Elep

=head1 Beyond make test

Test::Harness is responsible for running test scripts, analysing
their output and reporting success or failure. When I type 
F<make test> (or F<./Build test>) for a module, Test::Harness is usually
used to run the tests (not all modules use Test::Harness but the
majority do).

To start exploring some of the features of Test::Harness I need to
switch from F<make test> to the F<prove> command (which ships with
Test::Harness). For the following examples I'll also need a recent
version of Test::Harness installed; 3.14 is current as I write.

For the examples I'm going to assume that we're working with a
'normal' Perl module distribution. Specifically I'll assume that
typing F<make> or F<./Build> causes the built, ready-to-install module
code to be available below ./blib/lib and ./blib/arch and that
there's a directory called 't' that contains our tests. Test::Harness
isn't hardwired to that configuration but it  saves me from explaining
which files live where for each example.

Back to F<prove>; like F<make test> it runs a test suite - but it
provides far more control over which tests are executed, in what
order and how their results are reported. Typically F<make test>
runs all the test scripts below the 't' directory. To do the same
thing with prove I type:

prove -rb t

The switches here are -r to recurse into any directories below 't'
and -b which adds ./blib/lib and ./blib/arch to Perl's include path
so that the tests can find the code they will be testing. If I'm
testing a module of which an earlier version is already installed
I need to be careful about the include path to make sure I'm not
running my tests against the installed version rather than the new
one that I'm working on.

Unlike F<make test>, typing F<prove> doesn't automatically rebuild
my module. If I forget to make before prove I will be testing against
older versions of those files - which inevitably leads to confusion.
I either get into the habit of typing

make && prove -rb t

or - if I have no XS code that needs to be built I use the modules
below F<lib> instead

prove -Ilib -r t

So far I've shown you nothing that F<make test> doesn't do. Let's
fix that.

=head2 Saved State

If I have failing tests in a test suite that consists of more than
a handful of scripts and takes more than a few seconds to run it
rapidly becomes tedious to run the whole test suite repeatedly as
I track down the problems.

I can tell prove just to run the tests that are failing like this:

prove -b t/this_fails.t t/so_does_this.t

That speeds things up but I have to make a note of which tests are
failing and make sure that I run those tests. Instead I can use
prove's --state switch and have it keep track of failing tests for
me. First I do a complete run of the test suite and tell prove to
save the results:

prove -rb --state=save t

That stores a machine readable summary of the test run in a file
called '.prove' in the current directory. If I have failures I can
then run just the failing scripts like this:

prove -b --state=failed

I can also tell prove to save the results again so that it updates
its idea of which tests failed:

prove -b --state=failed,save

As soon as one of my failing tests passes it will be removed from
the list of failed tests. Eventually I fix them all and prove can
find no failing tests to run:

Files=0, Tests=0, 0 wallclock secs ( 0.00 usr + 0.00 sys = 0.00 CPU)
  Result: NOTESTS

As I work on a particular part of my module it's most likely that
the tests that cover that code will fail. I'd like to run the whole
test suite but have it prioritize these 'hot' tests. I can tell
prove to do this:

prove -rb --state=hot,save t

All the tests will run but those that failed most recently will be
run first. If no tests have failed since I started saving state all
tests will run in their normal order. This combines full test
coverage with early notification of failures.

The --state switch supports a number of options; for example to run
failed tests first followed by all remaining tests ordered by the
timestamps of the test scripts - and save the results - I can use

prove -rb --state=failed,new,save t

See the prove documentation (type prove --man) for the full list
of state options.

When I tell prove to save state it writes a file called '.prove'
('_prove' on Windows) in the current directory. It's a YAML document
so it's quite easy to write tools of your own that work on the saved
test state - but the format isn't officially documented so it might
change without (much) warning in the future.

=head2 Parallel Testing

If my tests take too long to run I may be able to speed them up by
running multiple test scripts in parallel. This is particularly
effective if the tests are I/O bound or if I have multiple CPU
cores. I tell prove to run my tests in parallel like this:

prove -rb -j 9 t

The -j switch enables parallel testing; the number that follows it
is the maximum number of tests to run in parallel. Sometimes tests
that pass when run sequentially will fail when run in parallel. For
example if two different test scripts use the same temporary file
or attempt to listen on the same socket I'll have problems running
them in parallel. If I see unexpected failures I need to check my
tests to work out which of them are trampling on the same resource
and rename temporary files or add locks as appropriate.

