Write blurb about C3 fail

[catagits/Catalyst-Runtime.git] / lib / Catalyst / Upgrading.pod

=head1 Upgrading to Catalyst 5.80

Most applications and plugins should run unaltered on Catalyst 5.80.

However as a lot of refactoring work has taken place, and several changes have
been made which could cause incompatibilities. If your application or plugin
is using deprecated code, or relying on side-effects, then you could have
issues upgrading to this release.

Most issues found with pre-existing components have been easy to solve, and a
complete description of behaviour changes which may cause compatibility issues,
or warnings which are now emitted is included below to help if you have problems.

If you think you have found an upgrade related issue which is not covered in
this document, then please email the Catalyst list to discuss the problem.

=head1 Known backwards compatibility breakages.

=head2 Issues with Class::C3

Catalyst 5.80 uses L<Algorithm::C3> method dispatch order. This is built into
perl 5.10, and comes via L<Class::C3> for perl 5.8. This replaces L<NEXT>
with L<Class::C3::Adopt::NEXT>, forcing all components to resolve methods using
C3, rather than the unpredictable dispatch order of L<NEXT>.

To be able to do this, however, entails that the graph of superclasses for each
class must be linearizable using the C3 algorithm. Unfortunately, when
superclasses are being used as mixins, it is easy to get this wrong.

Most common is the case of:

    package Component1; # Note, this is the common case
    use base qw/Class::Accessor::Fast Class::Data::Inheritable/;

    package Component2; # Accidentally saying it this way round causes fail.
    use base qw/Class::Data::Inheritable Class::Accessor::Fast/;

    package GoesBang;
    use base qw/Component1 Component2/;

And the Catalyst plugin most often causing this, is
L<Catalyst::Plugin::Sesssion::Store::FastMMap> - if you are using this plugin
and see issues, then please upgrade!

This can, however, be found in your own application - the only solution is to
go through each base class of the class the error was reported against, until
you identify the ones in conflict, and resolve them.

=head2 Components which inherit from Moose::Object before Catalyst::Component

Moose components which say:

    package TestApp::Controller::Example;
    use Moose;
    extends qw/Moose::Object Catalyst::Component/;

to use the constructor provided by Moose, whilst working (if you do some hacks
with the C< BUILDARGS > method), will not work with Catalyst 5.80 as
C<Catalyst::Component> inherits from C<Moose::Object>, and so C< @ISA > fails
to linearise.

The fix for this is to not inherit directly from C<Moose::Object>
yourself. Having components which do not inherit their constructor from
C<Catalyst::Component> is B<unsupported>, and has never been recommended,
therefore you're on your own if you're using this technique. You'll need
to detect the version of Catalyst your application is running with and deal
with it appropriately.

You will also see this issue if you do the following:

    package TestApp::Controller::Example;
    use Moose;
    use base 'Catalyst::Controller';

as C< use base > appends to @ISA.

The correct way to use Moose in a component in a both forward and backwards
compatible way is:

    package TestApp::Controller::Root;
    use Moose;
    BEGIN { extends 'Catalyst::Component' }; # Or ::Controller, or whatever

Note that the C< extends > declaration needs to occur in a begin block for
L<attributes> to operate correctly. You also don't get the L<Moose::Object>
constructor, and therefore attribute initialization will not work as normally
expected. If you want to use Moose attributes, then they need to be made lazy
to correctly initialize.

=head3 use Moose in MyApp

Similar to the above, this will also fail:

    package MyApp;
    use Moose;
    use Catalyst qw/
      ConfigLoader
    /;
    __PACKAGE__->setup;

If you need to use Moose in your application class (e.g. for method modifiers
etc) then the correct technique is:

    package MyApp;
    use Moose;
    extends 'Catalyst';
    __PACKAGE__->setup(qw/
        ConfigLoader
    /);

=head2 Anonymous closures installed directly into the symbol table

If you have any code which installs anonymous subroutine references directly
into the symbol table, you may encounter breakages. The simplest solution is
to use L<Sub::Name> to name the subroutine. Example:

    # Original code, likely to break:
    my $full_method_name = join('::', $package_name, $method_name);
    *$full_method_name = sub { ... };

    # Fixed Code
    use Sub::Name 'subname';
    my $full_method_name = join('::',$package_name, $method_name);
    *$full_method_name = subname $full_method_name, sub { ... };

Additionally, you can take advantage of Catalysts use of L<Class::MOP> and
install the closure using the appropriate metaclass. Example:

    use Class::MOP;
    my $metaclass = Moose::Meta::Class->initialize($package_name);
    $metaclass->add_method($method_name => sub { ... });

=head2 Hooking into application setup

To execute code during application startup the following snippet in MyApp.pm
used to work:

    sub setup {
        my ($class, @args) = @_;
        $class->NEXT::setup(@args);
        ... # things to do after the actual setup
    }

With Catalyst 5.80 this won't work anymore. Due to the fact that Catalyst is
no longer using NEXT.pm for method resolution, this no longer works. The
functionality was only ever originally operational as L<NEXT> remembers what
methods have already been called, and will not call them again.

