I'm currently developing a small single-page Web app using Mojolicious. The app has a Javascript frontend (using Backbone) that talks to a REST-ish API; the layout of the source is roughly:
use Mojolicious::Lite;
# ... setup code ...
get '/' => sub {
my $c = shift;
# fetch+stash data for bootstrapped collections...
$c->render('app_template');
};
get '/api_endpoint' => sub {
my $c = shift;
# fetch appropriate API data...
$c->render(json => $response);
};
# ... more API endpoints ...
app->start;
The app template uses EP, but very minimally; the only server-side template directives just insert JSON for bootstrapped collections. It's deployed via Apache as a plain CGI script. (This isn't optimal, but it's for low-traffic internal use, and more intricate server configuration is problematic in context.) Perl CGI is configured via mod_perl.
This works most of the time, but occasionally the renderer somehow gets the idea that it should cache the template and ignore changes to it. The debug records in error_log show "Rendering cached template" rather than the normal "Rendering template", and my new changes to the template stop appearing in the browser. I can't find a reliable way to stop this, though it will eventually stop on its own according to conditions I can't discern.
How can I make the app notice template changes reliably? Alternatively, how can I disable template caching completely?
How can I make the app notice template changes reliably?
This is what the morbo development server is for. Morbo wouldn't be used for your live code deployment, but for a development environment where you are continually changing your code and templates. Generally changes to live code and templates are meant to be handled by restarting the application server, or Apache in your case. (Hypnotoad has a hot-restart capability for this purpose)
Alternatively, how can I disable template caching completely?
To do this, add the following setup code (outside of routes, after use Mojolicious::Lite):
app->renderer->cache->max_keys(0);
For old answer see below.
I turned the findings of this answer into a plugin and released it on CPAN as Mojolicious::Plugin::Renderer::WithoutCache after discussing wit Grinnz on IRC, where they encouraged a release.
You can use it like this:
use Mojolicious::Lite;
plugin 'Renderer::WithoutCache';
It will create a new Cache object that does nothing, and install that globally into the renderer. That way, it doesn't need to be created every time like my initial answer below did.
In theory, this should be faster than Grinnz' approach (which is more sensible), and since you explicitly don't want to cache, you obviously want things to be as fast as possible, right? It's supposedly faster because the real Mojo::Cache would still need to go and try to set the cache, but then abort because there are no more free keys, and it also would try to look up the values from the cache every time.
I benchmarked this with both Dumbbench and Benchmark. Both of them showed negligible results. I ran them each a couple of times, but they fluctuated a lot, and it's not clear which one is faster. I included output of a run where my implementation was faster, but it still shows how minuscule the difference is.
Benchmark with Dumbbench:
use Dumbbench;
use Mojolicious::Renderer;
use Mojolicious::Controller;
use Mojolicious::Plugin::Renderer::WithoutCache::Cache;
my $controller = Mojolicious::Controller->new;
my $renderer_zero_keys = Mojolicious::Renderer->new;
$renderer_zero_keys->cache->max_keys(0);
my $renderer_nocache = Mojolicious::Renderer->new;
$renderer_nocache->cache( Mojolicious::Plugin::Renderer::WithoutCache::Cache->new );
my $bench = Dumbbench->new(
target_rel_precision => 0.005,
initial_runs => 5000,
);
$bench->add_instances(
Dumbbench::Instance::PerlSub->new(
name => 'max_keys',
code => sub {
$renderer_zero_keys->render( $controller, { text => 'foobar' } );
}
),
Dumbbench::Instance::PerlSub->new(
name => 'WithoutCache',
code => sub {
$renderer_nocache->render( $controller, { text => 'foobar' } );
}
),
);
$bench->run;
$bench->report;
__END__
max_keys: Ran 8544 iterations (3335 outliers).
max_keys: Rounded run time per iteration: 5.19018e-06 +/- 4.1e-10 (0.0%)
WithoutCache: Ran 5512 iterations (341 outliers).
