Re^3: Moose: return null object

[stvn]

So all that's fine, just it seems (very) silly (as in: pointlessly silly, silly oversight kind of silly) that you cannot return undef from a "new" call, and thot perhaps I'd missed something. But I'm not here to pick on Moose either ;)

Well, basically it is not an oversight at all, it is *very* deliberate. This is basically a variant of the "return value" vs. "exception" debate which happens any time you lock a C programmer and Java programmer together in a room for a long enough period of time. We (myself and the other core Moose devs) believe that exceptions are the way to go, so therefore Moose reflects this opinion. But it also goes a little deeper then that, to the philosophy of what is a "Class".

In non-Moose Perl OO, new was simply a method that you could make do whatever you wanted, convention said it was the method that constructed a new instance, but that was simply just convention. There was nothing special about new either, in fact it was a common idiom for a while to do $instance->new as a sort of cloning technique. But in the end, a "Class" in non-Moose Perl OO is not really a rigorously defined concept, but instead really just a loose set of conventions accumulated over time and borrowed from other languages. Now, there is nothing at all wrong with this, it is very powerful and extremely flexible, however it also suffers from the common "dark-side" of TIMTOWTDI.

In contrast, Moose defines "Class" much more rigorously, in particular it thinks that one of the primary functions of a "Class" is as an instance construction factory. So much so that new in Moose will eventually call $class->meta->new_object to get the instance (delegating up to the Class meta-object). So keeping with this philosophy, new should *always* return an instance, and if it cannot for whatever reason return an instance, it should explode loudly with an exception.

[Corion]

If you, as core developer, believe that "exceptions are the way to go", why does Moose produce such horrible, unhelpful, Java-style backtraces when dieing? For example when a type constraint is violated, I get about 7 lines of output that contribute nothing to the error cause, and have to combine the first and the last line to make up something that might actually relate to the error at hand:

Q:\>perl -MMoose -e "package Moo;use Moose; has 'foo' => (is => 'ro', 
+isa => 'Int'); no Moose; package main; my $bar = Moo->new({foo => 'bo
+om'});"
Attribute (foo) does not pass the type constraint because: Validation 
+failed for 'Int' failed with value boom at g:/Systeme/tools/pe
rl-5.10.0/perl/site/lib/Moose/Meta/Attribute.pm line 774
        Moose::Meta::Attribute::_coerce_and_verify('Moose::Meta::Attri
+bute=HASH(0x190c5f4)', 'boom', 'Moo=HASH(0x190c934)') called a
t g:/Systeme/tools/perl-5.10.0/perl/site/lib/Moose/Meta/Attribute.pm l
+ine 428
        Moose::Meta::Attribute::initialize_instance_slot('Moose::Meta:
+:Attribute=HASH(0x190c5f4)', 'Moose::Meta::Instance=HASH(0x187
145c)', 'Moo=HASH(0x190c934)', 'HASH(0x3ea354)') called at g:/Systeme/
+tools/perl-5.10.0/perl/site/lib/Class/MOP/Class.pm line 365
        Class::MOP::Class::_construct_instance('Moose::Meta::Class=HAS
+H(0x187104c)', 'HASH(0x3ea354)') called at g:/Systeme/tools/pe
rl-5.10.0/perl/site/lib/Class/MOP/Class.pm line 352
        Class::MOP::Class::new_object('Moose::Meta::Class=HASH(0x18710
+4c)', 'HASH(0x3ea354)') called at g:/Systeme/tools/perl-5.10.0
/perl/site/lib/Moose/Meta/Class.pm line 205
        Moose::Meta::Class::new_object('Moose::Meta::Class=HASH(0x1871
+04c)', 'HASH(0x3ea354)') called at g:/Systeme/tools/perl-5.10.
0/perl/site/lib/Moose/Object.pm line 25
        Moose::Object::new('Moo', 'HASH(0x12ca3ec)') called at -e line
+ 1

If you think something is a good engineering practice when using your tool, wouldn't it be a good idea to design your tool in a way that your tool encourages the practice instead of hindering it? Sending the programmer on a merry chase along exception paths is not friendly to the programmer.

If you use one of the standard exception mechanisms, like for example Carp, judicious use of @CARP_NOT can slim down the backtrace tree easily. Also, either passing down the (top-level) error cause to the place where the error gets detected will improve the feedback instead of requiring the programmer to intuit the error cause from the message(s) they get.

[stvn]

There seems to be no consensus on stack traces. Personally I find them very useful, you obviously differ in that opinion but still do see some usefulness, while others think they are simply evil incarnate.

To be honest, I wanted originally to use exception objects, but so many people/tools/programs expect die to throw a string and so many existing exception-object modules jumped through so many hoops it just turned me off from the idea. That said, these days I use Throwable and love it.

If you use one of the standard exception mechanisms, like for example Carp, judicious use of @CARP_NOT can slim down the backtrace tree easily

Moose does use Carp actually.

I will the first to admit (and I suspect many of the other core devs would agree) that Moose could use some better errors. There are few modules out there that couldn't use some help with errors, it is all too common to leave them till the end. Any help you would like to provide, I am sure would be greatly appreciated.

[Corion]

I'm not actively using Moose except for one toy project. It's thus highly unlikely that I'll try to undertake the augeian task of modifying the layers upon layers of abstraction just to try to make the world better for people who won't appreciate my work.

[nxadm]

Thank you for the information. I have asked myself the same question as the original poster. Probably because I also write in Java and a null object is pretty common there. This should certainly go into the Moose FAQ.

[stvn]

Probably because I also write in Java and a null object is pretty common there

It is, but you will never see one returned from a constructor.