[RESOLVED] Warning if variable is re-assigned?

[richb-hanover]

Currently, PRQL sets a variable to its latest assignment, losing/ignoring any earlier value. For example:

select [
a = 1,
a = 2,
]

yields

SELECT
  2 AS a

It seems to me that this is almost always an error - either a simple typo or a deeper problem - and should be flagged as an error.

Is there ever a case where re-assignment could be useful? Thanks.

[max-sixty]

I'm less persuaded this is almost always an error. e.g. say we want to replace nulls with zero:

from employees
derive salary = (salary ?? 0)

Almost all programming languages allow reassignments andor mutations...

[richb-hanover]

Almost all programming languages allow reassignments and/or mutations...

Yes, but the mistake I'm trying to guard against is mistyping a complicated column name (I have plenty of similar ones in my application: SDAppTotal2021 and SDAssTotal2021) and having it disappear from the resulting table because it got reassigned...

That said, your example is a compelling use case. Furthermore, it "does the right thing" if extended like this:

from employees
derive [
  salary = (salary ?? 0),
  salary = salary-7,
]

generates:

SELECT
  employees.*,
  COALESCE(salary, 0) - 7 AS salary
FROM
  employees

It seems there's tension generated by the dual purpose of variable = expression:

• create a column in the resulting table
• give a name to that column to use in subsequent calculations.

I was initially thinking about the way Javascript makes a distinction between const and let in declarations:

const a = 1 // compiler can detect re-assignment to 'a' as an error
let b = 1   // re-assignment is allowed

I wonder if it's sufficient to require re-assignment to occur on one line: e.g., salary = (salary ?? 0) - 7, so that the appearance of another salary = ... could trigger a warning/error...

[max-sixty]

There was some discussion on that in #6.

In retrospect, I think we came to the correct decision given the current model's simplicity. Though I can empathize with the stricter model's approach.

[richb-hanover]

Thanks for that link to the previous discussion. I'll just have to be careful when typing variable names. (And why wouldn't I want to do this? :-)