WithoutCache: Rounded run time per iteration: 5.0802e-06 +/- 5.6e-09 (0.1%)
Benchmark with Benchmark:
use Benchmark 'cmpthese';
use Mojolicious::Renderer;
use Mojolicious::Controller;
use Mojolicious::Plugin::Renderer::WithoutCache::Cache;
my $controller = Mojolicious::Controller->new;
my $renderer_zero_keys = Mojolicious::Renderer->new;
$renderer_zero_keys->cache->max_keys(0);
my $renderer_nocache = Mojolicious::Renderer->new;
$renderer_nocache->cache( Mojolicious::Plugin::Renderer::WithoutCache::Cache->new );
cmpthese(
-5,
{
'max_keys' => sub {
$renderer_zero_keys->render( $controller, { text => 'foobar' } );
},
'WithoutCache' => sub {
$renderer_nocache->render( $controller, { text => 'foobar' } );
},
}
);
__END__
Rate max_keys WithoutCache
max_keys 190934/s -- -2%
WithoutCache 193846/s 2% --
I recon in a heavy load environment with lots of calls it would eventually make a difference, but that is very hard to prove. So if you don't like to think about the internals of the cache, this plugin might be useful.
Old answer:
Looking at Mojolicious::Plugin::EPRenderer I found out that there is a cache. It's a Mojo::Cache instance, which has the methods get, set and max_keys, and inherits from Mojo::Base (like probably everything in Mojolicious).
The ::EPRenderer gets a $renderer, which is a Mojolicious::Renderer. It holds the Mojo::Cache instance. I looked at $c with Data::Printer, and found out that there is a $c->app that holds all of those.
Knowing this, you can easily make your own cache class that does nothing.
package Renderer::NoCache;
use Mojo::Base -base;
sub get {}
sub set {}
sub max_keys {}
Now you stick it into $c.
package Foo;
use Mojolicious::Lite;
get '/' => sub {
my $c = shift;
$c->app->renderer->cache( Renderer::NoCache->new );
$c->render(template => 'foo', name => 'World');
};
app->start;
__DATA__
## foo.html.ep
Hello <%= $name =%>.
Now every attempt to get or set the cache simply does nothing. It will try caching, but it will never find anything.
Of course it's not great to make a new object every time. It would be better to make that object once at startup and get it into the internal permanent version of app. You have CGI, so it might not make a difference.
You could also just monkey-patch the get out of Mojo::Cache. This more hacky approach will do the same thing:
package Foo;
use Mojolicious::Lite;
*Mojo::Cache::get = sub { };
get '/' => sub {
my $c = shift;
$c->render(template => 'foo', name => 'World');
};
app->start;
But beware: we just disabled fetching from every cache in your application that uses Mojo::Cache. This might not be what you want.
Related
I'm hacking on Request Tracker, which is written in perl and uses Mason
for the web interface. I'm trying to make a customized page, that involves an autohandler, an html page, and pulls in some methods in other comps. I have a simple class that I want to use to track a few things that I need for my parts of the interface. Right now all it
tracks is a database handle.
package RTx::FooBar::Web;
use strict;
use warnings;
use RTx::FooBar::Handle;
sub new
{
my $proto = shift;
$RT::Logger->debug("creating new");
my $class = ref($proto) || $proto;
my $self = {};
bless( $self, $class);
my $handle = RTx::FooBar::Handle->new();
$handle->Connect();
$self->{cfHandle} = $handle;
return $self;
}
sub DESTROY {
my $self = shift;
$RT::Logger->debug("destroy");
delete $self->{cfHandle};
}
sub CFHandle
{
my $self = shift;
return $self->{cfHandle};
}
1;
I tried sticking that into the session so I could use it wherever I needed
it in the web interface. So I try to use it in one web page - the autohandler does:
% $m->call_next;
<%INIT>
$RT::Logger->debug("my autohandler");
use RTx::FooBar::Web;
$session{cfWeb} ||= RTx::FooBar::Web->new();
</%INIT>
The thing that's bugging me right now (other than the fact that it's not
working) is that the logging in the "new" method prints out once, but the
logging in the DESTROY method prints out 56 times. And each time, the
debug in RTx::FooBar::Handle->DESTROY prints out as well, indicating that
$self->{cfHandle} wasn't removed. Can anybody suggest why this might be
happening? Is it because session is a tied hash?
*Update* I'm no longer using $session, and it's STILL destroying my handle 56 times after creating it once.
Just a guess here - $session is re-creating your objects from an externalized version (perhaps via Storable or the like). So it gets the object back without calling new so no logging. Not sure why it would be getting destroyed every time tho.
It looks like $session is in fact a tied hash, to an Apache::Session or something implementing the same interface (the docs talk about db backing for sessions).
