rjbs forgot what he was saying

not logged in (root) | by date | tagcloud | help | login

Email::Sender and Any::Moose

by rjbs, created 2011-10-31 11:07
last modified 2011-10-31 11:07
tagged with: @markup:md email journal perl programming
At YAPC: :Asia, I gave my talk [Email Hates the
Living!](http://www.youtube.com/watch?v=JENdgiAPD6c), and at the end there was
just one question: "When will you make
[Email::Sender](https://metacpan.org/release/Email-Sender) use
[Any::Moose](https://metacpan.org/module/Any::Moose)?"

It's a reasonable question. Email::Sender uses Moose, right now, which gives it a startup (compile-time) cost that's something like a quarter second higher than it would be with Mouse. I've replied to these requests before with a very stone-faced "never!" I just didn't want to deal with the inevitable bug reports related to it. All my friends at YAPC::Asia were so hospitable, though, and so polite in their questions about this change, that I decided to give it a go. While we sat around having beer and shabu-shabu, I pulled out my laptop and powered through an attempt to make it happen. As it turned out, it was not easy to do, and I have put it back on my list of things that I am unlikely to do myself.

Finding the problems was fun, though, as was explaining the problem with as little English as possible. I promised to publish the explanation, but before I got a chance to, someone submitted some patches to do the conversion. Awesome! Unfortunately, it was someone who hadn't been there, and didn't get to hear the explanation of the problem.

Lest it happen again, here goes:

Email::Sender is meant to replace Email::Send. Email::Send has a goofy API. When you tell it to send a message, it returns a Return::Value object, which might be false in boolean context. If it's false, you can inspect the object to find out why things failed. There's a lot more wrong with the API, but this one is pretty bad. Probably you realized that as soon as I said there's an object that might be false. ("[P]urveyors of exception object classes should have the good taste to make exception objects behave as true and stringify to something non-empty." -- Zefram))

Email::Sender addresses this problem by always returning an Email::Sender::Success object or throwing an Email::Sender::Failure exception. Its exceptions use my simple role for Moose exceptions: Throwable. With that said, I can start explaining the problem.

I converted Email::Sender to use Any::Moose within fifteen minutes or so. It was mostly a job for s/// and a little poking around here and there. All the tests passed that didn't throw errors. The errors failed because Email::Sender::Failure said this:

package Email::Sender::Failure;
use Any::Moose;
extends 'Throwable::Error';

When the Failure class was being loaded, other Email::Sender had already loaded, setting Any::Moose to use Mouse, and now a Mouse class (Email::Sender::Failure) was trying to extend a Moose class. This is fatal.

So, now you'd think you could go and make Throwable an Any::Moose role. This is not really a good idea, and here is the code I used to explain it:

# Abc.pm
package Abc;
use lib 'lib';
use Moose;
with('Role::AnyThing');
1;

# Xyz.pm
package Xyx;
use lib 'lib';
use Any::Moose;
with( 'Role::AnyThing' );
no Any::Moose;
1;

# Role/AnyThing.pm
package Role::AnyThing;
use lib 'lib';
use Any::Moose 'Role';
sub foo { ... }
no Any::Moose;
1;

Here I have two classes, one Moose and one Any::Moose. They both consume an Any::Moose role. So, what's the problem? Well:

$ perl -I lib -MAbc -MXyz -e0
$ echo $?
0

Great! I can use these in the same program! Or can I?

$ perl -I lib -MXyz -MAbc -e0
You can only consume roles, Role::AnyThing is not a Moose role at … line 137

What happened? Well, in the first example, Abc is compiled first. It loads Moose. When Xyz is compiled, it loads Any::Moose, which sees that Moose is loaded, and uses Moose as its backend. Everything is Moose. In the second example, Xyz is compiled first. It sees that Moose is not loaded and that Mouse is available. It uses Mouse as its backend. This means that Role::AnyThing – an Any::Moose role – is now a Mouse role. When Abc is compiled and tries to compose it, it fails: Abc is Moose but the role is Mouse.

Throwable exists to be used all over the place by all kinds of things. Sometimes it will be very, very simple and its behavior would be easy to produce in Mouse. Other times, though, it is getting composed with Moose-only roles. I throw exceptions that use Throwable together with MooseX modules. These can't be converted to Any::Moose, and I don't want to convert them.

The real solution is to have some sort of Any::Throwable, so that if you're using Any::Moose, you can use Any::Throwable and get a Throwable built against Any::Moose's backend, but can otherwise be sure that Throwable itself is always Moose-based. This isn't trivial, because Throwable::Error uses StackTrace::Auto, which would then need to be made to work properly under both Moose and Mouse.

So, we get back to what I think the solution is: make Moose faster. Of course, I'm even less likely to personally do much work on speeding up Moose compile time. I just don't have a lot of code that needs to run very quickly and repeatedly from a standing start. I write daemons, and I don't mind an extra quarter second startup time on those.

It would be nice if everyone could use Email::Sender safely without Moose, but I think it's a bunch more work than I am interested in doing. I'll definitely review patches, though, so let me know if you want to take a swing at making it all work.