September 27, 2016

Replace Callbacks with Ecto.Multi

We all have logic in our applications like this:

  • When a user is created, send a notification to an admin
  • When a post is deleted, remove it from the search cache
  • When a password is reset, log out that user’s active sessions

These side effects need to be predictable and reliable. Often, they’re some of the key business logic of the whole application.

Many Rails applications store this kind of logic in their models, using ActiveRecord callbacks. However, since Ecto 2.0 removed callbacks, this is not a viable (or desirable) approach in Elixir.

Instead, Elixir follows a simple rule:

If you have shared business logic, create a shared module.

One of the ways you can do this is by building simple “Service” modules around Ecto.Multi.1 Since business logic usually revolves around an Ecto.Schema, I often name service modules after the schema they interact with, such as UserService or PostService.

defmodule MyApp.UserService do
  alias MyApp.{Mailer, NotificationEmail}
  alias Ecto.Multi

  def insert(user_changeset) do
    Multi.new
    |> Multi.insert(:user, user_changeset)
    |> Multi.run(:notify_admin, &notify_admin/1)
  end

  defp notify_admin(%{user: user}) do
    user
    |> NotificationEmail.notify_admin
    |> Mailer.deliver
  end
end

I often name the service module functions after the operations they enhance, such as insert, update, or delete. Of course, there’s nothing preventing you from using other names for more custom functionality.

As you can see in the example above, Ecto.Multi allows us to encapsulate our business logic into a struct which can then be passed around, appended to, and run as a database transaction.

We can add arbitrary functions to the Multi operation, such as notify_admin/1 which will cause the transaction to rollback if they return {:error, ...}. We can then call the whole thing with Repo.transaction/1:

result =
  %User{}
  |> User.changeset(user_params)
  |> UserService.insert
  |> Repo.transaction

case result do
  {:ok, %{user: user}} ->
    # handle success
  {:error, :notify_admin, _failed_value, _changes_so_far} ->
    # handle the case that the email didn't send
  {:error, _failed_operation, _failed_value, _changes_so_far} ->
    # handle the generic error case
end

The error states are very detailed and allow you to be as specific or general as you want to be with your error handling.

Since each service function returns an Ecto.Multi, services can be chained together to execute as a single database transaction: 2

user_operation = UserService.insert(user)
log_operation = LogService.insert(user)

user_operation
|> Multi.append(log_operation)
|> Repo.transaction

Ecto.Multi structs are also inspectable using Ecto.Multi.to_list/1. This makes them pretty easy to unit test.

user
|> UserService.insert
|> Multi.to_list
# => [
#   {:user, {:insert, user, []}}
#   {:notify_admin, {:run, fun, []}}
# ]

There are at least several other advantages to having a service layer built around Ecto.Multi:

  • Schemas can contain only pure functions, with no ActiveRecord bloat. All your external side effects can be kept in your service layer.
  • Side-effects are opt-in, making testing much easier. Don’t want all the side effects? Just Repo.insert! your data instead.
  • Transactions can be automatically rolled back on failure, leaving no inconsistent mess behind. 3

Those are some pretty nice features for such a simple abstraction. Even so, you don’t have to use Ecto.Multi for everything, as we’ll see in the next section.

A Guide for Rails Developers

For before_create or before_update logic, update your schema’s changeset function(s) rather than using a service. Usually, this type of business logic only affects data, so it’s a natural fit for your changeset function.

For example, if we want to automatically promote or demote articles to our home page based on a like_count field, we could do that with changeset functions like so:

def changeset(schema, params) do
  schema
  |> cast(params)
  |> promote_popular
end

defp promote_popular(changeset) do
  if get_field(changeset, :like_count) > 1000 do
    put_change(changeset, :promoted, true)
  else
    put_change(changeset, :promoted, false)
  end
end

Or if we wanted to automatically encrypt a user’s password:

def changeset(schema, params) do
  schema
  |> cast(params)
  |> validate_required([:password])
  |> validate_confirmation(:password)
  |> encrypt_password
end

defp encrypt_password(changeset) do
  password = get_change(changeset, :password)

  if password do
    encrypted = Comeonin.Bcrypt.hashpwsalt(password)
    changeset
    |> put_change(:encrypted_password, encrypted)
    |> delete_change(:password)
    |> delete_change(:password_confirmation)
  else
    changeset
  end
end

For after_create or after_update logic, use an Ecto.Multi service. This logic usually has to do with other database tables or external side effects, and therefore is a good fit for a service module. The documentation provides a good example:

defmodule UserService do
  alias Ecto.Multi
  import Ecto

  def password_reset(account, params) do
    Multi.new
    |> Multi.update(:account, Account.password_reset_changeset(account, params))
    |> Multi.insert(:log, Log.password_reset_changeset(account, params))
    |> Multi.delete_all(:sessions, assoc(account, :sessions))
  end
end

The only important convention for service modules is that they return an Ecto.Multi structs: what you name the modules and functions themselves is much less important. Do what makes sense for your application.

Conclusion

I’ve been looking for a viable alternative to ActiveRecord callbacks and their associated bloat for a long time, and I finally feel like I’ve found it with Ecto.Multi. Be sure to check out its documentation, and if you’re interested, I’ve also covered its API in more detail in this episode of LearnPhoenix.tv.

  1. This is only one available option. You could also share business logic with a Plug or a plain module, depending on your application.

  2. As long as every operation has a unique name.

  3. This automatic rollback behavior is also opt-in. If you don’t want one of your operations to ever cause a rollback, just ensure it always returns {:ok, ...}.

Did You Enjoy This Post?

Comments