Can not infer where a column is from but then the error message says that column is available?

[cottrell]

What happened?

I'm not sure if this is a real bug or just a strange error message?

PRQL input

from data
join side:left other (other.id== data.id)
into A

from codes
into S

from A
select {
A.val
}

SQL output

Error: 
   ╭─[:7:3]
   │
 7 │   A.val
   │   ──┬──  
   │     ╰──── Cannot infer where `A`.val is from. It could be any of [(119, {}), (116, {})]
   │ 
   │ Help: available columns: `A`.val
───╯

Expected SQL output

Any sql.

MVCE confirmation

• Minimal example
• New issue

Anything else?

No response

[kgutwin]

I did some experimentation with this example.

The from codes \n into S section isn't necessary to reproduce the problem, this produces the same output:

from data
join side:left other (==id)
into A

from A
select {A.val}

Error: 
   ╭─[:6:9]
   │
 6 │ select {A.val}
   │         ──┬──  
   │           ╰──── Cannot infer where `A`.val is from. It could be any of [(119, {}), (116, {})]
   │ 
   │ Help: available columns: `A`.val
───╯

The error goes from being confusing to clear when you remove the into A \n from A sequence and remove the A reference in the select:

from data
join side:left other (==id)
select {val}

Error: 
   ╭─[:3:9]
   │
 3 │ select {val}
   │         ─┬─  
   │          ╰─── Ambiguous name
   │ 
   │ Help: could be any of: data.val, other.val
───╯

So the root cause of the error message is that A.val is ambiguous in that it could come from either data or other. Unfortunately, the into A \n from A break makes it impossible to directly reference columns from the source relations:

from data
join side:left other (==id)
into A

from A
select {data.val}

Error: 
   ╭─[:6:9]
   │
 6 │ select {data.val}
   │         ────┬───  
   │             ╰───── Unknown name `data.val`
───╯

The most obvious solution to me is to recommend explicitly defining columns using a select prior to the into A. This resolves the error from the user perspective.

from data
join side:left other (==id)
select {data.val, other.col}
into A

from A
select {A.val}  # no error now

This might actually be the only true solution to this problem, since there doesn't seem to be any other way for the resolver to disambiguate a column reference when there are multiple source relations with wildcard columns.

[cottrell]

Ah, yes I was just thinking in general about explicit col defines via select and didn't think to try it in these exampels.

I wonder if this is perhaps the ONE COMPROMISE that PRQL should really push hard on it's users. The more I get into the weeds the more I understand that very little can be done without knowing the original column names.

[max-sixty]

This issue has been fixed. The query now compiles correctly without the confusing "Cannot infer" error. Tested with PRQL version 0.13.10.