sixpack:一个和语言无关的 A/B 测试框架
sixpack:一个和语言无关的 A/B 测试框架
Python 测试工具
共1646Star
详细介绍
Sixpack
Sixpack is a framework to enable A/B testing across multiple programming languages. It does this by exposing a simple API for client libraries. Client libraries can be written in virtually any language.
Sixpack has two main parts. The first, Sixpack-server, is responsible for responding to web requests. The second, Sixpack-web, is a web dashboard for tracking and acting on your A/B tests. Sixpack-web is optional.
Requirements
- Redis >= 2.6
- Python >= 2.7 (3.0 untested, pull requests welcome)
Getting Started
To get going, create (or don't, but you really should) a new virtualenv for your Sixpack installation. Follow that up with pip install
:
$ pip install sixpack
Note: If you get an error like src/hiredis.h:4:20: fatal error: Python.h: No such file or directory
you need to install the python development tools. apt-get install python-dev
on Ubuntu.
Next, create a Sixpack configuration. A configuration must be created for Sixpack to run. Here's the default:
redis_port: 6379 # Redis port redis_host: localhost # Redis host redis_prefix: sixpack # all Redis keys will be prefixed with this redis_db: 15 # DB number in redis metrics: false # send metrics to StatsD (response times, # of calls, etc)? statsd_url: 'udp://localhost:8125/sixpack' # StatsD url to connect to (used only when metrics: true) # The regex to match for robots robot_regex: $^|trivial|facebook|MetaURI|butterfly|google|amazon|goldfire|sleuth|xenu|msnbot|SiteUptime|Slurp|WordPress|ZIBB|ZyBorg|pingdom|bot|yahoo|slurp|java|fetch|spider|url|crawl|oneriot|abby|commentreader|twiceler ignored_ip_addresses: [] # List of IP asset_path: gen # Path for compressed assets to live. This path is RELATIVE to sixpack/static secret_key: '<your secret key here>' # Random key (any string is valid, required for sixpack-web to run)
You can store this file anywhere (we recommend /etc/sixpack/config.yml
). As long as Redis is running, you can now start the Sixpack server like this:
$ SIXPACK_CONFIG=<path to config.yml> sixpack
Sixpack-server will be listening on port 5000 by default but can be changed with the SIXPACK_PORT
environment variable. For use in a production environment, please see the "Production Notes" section below.
Alternatively, as of version 1.1, all Sixpack configuration can be set by environment variables. The following environment variables are available:
SIXPACK_CONFIG_ENABLED
SIXPACK_CONFIG_REDIS_PORT
SIXPACK_CONFIG_REDIS_HOST
SIXPACK_CONFIG_REDIS_PASSWORD
SIXPACK_CONFIG_REDIS_PREFIX
SIXPACK_CONFIG_REDIS_DB
SIXPACK_CONFIG_ROBOT_REGEX
SIXPACK_CONFIG_IGNORE_IPS
- comma separatedSIXPACK_CONFIG_ASSET_PATH
SIXPACK_CONFIG_SECRET
SIXPACK_METRICS
STATSD_URL
Using the API
All interaction with Sixpack is done via HTTP GET
requests. Sixpack allows for cross-language testing by accepting a unique client_id
(which the client is responsible for generating) that links a participation to a conversion. All requests to Sixpack require a client_id
.
Participating in an Experiment
You can participate in an experiment with a GET
request to the participate
endpoint:
$ curl http://localhost:5000/participate?experiment=button_color&alternatives=red&alternatives=blue&client_id=12345678-1234-5678-1234-567812345678
If the test does not exist, it will be created automatically. You do not need to create the test in Sixpack-web.
Experiment names are not validated, so it is possible to explode the Redis keyspace. If you need to validate that the experiments being created are only those you wish to whitelist, consider fronting Sixpack with either Nginx+Lua/Openresty or Varnish, and performing your whitelisting logic there.