If you want to have a global persistent object (i.e., not tied to a single request) then give it a fully-qualified name, e.g., $RTx::FooBar::Web::TheObject ||= RTx::FooBar::Web->new(); or use something like Cache::MemoryCache. Alternately, set up your own global variable in the MasonAllowGlobals setting, although you might need to get into RT's config files to change that.
I'm currently working on internationalizing a large Perl/Mason web application (Perl 5.8.0, Mason 1.48, mod_perl & Apache). In choosing a localization module, I decided to go with Locale::TextDomain over Locale::Maketext, mostly because the latter's plural form support isn't as nice as I'd like.
The hang-up I'm having with Locale::TextDomain is that it resolves which catalog to use for translations based on the process' locale. When I realized this, I got worried about how this would affect my application if I wanted users to be able to use different locales -- would it be possible that a change in locale to suit one user's settings would affect another user's session? For example, could there be a situation in which an English user received a page in German because a German user's session changed the process' locale? I'm not very knowledgeable about how Apache's thread/process model works, though it seems that if multiple users can be served by the same thread, this could happen.
This email thread would indicate that this is possible; here the OP describes the situation I'm thinking about.
If this is true, is there a way I can prevent this scenario while still using Locale::TextDomain? I suppose I could always hack at the module to load the catalogs in a locale-independent (probably using DBD::PO), but hopefully I'm just missing something that will solve my problem...
You entirely avoid the setlocale problems by using web_set_locale instead.
(That message on the mailing list predates the addition of that function by about 4 years.)
Edit: You are correct that global behaviour persists in Apache children, leading to buggy behaviour.
I wrote up a test case:
app.psgi
use 5.010;
use strictures;
use Foo::Bar qw(run);
my $app = sub {
my ($env) = #_;
run($env);
};
Foo/Bar.pm
package Foo::Bar;
use 5.010;
use strictures;
use Encode qw(encode);
use File::Basename qw(basename);
use Locale::TextDomain __PACKAGE__, '/tmp/Foo-Bar/share/locale';
use Locale::Util qw(web_set_locale);
use Plack::Request qw();
use Sub::Exporter -setup => { exports => [ 'run' ] };
our $DEFAULT_LANGUAGE = 'en'; # untranslated source strings
sub run {
my ($env) = #_;
my $req = Plack::Request->new($env);
web_set_locale($env->{HTTP_ACCEPT_LANGUAGE}, undef, undef, [
map { basename $_ } grep { -d } glob '/tmp/Foo-Bar/share/locale/*'
]); # XXX here
return $req
->new_response(
200,
['Content-Type' => 'text/plain; charset=UTF-8'],
[encode('UTF-8', __ 'Hello, world!')],
)->finalize;
}
The app runs as a PerlResponseHandler. When the user request a language that cannot be set, the call fails silently and the language that was used last successfully is still enabled.
The trick to fix this is to always set to a language that exists with a fallback mechanism. At the spot marked XXX, add the code or web_set_locale($DEFAULT_LANGUAGE), so that despite using a global setting, the behaviour cannot persist because we guarantee that it is set/changed once per request.
Edit 2: Further testing reveals that it's not thread-safe, sorry. Use the prefork MPM only which isolates requests as processes; however worker and event are affected because they are thread-based.
When you run perl -e "Bla->new", you get this well-known error:
Can't locate object method "new" via package "Bla"
(perhaps you forgot to load "Bla"?)
Happened in a Perl server process the other day due to an oversight of mine. There are multiple scripts, and most of them have the proper use statements in place. But there was one script that was doing Bla->new in sub blub at line 123 but missing a use Bla at the top, and when it was hit by a click without any of the other scripts using Bla having been loaded by the server process before, then bang!
Testing the script in isolation would be the obvious way to safeguard against this particular mistake, but alas the code is dependent upon a humungous environment. Do you know of another way to safeguard against this oversight?
Here's one example how PPI (despite its merits) is limited in its view on Perl:
use strict;
use HTTP::Request::Common;
my $req = GET 'http://www.example.com';
$req->headers->push_header( Bla => time );
my $au=Auweia->new;
__END__
PPI::Token::Symbol '$req'
PPI::Token::Operator '->'
PPI::Token::Word 'headers'
PPI::Token::Operator '->'
PPI::Token::Word 'push_header'
PPI::Token::Symbol '$au'
PPI::Token::Operator '='
PPI::Token::Word 'Auweia'
PPI::Token::Operator '->'
PPI::Token::Word 'new'
Setting the header and assigning the Auweia->new parse the same. So I'm not sure how you can build upon such a shaky foundation. I think the problem is that Auweia could also be a subroutine; perl.exe cannot tell until runtime.
