Totally agree with you. Especially about point 1. Codegen is one of the bluntest tools there is and should be almost always avoided.

Why is that? I’ve done a few toy attempts to build out ORM-like model classes in Typescript that wrap an input schema type using typelevel features, and my profiles of the Typescript compiler show a lot of slowdowns coming from combinatoric explosions in the typechecker. In this language specifically, the type system has quite advanced expressiveness but using those features in practice runs into limitations of the checker’s implementation frequently, and more so once you have many input types. We have ~80 table types and often a single table type is a union of variants - our biggest, the type for a Notion block is a union with 64 variants. I wrote a type that takes such row type as input, and outputs an OO-style Model interface for it with accessor methods, relational shortcuts, etc, and a matching Proxy that does the same at runtime. It works, and for smaller types seems like a fine solution. But in practice all the magic gums up Typescript’s inference and makes the language service crash into memory limits. Or the compiler just gives up and says “this union type is too complex to represent”.

If I move the same type magic into codegen, the compiler is much happier; plus stack traces lead directly to a simple method implementation instead of jumping into difficult-to-follow proxy machinery. The downside is that the codegen logic that consumes the Typescript compiler API to understand the schema types is more complex than the mapped types before, but it does make it feasible to have my cake (derive an ORM/model API from a schema) and eat it too (not make the compiler so unstable and touchy).

While in that specific case codegen does sound like the right approach, I think in most regular cases you would be able to get away without it. Honestly, a 64-variant type is not really that common.

I think codegen is OK as long as

- its designed in a way where it doesn't cause users to want or need to modify the generated code (this is the biggest downside of codegen)

- its against "relatively stable" API thing (protobuf / swagger descriptions, DB schema etc).

I said almost never not never, 64 variants is quite esoteric I'd say. The reason why I dislike codegen so much is that you fix one problem but introduce multiple new problems.

How do you handle the generated code? Do you check it into source control? If so, how do you ensure people do not touch it. Manually modified code generated code is one of the worst places to be in for maintenance. You mentioned you wanted language server support this means the generated code must not be ephemeral and continually visible to the rest of the project. When will this generated code need to be regenerated? On a schema change? How do you ensure the generated code always matches the schema?

Like I said all of these question can be answered, but it's quite a burden. In fact if I were you I would have looked into fixing the typescript compiler since the problems your ran into probably weren't fundamental. -

My general tactic for codegen hygiene like this is to check in the generated files, and on every CI re-run the code generator. After the generator runs, CI asserts there are no changes in Git. If there are any changes, the job fails.

This ensures that generated code is consistent with its inputs (including the generator logic) for every merge to main. It prevents people from editing the generated code, since their edits will cause a diff in CI and fail the build, but no one ever actually loses their work since the codegen does not run automatically/continuously during local development.

Checking in the files also makes it very easy to review changes to the generator since you can always tell how the output is changing.

The total burden for the above system is:

- a 30 line command called `Notion assert-clean`

- A 30 line CircleCI job that loops over the list of code generator commands, calls the command, then calls assert-clean.

We’ve used that tactic for years for simpler stuff like “make sure the SQL dump is consistent with the SQL migrations” and “put all the file names in this directory tree into a typescript file so we can tab complete them”.

The only necessary bit is that your code generator shouldn’t be spitting out enormous mountains of unreadable code.

I guess you are aware of Prisma. What is your opinion of it in this context?

Prisma is a nice DB client library but does't do any of the things I'm interested in - reactive queries on the client, recursive graph traversal, raising level of abstraction for the organization. Schematizing the DB & a type-safe DB client is nice, but I'm more interested in stuff one or two levels of abstraction up. Like, after the data comes into memory Prisma is done and out of the picture. But getting the data into memory is the easy part IMO. Traversing it, adding permissions and business logic, managing & composing mutations, dealing with caching and reactivity.. that's the good stuff, and I'm not sure what if any Prisma offers there.

Codegen is a super power. It has consistently made me 5 to 10 times more productive compared with experienced developers I have worked with.

I have more than once implemented production code where 80% of the code is generated. With only the biz logic hand coded. Including a complex 50+ screen Typescript application developed from scratch in just 3 months.

However, as with any other tool, you have to use Codegen intelligently to get the benefits. I have more than once seen Codegen being used really badly. Making the people using it think that Codegen is always a bad idea. Codegen is a complex tool and there are many ways to use it incorrectly.

> You can never solve those problems in a generic way for any domain.

This is rather final. Honestly curious if this was shown to be the case or just the result of failed earlier attempts.

