ABLincoln is a PHP-based implementation of Facebook's PlanOut, a framework
for online field experimentation. ABLincoln makes it easy to deploy and
maintain sophisticated randomized experiments and to quickly iterate on these
experiments, while satisfying the constraints of large-scale Internet services
with many users., (*1)
Developers integrate ABLincoln by defining experiments that detail how inputs
(e.g. users, cookie IDs) should get mapped onto conditions. To set up an
experiment randomizing both the text and color of a button, you would create a
class like this:, (*2)
use \Vimeo\ABLincoln\Experiments\SimpleExperiment;
use \Vimeo\ABLincoln\Operators\Random as Random;
class MyExperiment extends SimpleExperiment
{
public function assign($params, $inputs)
{
$params->button_color = new Random\UniformChoice(
['choices' => ['#ff0000', '#00ff00']],
$inputs
);
$params->button_text = new Random\WeightedChoice(
[
'choices' => ['Join now!', 'Sign up.'],
'weights' => [0.3, 0.7]
],
$inputs
);
}
}
Then, in the application code, you query the Experiment object to find out what
values the current user should be mapped onto:, (*3)
$my_exp = new MyExperiment(['userid' => 42]);
$my_exp->get('button_color');
$my_exp->get('button_text');
Querying the experiment parameters automatically generates an exposure log that
we can direct to a location of our choice:, (*4)
{"name": "MyExperiment", "time": 1421622363, "salt": "MyExperiment", "inputs": {"userid": 42}, "params": {"button_color": "#ff0000", "button_text": "Join now!"}, "event": "exposure"}
The basic SimpleExperiment class logs to a local file by default. More
advanced behavior, such as Vimeo's methodology described below, can
easily be introduced to better integrate with your existing logging stack., (*5)
Installation
ABLincoln is maintained as an independent PHP Composer package hosted on
Packagist. Include it in in your composer.json file for nice autoloading
and dependency management:, (*6)
{
"require": {
"vimeo/ablincoln": "~1.0"
}
}
Comparison with the PlanOut Reference Implementation
ABLincoln and the original Python release of PlanOut are very similar in both
functionality and usage. Both packages implement abstract and concrete versions
of the Experiment and Namespace classes, parameter overrides to facilitate
testing, exposure logging systems, and various random assignment operators
tested and confirmed to produce identical outputs., (*7)
Notable differences between the two releases currently include:
- ABLincoln features native support for logging either to local files or
to any PSR-compliant stack through the use of included PHP logging traits
- ABLincoln does not currently include an interpreter for the
PlanOut language., (*8)
Usage in Production Environments
ABLincoln was ported and designed with scalability in mind. Here are a couple
ways that Vimeo has chosen to extend it to meet the needs of our testing
process:, (*9)
Application to an Existing Logging Stack
The Experiment logging traits provided with this port make it easy to log
exposure data in the most convenient way possible for your existing stack. A
quick implementation of the plug-and-play FileLoggerTrait andtail -f of
the log file is all you need to monitor parameter exposures in real-time.
Alternatively, the PSRLoggerTrait allows more customizable integration with
existing PSR-compliant logging code. Here at Vimeo, we use a basic
Monolog Handler to enforce PSR-3 compliance and allow PlanOut to talk
nicely to our existing logging infrastructure., (*10)
URL Overrides
ABLincoln already supports parameter overrides for quickly examining the
effects of difficult-to-test experiments. A simple way to integrate this
behavior with a live site is to pass overrides into your endpoint as a query
parameter:, (*11)
http://my.site/home.php?overrides=button_color:green,button_text:hello
Then it's a relatively simple task to parse the overrides from the query and
pass them into the PHP Experiment API override methods after instantiation., (*12)
Thanks
PlanOut, the software from which ABLincoln was ported, was originally
developed by Eytan Bakshy, Dean Eckles, and Michael S. Bernstein, and is
currently maintained by Eytan Bakshy at Facebook. Learn more
about good practice in large-scale testing by reading their paper,
Designing and Deploying Online Field Experiments., (*13)
Wallogit.com
