Spurious cast to Uint8 inserted when coalesce is used in a context of a join
- Dominant language
- Rust
- Stars
- 9.3k
- Forks
- 2.4k
- Avg merge
- 3d 7h
- Merged PRs (30d)
- 344
Description
### Describe the bug
Following is a valid (as far as I can tell) SQL that blows up in DataFusion:
```
DataFusion CLI v20.0.0
❯ with customer as (
select 11111 as c_custkey
), cust as (
select case when c_custkey = 0 then null else c_custkey end as c_custkey from customer
)
select coalesce(a.c_custkey, b.c_custkey) as custkey
from cust a inner join cust b
on a.c_custkey = b.c_custkey;
Arrow error: Cast error: Can't cast value 11111 to type UInt8
```
The culprit seems to be a spurious cast to UInt8 that gets inserted in coalesce arguments:
```
❯ explain with customer as (
select 11111 as c_custkey
), cust as (
select case when c_custkey = 0 then null else c_custkey end as c_custkey from customer
)
select coalesce(a.c_custkey, b.c_custkey) as custkey
from cust a inner join cust b
on a.c_custkey = b.c_custkey;
+---------------+----------------------------------------------------------------------------------------------------------------------------------------------+
| plan_type | plan |
+---------------+----------------------------------------------------------------------------------------------------------------------------------------------+
| logical_plan | Projection: coalesce(CAST(a.c_custkey AS UInt8), CAST(b.c_custkey AS UInt8)) AS custkey |
| | Inner Join: a.c_custkey = b.c_custkey |
| | SubqueryAlias: a |
| | SubqueryAlias: cust |
| | Projection: CASE WHEN customer.c_custkey = Int64(0) THEN Int64(NULL) ELSE customer.c_custkey END AS c_custkey |
| | SubqueryAlias: customer |
| | Projection: Int64(11111) AS c_custkey |
| | EmptyRelation |
| | SubqueryAlias: b |
| | SubqueryAlias: cust |
| | Projection: CASE WHEN customer.c_custkey = Int64(0) THEN Int64(NULL) ELSE customer.c_custkey END AS c_custkey |
| | SubqueryAlias: customer |
| | Projection: Int64(11111) AS c_custkey |
| | EmptyRelation |
...
```
### To Reproduce
See above repro in the CLI
### Expected behavior
The query above is expected to succeed
### Additional context
Any simplification to the above query I tried (e.g. removing join, case, or coalesce) makes the issue go away.
E.g. here is a query where b.custkey is replaced by 0 in the coalesce call:
```
❯ with customer as (
select 11111 as c_custkey
), cust as (
select case when c_custkey = 0 then null else c_custkey end as c_custkey from customer
)
select coalesce(a.c_custkey, 0) as custkey
from cust a inner join cust b
on a.c_custkey = b.c_custkey;
+---------+
| custkey |
+---------+
| 11111 |
+---------+
1 row in set. Query took 0.021 seconds.
```
Contributor guide
Research direction
Run the reproducer in the DataFusion CLI and compare its output with the EXPLAIN plan, focusing on the UInt8 casts added to coalesce arguments after the join. Trace the logical-plan and type-coercion path that produces that plan, then add a regression test for the query and verify it succeeds without the spurious cast.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- rust, sql
- Domain
- databases
- Issue type
- Bug
- Difficulty
- 4/5
- Estimated time
- 3-5 days
- Activity status
- Active
- Clarity
- Mostly clear
- Newbie friendliness
- 55/100