Question / comment / maybe bug report

I was surprised that this didn’t work:

def validate_not_cleared(changeset) do
  validate_change(changeset, :some_field, fn _, _new_value ->
    if changeset.data.some_field do
      [some_field: "can't be changed once set"]
    else
      []
    end
  end)
end

It successfully adds an error to the changeset when some_field is being changed to a non-nil value, but if the changeset is clearing the value it doesn’t run:

%SomeStruct{some_field: "foo"}
|> Ecto.Changeset.change(%{foo: nil})
|> validate_not_cleared()
# gives valid?: true

Here’s what I’m using instead, for the moment:

def validate_not_cleared(changeset) do
  case fetch_change(changeset, :some_field) do
    {:ok, _new_value} ->
      if changeset.data.some_field do
        add_error(changeset, :some_field, "can't be changed once set")
      else
        changeset
      end

    :error ->
      changeset
  end
end

EDIT: this is also distinct from how Ecto.Changeset.update_change works, which uses fetch_change…