config/initializers/multi_db doesn't work

[nate-at-gusto]

It doesn't insert the database selector middleware, because the railtie which inserts it has already run by the time the initializer is run.

Repro

See repo here.

bin/rails test test/integration/replica_test.rb 

[ghiculescu]

Related: #44996

If that PR isn't accepted, then config.after_initializer might fix it, here:

rails/activerecord/lib/active_record/railtie.rb

Line 98 in b24c441

if options = config.active_record.database_selector

In the meantime, moving the config to application.rb (like it was before #43796) definitely works.

[eileencodes]

I know that we discussed in chat that order of config initializers has been a problem for awhile but I am 100% sure I tested this in an app when I implemented it. This leads me to believe something else changed between 6.1 and 7.0 that is causing all these initializers to now not work. 😕

[ghiculescu]

This leads me to believe something else changed between 6.1 and 7.0 that is causing all these initializers to now not work. 😕

I haven't dug into it much but I have a feeling #42499 was working around the same regression.

[ghiculescu]

Trying to bisect this, but not having much luck yet.

railties/test/application/configuration_test.rb:

    test "can get multi-db middleware from initializer" do
      app_file "config/initializers/multi_db.rb", <<-RUBY
        Rails.application.configure do
          config.active_record.database_selector = { delay: 2.seconds }
        end
      RUBY

      app "development"

      assert_includes Rails.application.config.middleware.map(&:name), "ActiveRecord::Middleware::DatabaseSelector"
    end

    test "can get multi-db middleware from application.rb" do
      add_to_config <<-RUBY
        config.active_record.database_selector = { delay: 2.seconds }
      RUBY

      app "development"

      assert_includes Rails.application.config.middleware.map(&:name), "ActiveRecord::Middleware::DatabaseSelector"
    end

On both main and 6-1-stable the first test fails and the second test passes. I'm having some trouble getting 6-0-stable to build atm so can't confirm if it works there.

[ghiculescu]

I did some poking around. I don't know if this helps at all but i want to brain dump so I can wrap up for the day. The main interesting thing is that the order that railties get evaluated in is defined here

rails/railties/lib/rails/application.rb

Line 558 in 55cee03

def railties_initializers(current) # :nodoc:

This is the order on Rails 6.1, on 7 it's basically the same: ["I18n::Railtie", "ActiveSupport::Railtie", "ActionDispatch::Railtie", "ActiveModel::Railtie", "ActionController::Railtie", "ActiveRecord::Railtie", "GlobalID::Railtie", "ActiveJob::Railtie", "ActionMailer::Railtie", "Rails::TestUnitRailtie", "Sprockets::Railtie", "ActionView::Railtie", "ActiveStorage::Engine", "ActionCable::Engine", "ActionMailbox::Engine", "ActionText::Engine", "Turbolinks::Engine", "Webpacker::Engine", "AppTemplate::Application"]

So all the railties, and then all the engines, as defined here.

https://gist.github.com/ghiculescu/8d36b09402daff400d3786f89964c7ca lists each railtie/engine and which initializers it adds. My debug code was

    def railties_initializers(current) #:nodoc:
      initializers = []
      puts "railties is #{railties.to_a.map(&:class).map(&:name)}"
      puts "ordered_railties is #{ordered_railties.reverse.flatten.map(&:class).map(&:name)}"
      ordered_railties.reverse.flatten.each do |r|
        puts "r is #{r}"
        if r == self
          puts "adding #{current.map(&:name)}"
          initializers += current
        else
          puts "adding #{r.initializers.map(&:name)}"
          initializers += r.initializers
        end
      end
      initializers
    end

What's interesting is that load_config_initializers is not explicitly called by any of the railties. It is called basically as a side effect of all the engines. As a result it is called relatively late.

It seems to me like it should probably be called at a specified point, not randomly halfway through the loading of the frameworks.

[eileencodes]

I spent a long time on this today trying to determine when and how certain configs broke and the only thing I've come up with is that certain config initializations have broken since at least 6.0 and in Ruby 2.6 to to 3.0. This initializer predates 6.0 so I was testing with others that have been around for awhile.

For example calling Rails.application.config.active_record.use_schema_cache_dump = false in 6.1 and 6.0 raises an error that use_schema_cache_dump is undefined. I also see this if I set up the multi_db initializer - database_selector is undefined.

I can see why this is happening, the initializer in the Railtie runs before the initializer is loaded from config. options is never set so nothing else is set either. I attempted to use after_initialize with the Railtie and that didn't improve the situation.

I think the best course going forward is to drop the config.active_record.database_selector initializer that installs the middleware in the Railtie, and install it directly with the config initializer. PR incoming.

[eileencodes]

I've come to the conclusion that no one noticed before because we were using initializers less for this kind of thing. That's the only reasonable explanation I can come up with since it doesn't look like anything changed. That or this particular issue isn't related to the other issues we're seeing.

[eileencodes]

This is what I've come up with as a fix #45353

It has the unfortunate side effect that applications using these in places where they do work (application.rb) will still throw a deprecation warning.