Using this now causes infinite recursion between MyApp::setup and
Catalyst::setup, due to other backwards compatibility issues related to how
plugin setup works. Moose method modifiers like C<< before|after|around 'setup
=> sub { ... }; >> also will not operate correctly on the setup method.

The right way to do it is this:

    after setup_finalize => sub {
        ... # things to do after the actual setup
    };

The setup_finalize hook was introduced as a way to void this issue.

=head2 Components with a new method which returns false

Previously, if you had a component which inherited from Catalyst::COMPONENT,
but overrode the new method to return false, then your class' configuration
would be blessed into a hash on your behalf, and this would be returned from
the COMPONENT method.

This behaviour makes no sense, and so has been removed. Implementing your own
C< new > method in components is B<highly> discouraged, instead, you should
inherit the new method from Catalyst::Component, and use Mooses BUILD
functionality and/or Moose attributes to perform any construction work
necessary for your class.

=head2 __PACKAGE__->mk_accessor('meta');

Won't work due to a limitation of L<Moose>. This is currently being fixed
inside Moose.

=head2 Class::Data::Inheritable side effects

Previously, writing to a class data accessor would copy the accessor method
down into your package.

This behaviour has been removed. Whilst the class data is still stored
per-class, it is stored on the metaclass of the class defining the accessor.

Therefore anything relying on the side-effect of the accessor being copied down
will be broken.

The following test demonstrates the problem:

    {
        package BaseClass;
        use base qw/Class::Data::Inheritable/;
        __PACKAGE__->mk_classdata('foo');
    }

    {
        package Child;
        use base qw/BaseClass/;
    }

    BaseClass->foo('base class');
    Child->foo('sub class');
    
    use Test::More;
    isnt(BaseClass->can('foo'), Child->can('foo'));

=head2 Extending Catalyst::Request or other classes in an ad-hoc manor using mk_accessors

Previously, it was possible to add additional accessors to Catalyst::Request
(or other classes) by calling the mk_accessors class method.

This is no longer supported - users should make a sub-class of the class whose
behaviour they would like to change, rather than globally polluting the
Catalyst objects.

=head2 Confused multiple inheritance with Catalyst::Component::COMPONENT

Previously, Catalyst's COMPONENT method would delegate to the method on the
right hand side, which could then delegate back again with NEXT. This (as it
is insane AND makes no sense with C3 method dispatch order), and is therefore
no longer supported.

If a COMPONENT method is detected in the inheritance hierarchy to the right
hand side of Catalyst::Component::COMPONENT, then the following warning
message will be emitted:

    There is a COMPONENT method resolving after Catalyst::Component
    in ${next_package}.

The correct fix is to re-arrange your class' inheritance hierarchy so that the
COMPONENT method you would like to inherit is the first (left-hand most)
COMPONENT method in your @ISA.

=head1 WARNINGS

=head2 Methods in Catalyst::Dispatcher

The following methods in Catalyst::Dispatcher are both an implementation
detail, which may change in the 5.8X release series, and therefore their use
is highly deprecated.

=over

=item tree

=item dispatch_types

=item registered_dispatch_types

=item method_action_class

=item action_hash

=item container_hash

=back

The first time one of these methods is called, a warning will be emitted:

    Class $class is calling the deprecated method Catalyst::Dispatcher::$public_method_name,
    this will be removed in Catalyst 5.9X

You should B<NEVER> be calling any of these methods from application code.

Plugins authors and maintainers whose plugins currently call these methods
should change to using the public API, or, if you do not feel the public API
adequately supports your use-case, please email the development list to
discuss what API features you need so that you can be appropriately supported.

=head2 Class naming to packages defined does not correspond.

In this version of Catalyst, if a component is loaded from disk, but no
symbols are defined in that component's name space after it is loaded, this
warning will be issued:

    require $class was successful but the package is not defined.

This is to protect against confusing bugs caused by mis-typing package names,
and will become a fatal error in a future version.

Please note that 'inner packages' (via L<Devel::InnerPackage>) are still fully
supported, this warning is only issued when component file naming does not map
to B<any> of the packages defined within that component.

=head2 $c->plugin method

Calling the plugin method is deprecated, and calling it at runtime is B<highly
deprecated>.

Instead you are recommended to use L< Catalyst::Model::Adaptor > or similar to
compose the functionality you need outside of the main application name space.

Calling the plugin method will not be supported past Catalyst 5.81.

=cut