To get the most performance benefit I want to have the test scripts
that take the longest to run start first - otherwise I'll be waiting
for the one test that takes nearly a minute to complete after all
the others are done. I can use the --state switch to run the tests
in slowest to fastest order:

prove -rb -j 9 --state=slow,save t

=head2 Non-Perl Tests

The Test Anything Protocol (http://testanything.org/) isn't just
for Perl. Just about any language can be used to write tests that
output TAP. There are TAP based testing libraries for C, C++, PHP,
Python and many others. If I can't find a TAP library for my language
of choice it's easy to generate valid TAP. It looks like this:

1..3 
  ok 1 - init OK 
  ok 2 - opened file 
  not ok 3 - appended to file

The first line is the plan - it specifies the number of tests I'm
going to run so that it's easy to check that the test script didn't
exit before running all the expected tests. The following lines are
the test results - 'ok' for pass, 'not ok' for fail. Each test has
a number and, optionally, a description. And that's it. Any language
that can produce output like that on STDOUT can be used to write
tests.

Recently I've been rekindling a two-decades-old interest in Forth.
Evidently I have a masochistic streak that even Perl can't satisfy.
I want to write tests in Forth and run them using prove (you can
find my gforth TAP experiments at
https://svn.hexten.net/andy/Forth/Testing/). I can use the --exec
switch to tell prove to run the tests using gforth like this:

prove -r --exec gforth t

Alternately, if the language used to write my tests allows a shebang
line I can use that to specify the interpreter. Here's a test written
in PHP:

#!/usr/bin/php 
  <?php
    print "1..2\n"; 
    print "ok 1\n"; 
    print "not ok 2\n";
  ?>

If I save that as t/phptest.t the shebang line will ensure that it
runs correctly along with all my other tests.

=head2 Mixing it up

Subtle interdependencies between test programs can mask problems -
for example an earlier test may neglect to remove a temporary file
that affects the behaviour of a later test. To find this kind of
problem I use the --shuffle and --reverse options to run my tests
in random or reversed order.

=head2 Rolling My Own

If I need a feature that prove doesn't provide I can easily write my own.

Typically you'll want to change how TAP gets I<input> into and I<output>
from the parser.  L<App::Prove> supports arbitrary plugins, and L<TAP::Harness>
supports custom I<formatters> and I<source handlers> that you can load using
either L<prove> or L<Module::Build>; there are many examples to base mine on.
For more details see L<App::Prove>, L<TAP::Parser::SourceHandler>, and
L<TAP::Formatter::Base>.

If writing a plugin is not enough, you can write your own test harness; one of
the motives for the 3.00 rewrite of Test::Harness was to make it easier to
subclass and extend.

The Test::Harness module is a compatibility wrapper around TAP::Harness.
For new applications I should use TAP::Harness directly. As we'll
see, prove uses TAP::Harness.

When I run prove it processes its arguments, figures out which test
scripts to run and then passes control to TAP::Harness to run the
tests, parse, analyse and present the results. By subclassing
TAP::Harness I can customise many aspects of the test run.

I want to log my test results in a database so I can track them
over time. To do this I override the summary method in TAP::Harness.
I start with a simple prototype that dumps the results as a YAML
document:

package My::TAP::Harness;

use base qw( TAP::Harness ); use YAML;

sub summary {
    my ( $self, $aggregate ) = @_; 
    print Dump( $aggregate );
    $self->SUPER::summary( $aggregate );
  }

I need to tell prove to use my My::TAP::Harness. If My::TAP::Harness
is on Perl's @INC include path I can

prove --harness=My::TAP::Harness -rb t

If I don't have My::TAP::Harness installed on @INC I need to provide
the correct path to perl when I run prove:

perl -Ilib `which prove` --harness=My::TAP::Harness -rb t

I can incorporate these options into my own version of prove. It's
pretty simple. Most of the work of prove is handled by App::Prove.
The important code in prove is just:

use App::Prove;

my $app = App::Prove->new; 
  $app->process_args(@ARGV); 
  exit( $app->run ? 0 : 1 );

If I write a subclass of App::Prove I can customise any aspect of
the test runner while inheriting all of prove's behaviour. Here's
myprove:

#!/usr/bin/env perl use lib qw( lib );      # Add ./lib to @INC
  use App::Prove;

my $app = App::Prove->new;

# Use custom TAP::Harness subclass
  $app->harness( 'My::TAP::Harness' );

$app->process_args( @ARGV ); exit( $app->run ? 0 : 1 );

Now I can run my tests like this

./myprove -rb t

=head2 Deeper Customisation

Now that I know how to subclass and replace TAP::Harness I can
replace any other part of the harness. To do that I need to know
which classes are responsible for which functionality. Here's a
brief guided tour; the default class for each component is shown
in parentheses. Normally any replacements I write will be subclasses
of these default classes.