Further Update
Okay, from #Schwern's instructive comments below I learnt that PPI is just a tokenizer, and you can build upon it if you accept its limitations.
Testing is the only answer worth the effort. If the code contains mistakes like forgetting to load a class, it probably contains other mistakes. Whatever the obstacles, make it testable. Otherwise you're patching a sieve.
That said, you have two options. You can use Class::Autouse which will try to load a module if it isn't already loaded. It's handy, but because it affects the entire process it can have unintended effects.
Or you can use PPI to scan your code and find all the class method calls. PPI::Dumper is very handy to understand how PPI sees Perl.
use strict;
use warnings;
use PPI;
use PPI::Dumper;
my $file = shift;
my $doc = PPI::Document->new($file);
# How PPI sees a class method call.
# PPI::Token::Word 'Class'
# PPI::Token::Operator '->'
# PPI::Token::Word 'method'
$doc->find( sub {
my($node, $class) = #_;
# First we want a word
return 0 unless $class->isa("PPI::Token::Word");
# It's not a class, it's actually a method call.
return 0 if $class->method_call;
my $class_name = $class->literal;
# Next to it is a -> operator
my $op = $class->snext_sibling or return 0;
return 0 unless $op->isa("PPI::Token::Operator") and $op->content eq '->';
# And then another word which PPI identifies as a method call.
my $method = $op->snext_sibling or return 0;
return 0 unless $method->isa("PPI::Token::Word") and $method->method_call;
my $method_name = $method->literal;
printf "$class->$method_name seen at %s line %d.\n", $file, $class->line_number;
});
You don't say what server enviroment you're running under, but from what you say it sounds like you could do with preloading all your modules in advance before executing any individual pages. Not only would this prevent the problems you're describing (where every script has to remember to load all the modules it uses) but it would also save you memory.
In pre-forking servers (as is commonly used with mod_perl and Apache) you really want to load as much of your code before your server forks for the first time so that the code is stored once in copy-on-write shared memory rather than mulitple times in each child process when it is loaded on demand.
For information on pre-loading in Apache, see the section of Practical mod_perl
I have multiple Catalyst applications running as FCGI.
Is there a benefit in consolidating them into a single one with multiple constrollers?
Thanks,
Simone
RAM, probably? I think the minimum each server is going to hold onto is about 15MB so you might be able to save something like 100MB if you’re running 3 apps with with 3 servers. But that’s pure back of the napkin speculation.
Another option, which would likely achieve most of the same savings would be to move to Plack deployment. E.g., the same three apps, without consolidation, deployed on the same server (this is untested but seems right)–
# file: mutli-app.psgi
use Plack::Builder;
use YourApp;
use OurApp;
use MyApp;
MyApp->setup_engine('PSGI');
my $mine = sub { MyApp->run(#_) };
YourApp->setup_engine('PSGI');
my $your = sub { YourApp->run(#_) };
OurApp->setup_engine('PSGI');
my $our = sub { OurApp->run(#_) };
builder {
mount "/mine" => builder {
enable "Plack::Middleware::Foo";
$mine;
};
mount "/secondperson" => $your;
mount "/shared" => $our;
};
And then run it with–
plackup multi-app.psgi
I want a platform independent utility to take screenshots (not just within the browser).
The utility would be able to take screenshots after fixed intervals of time and be easily configurable by the user in terms of
time between successive shots,
the format the shots are stored,
till when (time, event) should the script run, etc
Since I need platform independence, I think Perl is a good choice.
a. Before I start out, I want to know whether a similar thing already exists, so I can start from there?
Searching CPAN gives me these two relevant results :
Imager-Screenshot-0.009
Imager-Search-1.00
From those pages, the first one looks easier.
b. Which one of these Perl modules should I use?
Taking a look at the sources of both, Imager::Search isn't much more than a wrapper to Imager::Screenshot.
Here's the constructor:
sub new {
my $class = shift;
my #params = ();
#params = #{shift()} if _ARRAY0($_[0]);
my $image = Imager::Screenshot::screenshot( #params );
unless ( _INSTANCE($image, 'Imager') ) {
Carp::croak('Failed to capture screenshot');
}
# Hand off to the parent class
return $class->SUPER::new( image => $image, #_ );
}
Given that Imager::Search does not really extend Imager::Screenshot much more, I'd say you're looking at two modules that are essentially the same.