I'm reluctant to say that it is impossible, because there are so many clever things that have been done with Haskell's type system that I can barely understand how to use, much less how to design those systems (like Servant's typed URLs). But from a naive user point of view, I don't see how its remotely possible to develop Ecto-style (or ActiveRecord style) associations with selective eager pre-loading, and typed validations against arbitrary trees of data with such a strictly typed system. And I think I've looked at every SQL library for Haskell and none of them really approach it.

I'm sure its possible to build something that compiles, but it would be so awful to use that you are better off with hand-written SQL where you need it, and N+1 stuff where you can get away with it.

Rel8 allows you to return trees of data:

https://hackage.haskell.org/package/rel8-1.3.0.0/docs/Rel8.h...

That's awesome, thanks for linking it to me. I don't think I have seen this library before, looks like its fairly new. I have to admit from the signature I can't tell if I can nest these arbitrarily deep, it looks like I can?

Arbitrary nesting is supported

    module Main where
    
    import Rel8
    import Data.Int
    import Hasql.Statement
    import Hasql.Session
    import Hasql.Connection
    import Database.Postgres.Temp
    
    main = Database.Postgres.Temp.with $ \db -> do
      Right conn <- acquire (toConnectionString db)
    
      flip run conn $ statement () $ select $ do
        q1 <- many $ do
          r1 <- many (values [1 , 2 :: Expr Int16])
          r2 <- many (values [3 , 4 :: Expr Int16])
    
          values [(litExpr "one", r1), (litExpr "two", r2)]
    
        q2 <- many $ do
          r1 <- many (values [10 , 20 :: Expr Int16])
          r2 <- many (values [30 , 40 :: Expr Int16])
    
          values [(litExpr "ten", r1), (litExpr "twenty", r2)]
    
        values [(litExpr "X", q1), (litExpr "Y", q2)]
    

    > main
    Right (Right [("X",[("one",[1,2]),("two",[3,4])]),("Y",[("ten",[10,20]),("twenty",[30,40])])])

Having said that, Postgres supports JSON fine, and Opaleye and Rel8 support Postgres's JSON fine, so you can just use that if you want arbitrarily structured and nested stuff.

Rel8 is pretty new, but it's built on top of opaleye which is much older.

If I understand this correctly: you are saying that strict-typing prevents arbitrary depth record requests to be made to SQL because Haskell cannot represent an arbitrary tree of types...

I presume that the accessors in Haskell would be 1-1 with SQL language requests, but the problem is that SQL queries are not typed? Or are dynamically typed?

So, instead of a generic library for SQL queries, one would have to implement a library for a specific SQL DB structure?

There are typed query libraries, like esqueleto which give you well typed and flexible query semantics, my complaint with it is that it just returns a tuple of lists, with a separate list for each type, you cannot return the data structured in a tree.

So if you have a companies table joined to a contacts table, you'd get a two element tuple with a list of companies and a list of contacts, and then you have to iterate to do the joins and turn it into a JSON tree.

And there is no help at all for wanting to take in a tree of JSON data, and turn it into updates to existing records and inserts for new records.

You could have a library that to knows how to return a tree if you define a typeclass for the tree, but the only way I'd know to do it would be a separate tree data type for each result set. In Ecto you do have to define a structure that knows all the possible associations, but which associations you actually load in any given query is flexible, you can have associations that are not loaded eagerly for one query, and loaded eagerly in another. In Haskell I can't imagine doing that without two different tree structs, because there is nothing that would know how to take two differently typed result sets and make the same tree partially loaded.

> So if you have a companies table joined to a contacts table, you'd get a two element tuple with a list of companies and a list of contacts, and then you have to iterate to do the joins and turn it into a JSON tree.

Esqueleto has a function to avoid n+1's and return a map now:

https://hackage.haskell.org/package/esqueleto-3.5.3.0/docs/D...

https://www.foxhound.systems/blog/grouping-query-results-has...

> Haskell's works best when working with flat data structures

One if the things Haskell is best at is parsing and the resulting data structures are very nested.

What is an example data structure you think Haskell would struggle with? What language wouldn't struggle with that example?