aweltsch commented on PR #14283:
URL: https://github.com/apache/datafusion/pull/14283#issuecomment-2613791296
I also looked into this a little bit. Here's my observations:
The `data_type` of a `CaseExpr` is the first non-null `data_type`. In the
`expr_or_expr` specialization this is only the same type as the else expression
if the _then_expr is null_. My impression is that @jayzhan211 's suggestion to
handle only the null value would thus make sense.
I personally would have applied the transformation
```rust
let e = &self.when_then_expr[0].1;
let then_expr = try_cast(Arc::clone(e), &batch.schema(),
return_type.clone())?;
let then_value = then_expr
.evaluate_selection(batch, &when_value)?
.into_array(batch.num_rows())?;
```
since to me this seems more obvious what's the intended goal, i.e. "make
sure that the then_expr returns the correct datatype for the case expression"
and seems to be a little bit more generic.
If the suggested fix here is to be accepted I would rather remove the
`expr_or_expr` specialization again, because from what I have seen the current
suggestion regresses on the `case_when: expr_or_expr` benchmark.
Also: I was surprised to see that I can't reproduce this issue when using
datafusion-cli, so from an end-to-end perspective the type coercion seems to
work somehow :thinking:
This is what I did:
```sql
> CREATE TABLE batch2(a int[]) AS VALUES ([1, 2, 3]), (null);
> SELECT CASE WHEN a IS NULL THEN NULL ELSE a END FROM batch2;
+--------------------------------------------------------+
| CASE WHEN batch2.a IS NULL THEN NULL ELSE batch2.a END |
+--------------------------------------------------------+
| [1, 2, 3] |
| NULL |
+--------------------------------------------------------+
```
--
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]