Arguments
experiment
(required) is the name of the test. Valid experiment names must be alphanumeric and can contain _
and -
.
alternatives
(required) are the potential responses from Sixpack. One of them will be the bucket that the client_id
is assigned to.
client_id
(required) is the unique id for the user participating in the test.
user_agent
(optional) user agent of the user making a request. Used for bot detection.
ip_address
(optional) IP address of user making a request. Used for bot detection.
force
(optional) force a specific alternative to be returned, example:
$ curl http://localhost:5000/participate?experiment=button_color&alternatives=red&alternatives=blue&force=red&client_id=12345678-1234-5678-1234-567812345678
In this example, red will always be returned. This is used for testing only, and no participation will be recorded.
traffic_fraction
(optional) Sixpack allows for limiting experiments to a subset of traffic. You can pass the percentage of traffic you'd like to expose the test to as a decimal number here. (?traffic_fraction=0.10
for 10%)
Response
A typical Sixpack participation response will look something like this:
{ status: "ok", alternative: { name: "red" }, experiment: { name: "button_color" }, client_id: "12345678-1234-5678-1234-567812345678" }
The most interesting part of this is alternative
. This is a representation of the alternative that was chosen for the test and assigned to a client_id
. All subsequent requests to this experiment/client_id combination will be returned the same alternative.
Converting a user
You can convert a user with a GET
request to the convert
endpoint:
$ curl http://localhost:5000/convert?experiment=button_color&client_id=12345678-1234-5678-1234-567812345678
Conversion Arguments
experiment
(required) the name of the experiment you would like to convert on.client_id
(required) the client you would like to convert.kpi
(optional) sixpack supports recording multiple KPIs. If you would like to track conversion against a specfic KPI, you can do that here. If the KPI does not exist, it will be created automatically.
Notes
You'll notice that the convert
endpoint does not take an alternative
query parameter. This is because Sixpack handles that internally with the client_id
.
We've included a 'health-check' endpoint, available at /_status
. This is helpful for monitoring and alerting if the Sixpack service becomes unavailable. The health check will respond with either 200 (success) or 500 (failure) headers.
Clients
We've already provided clients in four languages. We'd love to add clients in additional languages. If you feel inclined to create one, please first read the CLIENTSPEC. After writing your client, please update and pull request this file so we know about it.
Algorithm
As of version 2.0 of Sixpack, we use a deterministic algorithm to choose which alternative a client with receive. The algorithm was ported from Facebook's Planout project, and more information can be found HERE.
Dashboard
Sixpack comes with a built in dashboard. You can start the dashboard with:
$ SIXPACK_CONFIG=<path to config.yml> sixpack-web
The Sixpack dashboard allows you to visualize how each experiment's alternatives are doing compared to the rest, select alternatives as winners, and update experiment descriptions to something more human-readable.
Sixpack-web defaults to run on port 5001
but can be changed with the SIXPACK_WEB_PORT
environment variable.
API
Sixpack web dashboard has a bit of a read-only API built in. To get a list of all experiment information you can make a request like:
$ curl http://localhost:5001/experiments.json
To get the information for a single experiment, you can make a request like:
$ curl http://localhost:5001/experiments/blue-or-red-header.json
Production Notes
We recommend running Sixpack on gunicorn in production. You will need to install gunicorn in your virtual environment before running the following.
To run the sixpack server using gunicorn/gevent (a separate installation) you can run the following:
gunicorn --access-logfile - -w 8 --worker-class=gevent sixpack.server:start
To run the sixpack web dashboard using gunicorn/gevent (a separate installation) you can run the following:
gunicorn --access-logfile - -w 2 --worker-class=gevent sixpack.web:start
Note: After selecting an experiment winner, it is best to remove the Sixpack experiment code from your codebase to avoid unnecessary http requests.
Contributing
Fork it
Start Sixpack in development mode with:
$ PYTHONPATH=. SIXPACK_CONFIG=<path to config.yml> bin/sixpack
and:
$ PYTHONPATH=. SIXPACK_CONFIG=<path to config.yml> bin/sixpack-web
We've also included a small script that will seed Sixpack with lots of random data for testing and development on sixpack-web. You can seed Sixpack with the following command:
$ PYTHONPATH=. SIXPACK_CONFIG=<path to config.yml> sixpack/test/seed
This command will make a few dozen requests to the
participate
andconvert
endpoints. Feel free to run it multiple times to get additional data.Note: By default the server runs in production mode. If you'd like to turn on Flask and Werkzeug debug modes set the
SIXPACK_DEBUG
environment variable totrue
.Create your feature branch (
git checkout -b my-new-feature
)Write tests
Run tests with
nosetests
Commit your changes (
git commit -am 'Added some feature'
)Push to the branch (
git push origin my-new-feature
)Create new pull request
Please avoid changing versions numbers; we'll take care of that for you.
Using Sixpack in production?
If you're a company using Sixpack in production, kindly let us know! We're going to add a 'using Sixpack' section to the project landing page, and we'd like to include you. Drop Jack a line at jack [at] seatgeek dot.com with your company name.
License
Sixpack is released under the BSD 2-Clause License.