Sharing External ActiveRecord Connections

I had a problem with a deployed Rails application today that Mike Clark helped me track down.

It started after Mike and I migrated a Rails app to a new, zippier server. Everything looked great (and the app ran a lot faster). However, after a while, I started getting problems with a database. To explain what went wrong, I need to chat a little about the application.

This piece of code is an attempt to migrate functionality away from a particularly nasty set of old CGIs, web applications, bailing wire, chewing gum, and other historical crud. There’s a lot of existing functionality, and I decided not to try to port it all across to Rails on day one. Instead, I gave the Rails application its own database to play with, and gave it access to the old legacy database for when it needed to look up stuff in the existing system. To do that, I created a set of ActiveRecord classes for the legacy tables, and explicitly connected them to the legacy database, rather than the default Rails application database. A typical legacy model class looked smething like this:

  class LegacyOrder < ActiveRecord::Base
    set_table_name  "orders"
    set_primary_key "o_id"

    establish_connection "legacy_#{RAILS_ENV}"
    ...
  end

(The ”legacy_#{RAILS_ENV}” stuff lets me have development, test, and production versions of the legacy database in my database.yml file.)

So far, so good. I had four different legacy tables I mapped this way, and everything worked fine.

However, after a while, other applications accessing the legacy database started failing, claiming they couldn’t connect. Sure enough, it turned out that there were hundreds of active connections to that database: the database wasn’t going to accept any more.

The problem was that I was creating way too many connections to the legacy database. Each spawned Rails fcgi process established a separate connection to the legacy database for each legacy ActiveRecord class. That factor of four killed us.

But fixing it also raised an interesting question: in my original code, I established a separate connection for each legacy table in each running Rails process. This was clearly dumb. But what’s the best way of sharing a single connection between all of the legacy model objects in a Rails instance?

Some chatting on #rails-core came up wih three suggestions:

Create the connection once (perhaps in one of the environment files), and then assign it to each of the legacy AR classes:

  legacy_connection = ActiveRecord::Base.mysql_connection(...)

  LegacyOrder.connection = legacy_connection
  LegacyLineItem.connection = legacy_connection
  ...

Establish the connection in one model, then use it in the others

 class LegacyOrder < ActiveRecord::Base
   establish_connection ...
 end

 class LegacyLineItem < ActiveRecord::Base
   connection = LegacyOrder.connection
 end

Somehow make all the legacy models share a common parent, and have that parent own the connection.

Options (1) and (2) seemed messy: they added a bunch of coupling to the code where it wasn’t really justified. Option (3) seemed like the way to go. The only problem was that no one on #rails-core was sure it would work…

It turns out it does.

  class LegacyBase < ActiveRecord::Base
    establish_connection ...
  end

  class LegacyOrder < LegacyBase
     ...
  end

  class LegacyLineItem < LegacyBase
    ...
  end

The parent class creates a single connection to the legacy database, and all the child classes share it. The reason it works is some smarts in Rails.

It turns out that Rails does just about everything lazily. That includes connecting to databases and reflecting on tables to extract the schema (needed to build the internals of the models). This improves performance, but it also makes this hack possible. In general, you’d expect the LegacyBase class to map to a database table called legacy_base. It would, if we ever tried to use it to access data. But because we don’t, and because Rails only reflects on the table the first time a data access occurs, we can safely create an ActiveRecord class with no underlying database table.

This scheme lets me specify the legacy connection once, and share that connection between all my legacy models. It’s tidy, expressive, and saves resources.