Gigitsu
How do you write Ecto schema typespecs with nullable fields?
Hi everyone ![]()
I’m trying to improve the typespecs in my application contexts, but I’m running into dialyzer errors when dealing with schemas that have fields that can be nil.
Here’s an example:
defmodule MyApp.User do
use Ecto.Schema
import Ecto.Changeset
@type t :: %__MODULE__{
id: integer(),
email: String.t(),
age: non_neg_integer()
}
schema "users" do
field :email, :string
field :age, :integer
end
def changeset(user, attrs) do
user |> cast(attrs, [:email, :age]) |> validate_required(:email)
end
end
defmodule MyApp.UsersContext do
alias MyApp.User
@spec change_user(user :: User.t(), attrs :: map()) :: Ecto.Changeset.t()
def change_user(%User{} = user, attrs \\ %{}) do
User.changeset(user, attrs)
end
end
defmodule MyApp.FakeController do
alias MyApp.User
def index() do
MyApp.UsersContext.change_user(%User{}, %{})
:ok
end
end
In this schema, there are no default values so %User{} will generate a struct where every field is nil.
Using this setup, dialyzer complains with:
The function call will not succeed.
MyApp.UsersContext.change_user(
%MyApp.User{
:__meta__ => %Ecto.Schema.Metadata{
:context => nil,
:prefix => nil,
:schema => MyApp.User,
:source => <<117, 115, 101, 114, 115>>,
:state => :built
},
:age => nil,
:email => nil,
:id => nil
},
%{}
)
breaks the contract
(user :: MyApp.User.t(), attrs :: map()) :: Ecto.Changeset.t()
To fix this error I have to explicitly put | nil in MyApp.User.t() type:
@type t :: %__MODULE__{
id: integer() | nil,
email: String.t() | nil,
age: non_neg_integer() | nil
}
But here are some things I’m unsure about:
- Is it idiomatic to be so explicit with
| nilfor every nullable field? - Is there a better or preferred way to declare the schema type (
@type t) that keeps it maintainable and readable, especially in large schemas?
Another possible solution I’ve found is to declare, somewhere in the codebase, a generic schema typespec:
@type schema_t(schema) :: %{
optional(atom) => any,
__struct__: schema,
__meta__: Ecto.Schema.Metadata.t(schema)
}
and use it in my specs:
@spec change_user(user :: schema_t(User), attrs :: map()) :: Ecto.Changeset.t()
Curious how others approach this.
Thanks!
Marked As Solved
benwilson512
The short answer is yes.
integer() is simply different from integer() | nil. Your code is using %User{} which absolutely does not have an id value yet, it hasn’t been saved to the DB, so it isn’t valid to claim it has an integer id in all cases.
If you had code which did user.id + 10 that’s going to raise an exception, and you WANT dialyzer to give you heads up about such things.
Also Liked
billylanchantin
Curious how others approach this.
There’s also the TypedEctoSchema library:
It auto-generates the typespec boiler plate for your Ecto schemas. It has sensible defaults, and you can use :: inline to override the typespec for a specific fields.
dogweather
TypedEctoSchema and TypedStruct are life-savers. They eliminate unnecessary verbosity and make schema definitions readable.
Gigitsu
I like this approach, and your struct_t resembles my idea of a schema_t type, though with less enforcement. Thank you.
I agree. I once wrote a typed_schema macro myself, but it started to feel like it was adding unnecessary complexity and non-standard conventions on top of Ecto, so I went back to manually crafted typespecs.
Gigitsu
@benwilson512 so being explicit is the only option here. I was hoping there’d be a shortcut for defining nullable fields. Thanks!
@billylanchantin thank you for the answer. I’ve seen some libraries (like the one you provided) used to help with this, but I’m trying to keep things vanilla Ecto if possible.
zorn
In some of my recent projects, I’ve preferred to model the schema type to be a description of a repo-sourced value, thus allowing me to type things like inserted_at: DateTime.t() instead of inserted_at: DateTime.t() | nil.
To allow for some functions which expected a non-repo-sourced value I would make a @type struct_t() :: %__MODULE__{}.
I can’t say I observe many other Elixir codebases doing this, and my quick round-the-room check with some other devs says this nuance was not on their radar.








