortOptions::default())?;
+ Ok((0..v.len())
+ .map(|i| {
+ // SQL three-valued logic: null IN (...) is always null
+ if needle_nulls.is_some_and(|nulls| nulls.is_null(i)) {
+ return None;
+ }
- Ok(ArrayIter::new(v)
- .map(|v| {
- v.and_then(|v| {
- let hash = v.hash_one(&self.hash_set.state);
+ let hash = hashes[i];
let contains = self
- .hash_set
.map
.raw_entry()
- .from_hash(hash, |idx|
in_array.value(*idx).is_equal(&v))
+ .from_hash(hash, |idx| cmp(i, *idx).is_eq())
Review Comment:
I mean specialize the entire thing (including hash table) - so that you pay
the dispatch once (either at InLIstExpr creation time or maybe once per batch),
rather than on each row
Another problem with the dyn comparator approach is that it prevents
inlining/vectorization
Here is one way to specialize the hashset:
- https://github.com/pydantic/datafusion/pull/45
(I haven't fully worked out the generics yet)
--
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: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
