Builder API for Creating Spark DSLs

[erikareads]

Is your feature request related to a problem? Please describe. Now, to create a DSL with Spark, you need to manually manipulate Entity and Section structs.

Describe the solution you'd like I propose a builder API that allows the specification of all the information that a DSL needs procedurally, leaning heavily on the pipe operator.

Describe alternatives you've considered A meta DSL would be another way to do this. As @jimsynz and I discussed in #47, a new API gives us the opportunity to improve the usability and understandability of Spark. Including radical changes that don't need to be backwards compatible, since the API is new. A Builder API will be easier to incrementally design and can serve as the backbone of what we do with the "Code as Data" from the meta-DSL.

Express the feature either with a change to resource syntax, or with a change to the resource interface

For example

entity = Spark.Builder.Entity.new(:my_entity, MyEntity)
  |> Spark.Builder.Entity.field(Spark.Builder.Field.new(...))
  |> Spark.Builder.Entity.field(...)
  |> Spark.Builder.Entity.subentity(...)
  ...

With builder API modules and functions for all of the data needed for a complete DSL.

[jimsynz]

I can't wait to see this @erikareads

[jimsynz]

An approach I've seen and I like for building nested structures, which you can take or leave as you please:

entity = Builder.Entity.new(:my_entity, MyEntity)
  |> Builder.Entity.field(name, fn field ->
    field
    |> Builder.Field.type(...)
    |> Builder.Field.default(nil)
  end)
  |> Builder.Entity.field(...)
  ...

[erikareads]

My plan was to use new at each layer:

entity = Builder.Entity.new(:my_entity, MyEntity)
  |> Builder.Entity.required_field(
    Builder.Field.new(:field_name, Builder.Type.any())
    |> Builder.Field.default(...)
  )
  |> Builder.Entity.field(...)
  ...

Either way, Field will need both a name and a type as part of its constructor:

entity = Builder.Entity.new(:my_entity, MyEntity)
  |> Builder.Entity.required_field(:field_name, Builder.Type.any(), fn field ->
    field
    |> Builder.Field.default(...)
  end)
  |> Builder.Entity.field(...)
  ...

I'll need to think about what that looks like for a complex type specification.

[erikareads]

Perhaps a combination, the new constructor of the Field will take an optional anonymous function:

entity = Builder.Entity.new(:my_entity, MyEntity)
  |> Builder.Entity.required_field(
    Builder.Field.new(:field_name, Builder.Type.type(
      Builder.Type.keyword_list(Builder.Type.string()),
      fn type ->
        type
        |> Builder.Type.or(
            Builder.Type.tuple(
              [Builder.Type.string(), Builder.Type.string()]
            )
           )
        |> Builder.Type.or(Builder.Type.string())
      end),
      fn field ->
        field
        |> Builder.Field.default(...)
        |> Builder.Field.documentation(...)
    end)
   )
   ...

[erikareads]

This opens the door for:

entity = Builder.Entity.new(:my_entity, MyEntity, fn entity ->
  ...
end)

Which is relevant since entities will be nested in Sections, sections nested in the top-level DSL.

[erikareads]

What are all the ways to fill fields in a targeted Entity struct?

I'm thinking about cleaning up the interface by forcing a type_field() convention.

For example:

Builder.Entity.required_field(entity, name, type, fn field -> ... end)

Builder.Entity.optional_field(entity, name, type, fn field -> ... end)

Builder.Entity.subentity_field(entity, subentity, relationship: :one_to_one | :one_to_many) 

# Alternatively:

Builder.Entity.subentity_field(entity, :one_to_one, subentity, fn subentity -> ... end)

Builder.Entity.recursive_field(entity, recursive_as)

Builder.Entity.preset_field(entity, literal)

# Alternatively:

Builder.Entity.required_field(entity, name, type, fn field ->
  field
  |> Builder.Field.default(literal)
end)

# Some variation of:

Builder.Field.argument() # Adds a field's name to the args list

With this arrangement, we can infer the schema directly while using a consistent interface.

[erikareads]

Looking through the fields of https://hexdocs.pm/spark/Spark.Dsl.Entity.html#t:t/0 that I don't understand:

• identifier: Is this an attribute on a field, or a field type unto itself?
• imports: What does this do?
• links: What does this do?
• modules: What does this do?
• no_depend_modules: What does this do?

[erikareads]

Do we have any invalid states represented between hide, identifier, deprecated, and auto_set_fields? Can a field have all of those at the same time?

[erikareads]

Looking through the fields of https://hexdocs.pm/spark/Spark.Dsl.Entity.html#t:t/0 that I don't understand:

* `identifier`: Is this an attribute on a field, or a field type unto itself?

* `imports`: What does this do?

* `links`: What does this do?

* `modules`: What does this do?

* `no_depend_modules`: What does this do?

Answering my own questions:

identifier is its own kind of field, with special logic for unique identifying every created entity from that DSL macro. I plan to handle it as Builder.Entity.identifier_field(entity, name)

imports This is a list of modules that is auto-imported when the given entity or section is brought into scope. For example, it's used by Ash.Resource to bring in helper functions: https://github.com/ash-project/ash/blob/main/lib/ash/resource/validation/builtins.ex. I plan to handle this as Builder.Entity.import(entity, Module1)

links does nothing, so far as I can tell. It is ignored by Spark.

modules and no_depend_modules are both special cases of the :module type for a field with special handling for expansion into a fully qualified module name from an alias. In the latter case with some workarounds avoiding adding a dependency on the named module. Since they're special handling for a type of field, I plan to handle them internally when a field is declared as Builder.Type.module().

[erikareads]

Question: Is there ever a case where a field_name can represent both a schema and a subentity? Is this something we would like to allow?

That is, is subentity_field a special case of required_field or optional_field, or a distinct case that is neither of the two?

[joshprice]

Mentioned this briefly to @jimsynz earlier. What's stopping us defining a Spark DSL in a Spark DSL?

Agree that it's a very strange feeling building a DSL in a struct when what you're trying to give your DSL users is all the bells and whistles you need to do that. I don't think a pipeline approach as described here goes far enough from struct editing, still seems less readable than a full blown DSL.

[joshprice]

Just answered my own question by immediately discovering #47 and #48 after I wrote the above comment. Is this issue still relevant?

[zachdaniel]

The ultimate goal is to support leveraging these structures for runtime things (like runtime dynamic resourcs), so I think the builder api is still very relevant.