None of the industrial languages I’ve looked at (Kotlin, Typescript, Go, Java, Swift, Rust) can do the things I want to do using only typelevel features, especially at practical scale. Eg, transform a schema type that describes my database tables into an easy-to-use query builder and data abstraction API with rich method chaining.

Often in Java/Kotlin projects people end up writing compiler plugins / annotation processors which is like moving the codegen to build time and making it harder to inspect/understand, while simultaneously spending more developer time on it since those systems may need to run for every compiler invocation. The same goes for Rust; rust’s various macro systems target the problem I have, but Rust users report long compile times due to macro magic. More esoteric languages like Scala and Haskell might have the expressive power I want, but don’t seem practical to implement for other reasons.

Of course there’s always the runtime only approach in Ruby/ActiveRecord but that runs into correctness problems at practical scales.

Codegen is usually brittle and is often ugly, but is guaranteed to have more metaprogramming power than in-language typelevel features. Ultimately any such system can be complex, I just hope to build one where the leverage I payed for with complexity is well worth it.

> More esoteric languages like Scala and Haskell might have the expressive power I want, but don’t seem practical to implement for other reasons.

Well, if you can't use languages that make it possible/feasible, then code generation might indeed be the best option. It's just that your claim sounded quite general, so I had to jump in. :)

> Eg, transform a schema type that describes my database tables into an easy-to-use query builder and data abstraction API with rich method chaining.

TypeScript can do this. It was done already halfway-well 6 years ago with way fewer features in TypeScript (https://github.com/brianc/node-sql/blob/master/lib/types.d.t...) - today we can do much better.

FWIW I've given up the SQL query builder route and instead went with GraphQL query builder / Hasura route (https://typed-graphql-builder.spion.dev/).

Its technically codegen but only from the schema; queries are fully typed on-the-fly.

I want to derive a Model type with methods from a Row type so I can write something like `const openDiscussions = blockModel.getContent({ where: c => c.getType() !== “page” }).andRecurse().getDiscussions({ where: d => !d.getResolvedAt() })` given an input row type like `type BlockRow = { id: BlockId, type: “page” | “text”, content?: BlockId[], discussions?: DiscussionId[] }; type DiscussionRow = { id: DiscussionId, resolvedAt?: Date }`

You cannot use the lambdas operators combo without a preprocessor (like ttypescript) mainly because closure captures cannot be accessed from the function AST.

You will also need to use a tuntime dsl to describe the schema, and then derive both the model classes and the model types from it (just because runtime parts cannot be derived from types)

Other than that, the rest is very possible.

> You will also need to use a tuntime dsl to describe the schema

This is kind of the whole ballgame right? How would you do any of this with types if the schema isn't defined until runtime?

I meant that since types are erased at compile-time in typescript, its much easier to make a schema description DSL that would then serve as the base to derive both the ORM DSL (runtime objects) and the types (compile time checks). If you go the types-first route, you will be forced to use proxies which can be more painful and constraining.

The popular choice here is to use classes plus decorators, but its not the only choice

https://typegraphql.com/docs/introduction.html#what

The convenient bit here is that classes already have both a type and a runtime representation, and decorators are also available to attach any extra metadata necessary to that runtime representation

The other benefit of codegen in situations like this (a company wide schema) is the potential for multi language bindings. Entgo demonstrates this by bridging a Go ORM, SQL migrations and a GraphQl schema.

Interesting reading along the same lines is Language Oriented Programming (1994) http://www.gkc.org.uk/martin/papers/middle-out-t.pdf

> But we still do it, because in the end you learn it once and have a powerful tool that makes you more productive for the rest of your life and allow you do to things that you otherwise would never have been able to.

Your own words are apt: I disagree here. And I also don't like your framing.

> Go has more inexperierenced developers using it compared to other languages. But over time, they will turn into experienced developers. They learn more concept and want to become more productive. So they will gradually ask for more and more features until Go is nothing like it was before.

Ignoring the rest of your comment, as well as the possibly snide remark regarding the experience of engineers who use Go, I believe this neither reflects reality nor the reality I wish to participate in.

Go eschews complexity, and is particularly paranoid of complexity sneaking in under the guise of convenience. Productivity is rarely hindered by having to hand-write a for-loop or some boilerplate to handle errors.

Go is designed to be easy to read and easy to write, to be comprehensible by both green and sage developers, and most importantly maintainable. This means preferring explicit behavior over implicit magic. The toolchain prioritizes fast compilation times over aggressive optimization.

I don't write Go to be productive (but I am very productive). I write Go so everyone can be productive. I don't want anyone spending time unraveling your life's work masterpiece. I don't want anyone to ever have to understand Scala or Haskell.

It's cool that you feel it necessary to reach for more powerful, expressive language constructs to be productive. I do not.