Redesign `TypeName`, `DynIden` and `Expr` for correctness

[Expurple]

See the thread starting from this comment.

Let's continue that discussion about redesigning these types. They should be able to represent most valid SQL type names, identifiers and expressions, and none of the invalid ones.

[Huliiiiii]

Expressions can be divided into two categories:

Terminal/Atomic nodes:

• Value
• Keyword
• Iden

Non-Terminal/Composite nodes:

• Unary
• Binary
• List (separated by commas / space)
• Wrapped (parenthesized, arrays)

A rough design might look like this:

enum Expr {
    Value(..),
    Keyword(..),
    Iden(..),

    Unary(UnOper, Self),
    Binary(Self, BinOper, Self),
    List(Vec<Self>, Separator),
    Wrapped(Self, Wrapper),

    Raw(..),
}

[Huliiiiii]

Or we can convert exprs into structs and turn it into Box<dyn Buildable>

[Expurple]

Your list of non-terminal nodes assumes that they can only be built of other expressions and operators/separators/parentheses that combine them; and that everything can be "lowered" to these primitives. But that's not the case. FunctionCall, SubQuery, AsEnum and Case can also contain other things like keywords (DISTINCT in function args, CASE control flow) and type names. We have Keyword and TypeName variants today, but that's a hack that needs to be removed. Those aren't valid expressions. We need to have specific, "special", high-level structs/variants (like FunctionCall, SubQueryStatement and CaseStatement) to properly handle such nodes.

I don't see how Buildable structs change anything. We still need to correctly design these structs, just like we need to correctly design Expr variants today. Separate structs may be useful to represent specific AST nodes in the type system, but for that we can "extract" those without getting rid of Expr. The corresponding Expr variant would be left with one field which is that struct

[Huliiiiii]

FunctionCalls can be represented as:

List(
  [
    Keyword, 
    Wrapped(
      List(
        ...
      ),
      Parenthesis
    ),
  ],
  Nothing
)

But this is a little bit complicated.

Higher-level structures or variants can make it easier to construct more complex expressions, but more variants will bring higher complexity.

[Expurple]

I understood this idea. Yes, this is possible. But the problem in my view is that Keyword is not actually a valid Expr. IMO, Expr::Keyword and Expr::TypeName should not exist. If we remove those, some SQL constructs can no longer be represented using your "common" low-level method

[Huliiiiii]

How about this?

enum Atom {
   Lit(..),
   Keyword(..),
   Iden(..),
   Value(..),
   ..
}

enum Expr {
  Atom(Atom),
  ..
}

[Expurple]

Doesn't change anything. These types still allow you to construct invalid expressions from arbitrary keywords: let keyword_expr = Expr::Atom(Atom::Keyword(..)). I don't want to allow expressions like IF + 3. I want to have a clear, obvious contract:

"An Expr is a syntactically correct SQL expression (unless the user has opted into custom raw SQL strings)."

Most developers would expect Expr to be a fuckin expression, not "an arbitrary AST node that can also be something like BEGIN or int[] because that was easier to implement". It's a usability and maintainability problem. If you blur several unrelated concepts into a single "flexible" type, eventually you hit problems like #827. Turns out, identifiers, type names and column names are three different things! They need to be quoted differently, support different syntax, support different conversion constructors, etc. Expressions, keywords and type names are also three different things that need to be handled differently.

[Huliiiiii]

Flexibility and correctness are often difficult to achieve at the same time.

I think one possible approach is to split different types of expressions into separate structs, and define their serialization to SQL through traits. However, this approach may either sacrifice some flexibility, or require complex trait bounds to preserve it, which can lead to confusing error messages.

[Expurple]

For flexibility, we already have cust* methods and Cust/Custom variants. I adore them. When something isn't supported natively, they usually provide a working solution. I'm not sure if we need a different, semi-structured "arbitrary node" workaround in addition to that.

I know that you need #926. But if we look closer at it, an "expression list" isn't itself a valid expression that can be freely composed with other expressions. Which means that ExprList should be a separate type, rather than a new variant of Expr. I think, we should add that type, then add a backend function to render that into an SQL string, and then you can safely format that rendered string into a bigger raw string that comprises your JSON_EXISTS function call.

(Update: I've reposted this comment under your original Expr::List proposal. We can continue this thread over there if that's more appropriate)

I think one possible approach is to split different types of expressions into separate structs, and define their serialization to SQL through traits.

I don't see how that's relevant here. We need some "general expression node" in order to compose expressions. That's just how the SQL grammar works. If the Expr non-terminal was expressed as Box<dyn Buildable> instead of enum Expr , I would still argue that ExprList, TypeName and Keyword shouldn't implement Buildable and shouldn't be usable as expression nodes.

(Although, they'd obviously have to implement some other trait in order to be accepted by the backend in other contexts. That's equivalent to having "a backend function to render ExprList" that I propose)

[Huliiiiii]

I think it’s fine to discuss this here.
In your view, which variants should remain in Expr, and which should be split out?
What are your thoughts on the overall design of the new API?

[Expurple]

It's really simple. Only the variants that are actually expressions should stay in Expr. Use expr + 1 as a quick test. "Real" expressions should be composable.

I don't have time to consult the grammar and check every variant that we have today. But it's obvious that Expr::TypeName doesn't pass this test and should be removed. In practice, it's only used to implement casts, so we should replace it with a Cast(Box<Expr>, TypeName) variant. It's very similar to the existing Expr::AsEnum variant, so we should dive into that and see if these variants can be unified.

Sorry, I caused some confusion about Keyword previously. I should've checked the Keyword contents. Turns out, it's already "only the keywords that are usable as expressions", not "arbitrary SQL keywords". Expr::Keyword can stay.

[Huliiiiii]

Cast sounds good.

AsEnum just... Cast

sea-query/src/backend/postgres/query.rs

Lines 39 to 51 in 8346526

	Expr::AsEnum(type_name, expr) => {
	write!(sql, "CAST(").unwrap();
	self.prepare_simple_expr_common(expr, sql);
	let q = self.quote();
	let type_name = type_name.to_string();
	let (ty, sfx) = if type_name.ends_with("[]") {
	(&type_name[..type_name.len() - 2], "[]")
	} else {
	(type_name.as_str(), "")
	};
	write!(sql, " AS {}{}{}{})", q.left(), ty, q.right(), sfx).unwrap();
	}
	_ => QueryBuilder::prepare_simple_expr_common(self, simple_expr, sql),