Fix(bigquery)!: parse information schema views into a single identifier

[georgesittas]

Behavior in main today:

>>> from sqlglot import parse_one
>>> parse_one("select * from db.information_schema.tables", "bigquery")
Select(
  expressions=[
    Star()],
  from=From(
    this=Table(
      this=Identifier(this=tables, quoted=False),
      db=Identifier(this=information_schema, quoted=False),
      catalog=Identifier(this=db, quoted=False))))
>>> parse_one("select * from c.db.information_schema.tables", "bigquery")
Select(
  expressions=[
    Star()],
  from=From(
    this=Table(
      this=Dot(
        this=Identifier(this=information_schema, quoted=False),
        expression=Identifier(this=tables, quoted=False)),
      db=Identifier(this=db, quoted=False),
      catalog=Identifier(this=c, quoted=False))))

This PR makes it so that we will always represent INFORMATION_SCHEMA.X views using a single Identifier expression for BigQuery. The motivation here is that, previously, we wouldn't properly qualify information schema references, because we'd think there's a db/catalog present.

Choosing Identifier over a Dot is motivated by the fact that we the schema module can't handle multiple mapping depths, so this should keep the nesting consistent even if there are information schema view references.

Xref: TobikoData/sqlmesh#3317

[barakalon]

I’m mostly familiar with MySQL’s information_schema. In MySQL, information_schema is a database, so I think this would be wrong?

Does this PR imply that Bigquery creates default information_schema tables in each database? And if so, I wonder how common that is.

[georgesittas]

I’m mostly familiar with MySQL’s information_schema. In MySQL, information_schema is a database, so I think this would be wrong?

Does this PR imply that Bigquery creates default information_schema tables in each database? And if so, I wonder how common that is.

Ah, got it - thanks for catching this. I did assume there would be an information schema under each db, but I'll double-check this assumption.

Regarding BigQuery, yes, they need the information schema to be qualified with a region or dataset:

An INFORMATION_SCHEMA view needs to be qualified with a dataset or region.

This means you can have up to 4 parts in a Table reference, if you were to include both the region and the project name.

[georgesittas]

@barakalon it turns out your hunch was right, none of the other engines treats the information schema views like BigQuery does, so this fix really needs to be specific to it. Thanks for catching this, should've investigated more thoroughly.

[georgesittas]

There's some more work to be done here around 4-part table references in the schema module, because right now we're expecting the references to have the same depth but information schema views break this invariant.

[georgesittas]

This change should be safe because BigQuery aliases are case-insensitive.