rails5 can't access session in ActionDispatch::IntegrationTest

[rubys]

Reproduction scenario:

rm -rf throwaway
rails new throwaway 
cd throwaway
rails generate scaffold Product name
rails db:migrate
rails test
sed -i -e '/patch product/ i\
    session[:userid] = 0\
' test/controllers/products_controller_test.rb
rails test
BACKTRACE=1 rails test

Test results: throwaway.txt

[this is with rails master]

Note: I reran rails test with BACKTRACE set as the actual error is inside action_dispatch/testing/test_process.

Example with similar code working in Rails 4.2:

Code: http://intertwingly.net/projects/AWDwR4/checkdepot-215-42/section-10.4.html#cmd3
Test: http://intertwingly.net/projects/AWDwR4/checkdepot-215-42/section-10.4.html#cmd8

Results for same example in Rails 5 (master):

http://intertwingly.net/projects/AWDwR4/checkdepot/section-10.4.html#cmd8

[dhh]

The session is an internal structure for the controller. Integration test is more black box than controller tests were. I don't think we should be testing internals of the controller like this. I would rework the test to test something visible in the UI instead.

[rubys]

I fully understand removing access to low level details. But in this case, that just exposes what is the real problem here: http://api.rubyonrails.org/classes/ActionDispatch/Integration/RequestHelpers.html is woefully deficient. Don't get me wrong, it is fine if you are testing a single controller action in isolation. But once you want to test anything that involves session logic in your controller, you end up needing things like this: https://github.com/jnicklas/capybara#clicking-links-and-buttons and https://github.com/jnicklas/capybara#interacting-with-forms.

I don't think that these methods would be hard to write (Rails already has access to response.body and css_selector logic). I don't think they could make the current Rails 5 schedule. I don't think that access to a session should be removed until the higher level functions are in place.

Looking at the problem I'm seeing, it appears that the issue is that the session isn't created yet. Doing an unrelated action (like a get) would create a session.

[dhh]

We tested all of Basecamp 3 with controller tests driven by ActionDispatch::IntegrationTest, and I didn't find a need for capybara-style button clickers etc. I think it'd be nice to offer this as an additional level-up, but it's not a blocker in my book. If that's something you'd like to take a swing at, please do! Would make a great addition to Rails 5.1.

[cannikin]

@dhh Can you explain your technique for this? Do you actually post to your login page at the top of your single action test in the controller test, then commence with testing the action? Something like:

test "should get new" do
  user = users(:john)
  post login_url, params: { email: user.email, password: user.password }

  get new_post_url
  assert_response :success
end

Sorry to bug you but I haven't found any docs in the wider world that explain this new recommended controller testing paradigm.

[dhh]

That's exactly it. We've wrapped that into a helper method that we usually then use in the setup of a controller. So:

  # helper method
  def sign_in_as(name)
    post login_url, params: { sig: users(name).perishable_signature )
  end

class SomeControllerTest
  setup { sign_in_as 'david' }

  test 'the truth' do
  ..

[cannikin]

Thank you sir! I'm liking the feel of these new controller-integration hybrid tests so far.

[nikolaokonesh]

@dhh Hi
how to test omniauth "should get new" do ? and "create"...
` helper_method :authenticate_user!

[ream88]

If you still need this (I did 😕), this is the solution I came up for (which is just copied from ActionDispatch::Cookies):

def session_header(session)
  session = session.reverse_merge(session_id: SecureRandom.hex)

  key_generator = ActiveSupport::KeyGenerator.new(ENV.fetch('SECRET_KEY_BASE'), iterations: 1000)
  secret = key_generator.generate_key('encrypted cookie')
  sign_secret = key_generator.generate_key('signed encrypted cookie')
  encryptor = ActiveSupport::MessageEncryptor.new(secret, sign_secret, digest: 'SHA1', serializer: ActiveSupport::MessageEncryptor::NullSerializer)

  { 'Cookie' => "_your_app_session=#{encryptor.encrypt_and_sign(session.to_json)}" }
end

You can use it like so:

get '/needs/session', headers: session_header('some' => 'data')

This works with the current 5.0 versions, but can break any time for newer versions. Use at your own risk! The better way would be to follow @dhh's advise:

I would rework the test to test something visible in the UI instead.

[evgenyfadeev]

When authentication depends on an external service, sometimes it might not be available for proper integration tests; for instance, it might be behind some firewall.

Therefore - one has to either stub the authentication service or create a fake test login controller. The second will need to assess the session and that would run contrary to the idea that session is not available for the purposes of running integration tests. Stubbing the service is also extra work.

Perhaps adding an official login method, available from the test cases and not giving full access to session would be a good compromise. Otherwise, we just have an extra stumbling block in the way of testing.

[ream88]

2 years later, and my custom code above bit me when upgrading to 5.2. But thankfully I can now use the technique suggested by @dhh. So, thx again man! 😄