alpaylan opened a new issue, #26620:
URL: https://github.com/apache/datafusion/issues/26620
### Describe the bug
Two window functions with the same `ORDER BY <expression>` make the physical
planner panic in debug builds, when
one of the window functions has an argument that gets constant-folded (`1 +
1`):
```
thread 'main' panicked at datafusion/core/src/physical_planner.rs:1116:21:
all window expressions shall have the same sort keys, as guaranteed by
logical planning
```
`EXPLAIN` of the query panics too. Release builds skip the assertion.
### To Reproduce
Run in `datafusion-cli` built with debug assertions (for example `cargo run
-p datafusion-cli`):
```sql
CREATE TABLE t (a BIGINT, b BIGINT) AS VALUES (1, 10), (2, 20);
SELECT sum(1 + 1) OVER (ORDER BY a + 1), sum(b) OVER (ORDER BY a + 1) FROM t;
```
```
thread 'main' panicked at datafusion/core/src/physical_planner.rs:1116:21:
all window expressions shall have the same sort keys, as guaranteed by
logical planning
```
It also panics with `max(1 + 1)`, `count(1 + 1)` or a different second
function, and when the two window functions
are in the other order. It does not panic with `max(2)` instead of `max(1 +
1)`. It does not panic with `INT`
columns either: there `a + 1` gets a cast, which changes the expression.
### Expected behavior
The query returns two rows, `2, 10` and `4, 30`. The logical planner put
both window functions in
one `WindowAggr` with the same `ORDER BY`, so the physical planner should
not hit its assertion.
### Additional context
- Tested on `main` at c0ec5127ac99 (2026-10-22), debug build: panics. Latest
release 55.2.0 built in debug mode: panics at
`physical_planner.rs:901` (same assertion). The 55.2.0 release binary (no
debug assertions) returns the correct rows.
- Where in the code: `simplify_expressions` folds `1 + 1` and wraps that
window expression in an alias to keep its
name. Then `common_sub_expression_eliminate` replaces the shared `a + 1`
with `__common_expr_1`. It adds an inner
alias (`__common_expr_1 AS t.a + Int64(1)`) only in the window expression
that has no outer alias. The
`debug_assert!` in `DefaultPhysicalPlanner` compares the sort keys
including aliases, so the two keys look
different. With that assertion relaxed to ignore aliases, `EXPLAIN` shows
the optimized node (frames trimmed):
`WindowAggr: windowExpr=[[sum(Int64(2)) ORDER BY [__common_expr_1 ASC
NULLS LAST] AS ..., sum(t.b) ORDER BY [__common_expr_1 AS t.a + Int64(1) ASC
NULLS LAST] AS ...]]`
Found with an automated differential/metamorphic testing tool; reproduced on
current `main`.
### Willingness to contribute
I would like to contribute a fix for this bug, but request guidance from the
DataFusion community
--
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]
