MRO's and you; how the distinction between C3 and DFS changed my life

Recently I fixed DBIx::Class::Helpers so that each helper would have a base class. This is actually something that ribasushi had been sorta hounding me to do for years but I could never figure out the case where it mattered. The reason I finally made the change was because a user ran into an issue that fixing the base class actually resolved. Unfortunately I neither documented what it was nor wrote a test. Such is life, live and learn.

So today I was documenting one of our products that I know all the guts of and, before I did anything, I ran the tests. A non-trivial number of them failed, something like 40% of them. I was astounded; nothing has changed on this thing in months. As you might be able to guess, the commit mentioned above caused the problem.

First off, let me illustrate the problem. You care about the longevity and overall quality of your codebase, so you carefully create a base class for your DBIx::Class resultsets that might look something like this:

package MyCompany::Schema::ResultSet;

use strict;
use warnings;

use parent 'DBIx::Class::ResultSet';

__PACKAGE__->load_components('Helper::ResultSet::IgnoreWantarray');

1;

And then in your project you do something like this:

package TeaTime::Schema::ResultSet;

use strict;
use warnings;

use parent 'MyCompany::Schema::ResultSet';

__PACKAGE__->load_components(qw(
   Helper::ResultSet::IgnoreWantarray
   Helper::ResultSet::Errors
));

1;

Ok, so it doesn’t do much, but at the very least it guards you against some pretty typical mistakes while using your DBIx::Class resultsets.

Next, you make a resultset class, maybe something like this:

package TeaTime::Schema::ResultSet::Milk;

use parent 'TeaTime::Schema::ResultSet';

sub in_order {
   shift->search(undef, { order_by => { -desc => 'when_expires'} })
}

1;

Now you have a handy method to get your results in a sensible order. Great!

So just to be clear, here is your inheritance hierarchy:

          ...
           |
         DBIC
           |
DBIC::ResultSet-*-*-*-*-*-*-*\-*-*-*-*-*-*-*-*\
           |                 |                 |
           |            ::IgnoreWantarray      *
           |    /------------/                 |
MC::Schema::ResultSet        |              ::Errors
           |                 |                   |
TT::Schema::ResultSet--------/-------------------/
           |
TT::Schema::ResultSet::Milk

Note the line with astrisks in it; that line was added by the commit mentioned in the first paragraph. Here’s where it gets scary. The linear isa of our ::ResultSet is:

  "TeaTime::Schema::ResultSet",
  "DBIx::Class::Helper::ResultSet::Errors",
  "MyCompany::Schema::ResultSet",
  "DBIx::Class::Helper::ResultSet::IgnoreWantarray",
  "DBIx::Class::ResultSet",
  "DBIx::Class",
  "DBIx::Class::Componentised",
  "Class::C3::Componentised",
  "DBIx::Class::AccessorGroup",
  "Class::Accessor::Grouped"

Perfect, that’s what we expect!

What about ::Milk?

  "TeaTime::Schema::ResultSet::Milk",
  "TeaTime::Schema::ResultSet",
  "DBIx::Class::Helper::ResultSet::Errors",
  "DBIx::Class::ResultSet",
  "DBIx::Class",
  "DBIx::Class::Componentised",
  "Class::C3::Componentised",
  "DBIx::Class::AccessorGroup",
  "Class::Accessor::Grouped",
  "MyCompany::Schema::ResultSet",
  "DBIx::Class::Helper::ResultSet::IgnoreWantarray"

WOAH! Why is ::IgnoreWantarray all the way at the bottom?

Well it turns out that ::Milk is using the dfs MRO. Basically what that means is that it picks a path, in this case the one via ::Errors, and traverses down (dfs stands for depth first search, afterall) instead of going side to side. Fundamentally the c3 resolution solves this problem.

Here’s where it gets even weirder. If I remove the parent classes from ::Errors and ::IgnoreWantarray both c3 and dfs work fine. The reason for that is that if there is no parent, dfs won’t traverse to it, obviously.

At some point I’d like to make a little web app that allows the user to create a hierarchy and see the different orderings based on algorithms, because I still don’t quite understand why the above hierarchy is needed (there needs to be two layers of base classes, to be clear.)

🔗 What’s wrong with my brain?

So why did I make this mistake? To be clear, this code has worked for years at MTSI, and suddenly, a relatively small change to make things “more correct” broke nearly everything! Well the problem was that I assumed that if your parent class used c3 you would too. That’s not the case at all. The reason the parent classes above use c3 is because they call load_components. Many of you may not have this problem because you are using either Moo or Moose in your resultsets. So basically, I assumed that c3 would be turned on for me, and it’s not.

🔗 What’s the next step?

This is a big problem. It needs to be trivially easy to do the right thing because there is no doubt in my mind that other people have made the same mistake. My plan is to make a DBIx::Class::Candy::ResultSet which will do the same type of things as DBIx::Class::Candy, but also set the c3 mro. I will also do some research and see if the original DBIx::Class::Candy needs to do the same thing.

In the meantime, beware of DBIx::Class::Helpers versions 2.025002 and 2.025003. I will have a short post for when the resultset sugar is available.

Posted Tue, Apr 14, 2015

If you're interested in being notified when new posts are published, you can subscribe here; you'll get an email once a week at the most.