In our Perl web code we use Template Toolkit, and we have extended the engine by writing various plugins where we have found them useful.

One such plugin extended Template::Plugin::Filter as follows (details have been changed to protect the innocent):

package CVLibrary::Template::SomeFilter;

use strict;
use warnings;

use parent 'Template::Plugin::Filter';

use SomeModule 'some_function';

sub init {
    my $self = shift;
    $self->{_DYNAMIC} = 1;
    return $self;

sub filter {
    return some_function( $_[1] );


This filter plugin has been in production for quite some time, and as far as we knew, it worked well. Recently, however, we started paying more attention to memory leak issues (the subject, I’m sure, of a future blog post).

We found that Template::Plugin::Filter was creating a circular reference, and tracked this down to a particular upstream bug:

As you can see, in the source the “weaken” call which would fix the leak has been commented out. The comments above that line detail the reason why, and ask us to wait until enough people report memory leak issues!

The unweakened reference means that if you “use” the filter more than once per template object, one of the instances will not get cleaned up. For example:

[% USE SomeFilter %]
[... and later, perhaps in an included template ...]
[% USE SomeFilter # this leaks the original plugin instance %]

At this point I’d love to be able to say that we fixed this upstream bug. Alas, we haven’t. However, we found a workaround, by rewriting our template plugin to avoid the module with the leak (and make it shorter!):

package CVLibrary::Template::SomeFilter;

use strict;
use warnings;

# Can't use "Template::Plugin::Filter" as it leaks:
use parent 'Template::Plugin';

use SomeModule ();

sub new {
    my ( $class, $context ) = @_;

        some_function => \&SomeModule::some_function );

    bless {}, $class;


It turns out that for short filters, the base class is not hiding much complexity after all, so we can inherit directly from Template::Plugin, avoiding the problematic code.

(Credit is due to James Raspass for actually doing the debugging/development work on this issue!)