GraphQL JIT – GraphQL execution using a JIT compiler

I remember seeing graphql-jit on a comparative benchmark of GQL server implementations, and it blowing everything out of the water, really impressive:

https://github.com/benawad/node-graphql-benchmarks

Depending on your usecase, you can go one level further and directly evaluate the query AST without delegating to resolvers.

(This generally only makes sense if you do programmatic schema generation. IE, read some metadata about a source and then generate GQL types/operations from it.)

So a query comes in, say something like:

    query {
        users(where: { name: { like: "foo" } }) {
            id
            email
            todos(where: { completed: { eq: true } }) {
                description
            }
        }
    }

Rather than having resolvers, and firing the "users" and "users.todos" resolvers, you interpret the AST of the query directly as if it were a regular value

In the case of a SQL database, you'd have a recursive function that takes a level of the query AST and generates a SQL string. Then you'd execute the query, format the results, and return it to the client.

IE, in this case you'd probably generate something like

  SELECT id, email, JSON_OBJECT('todos', ...) AS todos
  FROM users JOIN todos ON todos.user_id = users.id

As far as the client can tell, it's a "regular" GraphQL server, but the serverside implementation doesn't have any kind of real executable schema. It just behaves like it does.

The downside of this approach is that you lose the ability to use built-in functionality for things like tracing, query depth/complexity limiting, etc. You'd have to write this on your own.

It's good to see this happening, and happening in the open. In both JS and Python I've seen GraphQL execution take a ridiculous amount of time using the commonly used libraries, so long that we had fall back to raw JSON content in some parts of our schema that previously had a lot of small resolvers. There was a Python semi-JIT experiment a while ago, but it was paid and had some dodgy security implications os we couldn't use it, but the concept is a strong one.

This has been an effective approach for other areas like templating in Jinja2, so I'm not surprised that it's being used for GraphQL.

If you need to go to these steps to make optimizations, would it not make more sense to switch to a more bare metal programming language such as Rust for the entire stack?

Why not just go with graphql-aot and be done with it, plus it would work on serverless which I guess this wouldn't.

It'd be nice if they talked about what their compilation did.