Norm

Norm is a system for specifying the structure of data. It can be used for validation and for generation of data. Norm does not provide any set of predicates and instead allows you to re-use any of your existing validations.

import Norm

conform!(123, spec(is_integer() and &(&1 > 0)))
=> 123

conform!(-50, spec(is_integer() and &(&1 > 0)))
** (Norm.MismatchError) val: -50 fails: &(&1 > 0)
    (norm) lib/norm.ex:44: Norm.conform!/2

user_schema = schema(%{
  user: schema(%{
    name: spec(is_binary()),
    age: spec(is_integer() and &(&1 > 0))
  })
})

input = %{user: %{name: "chris", age: 30, email: "c@keathley.io"}

conform!(input, user_schema)
=> %{user: %{name: "chris", age: 30}}

user_schema
|> gen()
|> Enum.take(3)
=> [
  %{user: %{age: 0, name: ""}},
  %{user: %{age: 2, name: "x"}},
  %{user: %{age: -2, name: ""}}
]

Installation

Add norm to your list of dependencies in mix.exs. If you'd like to use Norm's generator capabilities then you'll also need to include StreamData as a dependency.

def deps do
  [
    {:stream_data, "~> 0.4"},
    {:norm, "~> 0.6"}
  ]
end

Validation and conforming values

Norm validates data by "conforming" the value to a specification. If the values don't conform then a list of errors is returned. There are 2 functions provided for this conform/2 and conform!/2. If you need to return a list of well defined errors then you should use conform/2. Otherwise conform!/2 is generally more useful. The input data is always passed as the 1st argument to conform so that calls to conform are easily chainable.

Predicates and specs

Norm does not provide a special set of predicates and instead allows you to convert any predicate into a spec with the spec/1 macro. Predicates can be composed together using the and and or keywords. You can also use anonymous functions to create specs.

spec(is_binary())
spec(is_integer() and &(&1 > 0))
spec(is_binary() and fn str -> String.length(str) > 0 end)

The data is always passed as the first argument to your predicate so you can use predicates with multiple values like so:

def greater?(x, y), do: x > y
conform!(10, spec(greater?(5)))
=> 10
conform!(3, spec(greater?(5)))
** (Norm.MismatchError) val: 3 fails: greater?(5)
    (norm) lib/norm.ex:44: Norm.conform!/2

Tuples and atoms

Atoms and tuples can be matched without needing to wrap them in a function.

:atom = conform!(:atom, :atom)

{1, "hello"} = conform!({1, "hello"}, {spec(is_integer()), spec(is_binary())})

conform!({1, 2}, {:one, :two})
** (Norm.MismatchError) val: 1 in: 0 fails: is not an atom.
val: 2 in: 1 fails: is not an atom.

Because Norm supports matching on bare tuples we can easily validate functions that return {:ok, term()} and {:error, term()} tuples.

# if User.get_name/1 succeeds it returns {:ok, binary()}
result = User.get_name(123)
{:ok, name} = conform!(result, {:ok, spec(is_binary())})

These specifications can be combined with one_of/1 to create union types.

result_spec = one_of([
  {:ok, spec(is_binary())},
  {:error, spec(fn _ -> true end)},
])

{:ok, "alice"} = conform!(User.get_name(123), result_spec)
{:error, "user does not exist"} = conform!(User.get_name(-42), result_spec)

Schemas

Norm provides a schema/1 function for specifying maps and structs:

user_schema = schema(%{
  user: schema(%{
    name: spec(is_binary()),
    age: spec(is_integer()),
  })
})

conform!(%{user: %{name: "chris", age: 31}}, user_schema)
=> %{user: %{name: "chris", age: 31}}

conform!(%{user: %{name: "chris", age: -31}}, user_schema)
** (Norm.MismatchError) in: :user/:age val: -31 fails: &(&1 > 0)
    (norm) lib/norm.ex:44: Norm.conform!/2

You can also create specs from structs:

defmodule User do
  defstruct [:name, :age]

  def s, do: schema(%__MODULE__{
      name: spec(is_binary()),
      age: spec(is_integer())
    }
end

This will ensure that the input is a User struct with the key that match the given specification. Its convention to provide a s() function in the module that defines the struct so that schema's can be shared throughout your system.

You don't need to provide specs for all the keys in your struct. Only the specced keys will be conformed. The remaining keys will be checked for presence.

defmodule User do
  defstruct [:name, :age]
end

conform!(%User{name: "chris"}, schema(%User{}))
=> %User{name: "chris", age: nil}

Key semantics

Atom and string keys are matched explicitly and there is no casting that occurs when conforming values. If you need to match on string keys you should specify your schema with string keys.

Schema's accomodate growth by disregarding any unspecified keys in the input map. This allows callers to start sending new data over time without coordination with the consuming function.

Selections

You may have noticed that there's no way to specify optional keys in a schema. This may seem like an oversite but its actually an intentional design decision. Whether a key should be present in a schema is determined by the call site and not by the schema itself. For instance think about the assigns in a plug conn. When are the assigns optional? It depends on where you are in the pipeline.

