Fix!: render audits at runtime

[georgesittas]

The motivation behind this PR is that, given this model:

MODEL (
  name test_model,
  kind INCREMENTAL_BY_UNIQUE_KEY (
    unique_key (a, d)
  ),
  audits (
    unique_combination_of_columns(
      columns = (a, d),
      condition = d between @start_date and @end_date
    )
  ),
);

SELECT
  1 AS a,
  CAST('2025-01-20' AS DATE) AS d

The audit condition renders as follows at runtime today:

"d" BETWEEN CAST('1970-01-01' AS DATE) AND CAST('1970-01-01' AS DATE)

The reason this happens is that the model meta is rendered at load time and we store the properties to the corresponding pydantic fields, e.g. audits, even though start_date is a variable that changes depending on the runtime stage.

Approach is similar to e104491.

[erindru]

LGTM

Would this change the fingerprint of existing models that had audits?

[georgesittas]

Would this change the fingerprint of existing models that had audits?

It does, but I didn't notice it during testing because the old, rendered version of audits had been cached locally and so there was no metadata diff!

EDIT: after discussing offline, I realized that we can't really heal this with a migration because the snapshot payload would previously contain the following entry for audits:

"audits":[["unique_combination_of_columns",{"columns":"(a, d)","condition":"d BETWEEN CAST(\'1970-01-01\' AS DATE) AND CAST(\'1970-01-01\' AS DATE)"}]]

So, there’s no information about using variables whatsoever. The snapshot will still have the rendered date in it, but the local model will now have the @start_date and @end_date variables and we’ll inevitably get a metadata diff.