When I run my tests TAP::Harness creates a scheduler
(TAP::Parser::Scheduler) to work out the running order for the
tests, an aggregator (TAP::Parser::Aggregator) to collect and analyse
the test results and a formatter (TAP::Formatter::Console) to display
those results.

If I'm running my tests in parallel there may also be a multiplexer
(TAP::Parser::Multiplexer) - the component that allows multiple
tests to run simultaneously.

Once it has created those helpers TAP::Harness starts running the
tests. For each test it creates a new parser (TAP::Parser) which
is responsible for running the test script and parsing its output.

To replace any of these components I call one of these harness
methods with the name of the replacement class:

aggregator_class 
  formatter_class 
  multiplexer_class 
  parser_class
  scheduler_class

For example, to replace the aggregator I would

$harness->aggregator_class( 'My::Aggregator' );

Alternately I can supply the names of my substitute classes to the
TAP::Harness constructor:

my $harness = TAP::Harness->new(
    { aggregator_class => 'My::Aggregator' }
  );

If I need to reach even deeper into the internals of the harness I
can replace the classes that TAP::Parser uses to execute test scripts
and tokenise their output. Before running a test script TAP::Parser
creates a grammar (TAP::Parser::Grammar) to decode the raw TAP into
tokens, a result factory (TAP::Parser::ResultFactory) to turn the
decoded TAP results into objects and, depending on whether it's
running a test script or reading TAP from a file, scalar or array
a source or an iterator (TAP::Parser::IteratorFactory).

Each of these objects may be replaced by calling one of these parser
methods:

source_class
  perl_source_class 
  grammar_class 
  iterator_factory_class
  result_factory_class

=head2 Callbacks

As an alternative to subclassing the components I need to change I
can attach callbacks to the default classes. TAP::Harness exposes
these callbacks:

parser_args      Tweak the parameters used to create the parser 
  made_parser      Just made a new parser 
  before_runtests  About to run tests 
  after_runtests   Have run all tests 
  after_test       Have run an individual test script

TAP::Parser also supports callbacks; bailout, comment, plan, test,
unknown, version and yaml are called for the corresponding TAP
result types, ALL is called for all results, ELSE is called for all
results for which a named callback is not installed and EOF is
called once at the end of each TAP stream.

To install a callback I pass the name of the callback and a subroutine
reference to TAP::Harness or TAP::Parser's callback method:

$harness->callback( after_test => sub {
    my ( $script, $desc, $parser ) = @_;
  } );

I can also pass callbacks to the constructor:

my $harness = TAP::Harness->new({
    callbacks => {
	    after_test => sub {
        my ( $script, $desc, $parser ) = @_; 
        # Do something interesting here
	    }
    }
  });

When it comes to altering the behaviour of the test harness there's
more than one way to do it. Which way is best depends on my
requirements. In general if I only want to observe test execution
without changing the harness' behaviour (for example to log test
results to a database) I choose callbacks. If I want to make the
harness behave differently subclassing gives me more control.

=head2 Parsing TAP

Perhaps I don't need a complete test harness. If I already have a
TAP test log that I need to parse all I need is TAP::Parser and the
various classes it depends upon. Here's the code I need to run a
test and parse its TAP output

use TAP::Parser;

my $parser = TAP::Parser->new( { source => 't/simple.t' } );
  while ( my $result = $parser->next ) {
    print $result->as_string, "\n";
  }

Alternately I can pass an open filehandle as source and have the
parser read from that rather than attempting to run a test script:

open my $tap, '<', 'tests.tap' 
    or die "Can't read TAP transcript ($!)\n"; 
  my $parser = TAP::Parser->new( { source => $tap } );
  while ( my $result = $parser->next ) {
    print $result->as_string, "\n";
  }

This approach is useful if I need to convert my TAP based test
results into some other representation. See TAP::Convert::TET
(http://search.cpan.org/dist/TAP-Convert-TET/) for an example of
this approach.

=head2 Getting Support

The Test::Harness developers hang out on the tapx-dev mailing
list[1]. For discussion of general, language independent TAP issues
there's the tap-l[2] list. Finally there's a wiki dedicated to the
Test Anything Protocol[3]. Contributions to the wiki, patches and
suggestions are all welcome.

[1] L<http://www.hexten.net/mailman/listinfo/tapx-dev>
[2] L<http://testanything.org/mailman/listinfo/tap-l>
[3] L<http://testanything.org/>