Schema's also force all keys to match at all times. This is generally useful as it limits your ability to introduce errors. But it also limits schema growth and turns changes that should be non-breaking into breaking changes.

In order to support both of these scenarios Norm provides the selection/2 function. selection/2 allows you to specify exactly the keys you require from a schema at the place where you require them.

user_schema = schema(%{
  user: schema(%{
    name: spec(is_binary()),
    age: spec(is_integer()),
  })
})
just_age = selection(user_schema, [user: [:age]])

conform!(%{user: %{name: "chris", age: 31}}, just_age)
=> %{user: %{age: 31}}

# Selection also disregards unspecified keys
conform!(%{user: %{name: "chris", age: 31, unspecified: nil}, other_stuff: :foo}, just_age)
=> %{user: %{age: 31}}

Patterns

Norm provides a way to specify alternative specs using the alt/1 function. This is useful when you need to support multiple schema's or multiple alternative specs.

create_event = schema(%{type: spec(&(&1 == :create))})
update_event = schema(%{type: spec(&(&1 == :update))})
event = alt(create: create_event, update: update_event)

conform!(%{type: :create}, event)
=> {:create, %{type: :create}}

conform!(%{type: :update}, event)
=> {:update, %{type: :update}}

conform!(%{type: :delete}, event)
** (Norm.MismatchError)
  in: :create/:type val: :delete fails: &(&1 == :create)
  in: :update/:type val: :delete fails: &(&1 == :update)

Generators

Along with validating that data conforms to a given specification, Norm can also use specificiations to generate examples of good data. These examples can then be used for property based testing, local development, seeding databases, or any other usecase.

user_schema = schema(%{
  user: schema(%{
    name: spec(is_binary()),
    age: spec(is_integer() and &(&1 > 0))
  })
})
conform!(%{user: %{name: "chris", age: 30}}, user_schema)
=> %{user: %{name: "chris", age: 30}}

user_schema
|> gen()
|> Enum.take(3)
=> [
  %{user: %{age: 0, name: ""}},
  %{user: %{age: 2, name: "x"}},
  %{user: %{age: -2, name: ""}}
]

Under the hood Norm uses StreamData for its data generation. This means you can use your specs in tests like so:

input_data = schema(%{"user" => schema(%{"name" => spec(is_binary())})})

property "users can update names" do
  check all input <- gen(input_data) do
    assert :ok == update_user(input)
  end
end

Built in generators

Norm will try to infer the generator to use from the predicate defined in spec. It looks specifically for the guard clauses used for primitive types in elixir. Not all of the built in guard clauses are supported yet. PRs are very welcome ;).

Guiding generators

You may have specs like spec(fn x -> rem(x, 2) == 0 end) which check to see that an integer is even or not. This generator expects integer values but there's no way for Norm to determine this. If you try to create a generator from this spec you'll get an error:

gen(spec(fn x -> rem(x, 2) == 0 end))
** (Norm.GeneratorError) Unable to create a generator for: fn x -> rem(x, 2) == 0 end
    (norm) lib/norm.ex:76: Norm.gen/1

You can guide Norm to the right generator by specifying a guard clause as the first predicate in a spec. If Norm can find the right generator then it will use any other predicates as filters in the generator.

Enum.take(gen(spec(is_integer() and fn x -> rem(x, 2) == 0 end)), 5)
[0, -2, 2, 0, 4]

But its also possible to create filters that are too specific such as this:

gen(spec(is_binary() and &(&1 =~ ~r/foobarbaz/)))

Norm can determine the generators to use however its incredibly unlikely that Norm will be able to generate data that matches the filter. After 25 consequtive unseccesful attempts to generate a good value Norm (StreamData under the hood) will return an error. In these scenarios we can create a custom generator.

Overriding generators

You'll often need to guide your generators into the interesting parts of the state space so that you can easily find bugs. That means you'll want to tweak and control your generators. Norm provides an escape hatch for creating your own generators with the with_gen/2 function:

age = spec(is_integer() and &(&1 >= 0))
reasonable_ages = with_gen(age, StreamData.integer(0..105))

Because gen/1 returns a StreamData generator you can compose your generators with other StreamData functions:

age = spec(is_integer() and &(&1 >= 0))
StreamData.frequencies([
  {3, gen(age)},
  {1, StreamData.binary()},
])

gen(age) |> StreamData.map(&Integer.to_string/1) |> Enum.take(5)
["1", "1", "3", "4", "1"]

This allows you to compose generators however you need to while keeping your generation co-located with the specification of the data.

Should I use this?

Norm is still early in its life so there may be some rough edges. But we're actively using this at my current company (Bleacher Report) and working to make improvements.

Contributing and TODOS

Norm is being actively worked on. Any contributions are very welcome. Here is a limited set of ideas that are coming soon.

Support generators for other primitive types (floats, etc.)
More streamlined specification of keyword lists.
selections shouldn't need a path if you just want to match all the keys in the schema
Support "sets" of literal values
specs for functions and anonymous functions
easier way to do dispatch based on schema keys

tylersamples / norm