path: root/docs/paste-deploy.txt

Paste Deployment
================

:author: Ian Bicking <ianb@colorstudy.com>
:revision: $Rev$
:date: $LastChangedDate$

.. contents::

Introduction
------------

Paste Deployment is a system for finding and configuring WSGI
applications and servers.  For WSGI application consumers it provides
a single, simple function (``loadapp``) for loading a WSGI application
from a configuration file or a Python Egg.  For WSGI application
providers it only asks for a single, simple entry point to your
application, so that application users don't need to be exposed to the
implementation details of your application.

The result is something a system administrator can install and manage
without knowing any Python, or the details of the WSGI application or
its container.

Paste Deployment currently does not require other parts of `Paste
<http://pythonpaste.org>`_, and is distributed as a separate package.

To see updates that have been made to Paste Deploy see the `news file
<news.html>`_.

Installation
------------

First install `setuptools
<http://peak.telecommunity.com/DevCenter/setuptools>`_.

You can install Paste Deployment by installing `easy_install
<http://peak.telecommunity.com/DevCenter/EasyInstall>`_ and running::

    $ sudo easy_install.py http://svn.pythonpaste.org/Paste/Deploy/trunk

If you want to track development, do::

    $ svn co http://svn.pythonpaste.org/Paste/Deploy/trunk Paste-Deploy
    $ cd Paste-Deploy
    $ sudo python setup.py develop

This will install the package globally, but will load the files in the
checkout.

Basic Usage
-----------

The basic way you'll use Paste Deployment is to load `WSGI
<http://www.python.org/peps/pep-0333.html>`_ applications.  Many
Python frameworks now support WSGI, so applications written for these
frameworks should be usable.

The primary function is ``paste.deploy.loadapp``.  This loads an
application given a URI.  You can use it like::

    from paste.deploy import loadapp
    wsgi_app = loadapp('config:/path/to/config.ini')

There's two URI formats currently supported: ``config:`` and ``egg:``.

``config:`` URIs
----------------

URIs that being with ``config:`` refer to configuration files.  These
filenames can be relative if you pass the ``relative_to`` keyword
argument to ``loadapp()``.  

.. note:: 
   
   Filenames are never considered relative to the current working
   directory, as that is a unpredictable location.  Generally when
   a URI has a context it will be seen as relative to that context;
   for example, if you have a ``config:`` URI inside another
   configuration file, the path is considered relative to the
   directory that contains that configuration file.

Config Format
~~~~~~~~~~~~~

Configuration files are in the INI format.  This is a simple format
that looks like::

    [section_name]
    key = value
    another key = a long value
        that extends over multiple lines

All values are strings (no quoting is necessary).  The keys and
section names are case-sensitive, and may contain punctuation and
spaces (though both keys and values are stripped of leading and
trailing whitespace).  Lines can be continued with leading whitespace.

Lines beginning with ``#`` (preferred) or ``;`` are considered
comments.

Applications
~~~~~~~~~~~~

You can define multiple applications in a single file; each
application goes in its own section.  Even if you have just one
application, you must put it in a section.

Each section name defining an application should be prefixed with
``app:``.  The "main" section (when just defining one application)
would go in ``[app:main]`` or just ``[app]``.

There's two ways to indicate the Python code for the application.  The
first is to refer to another URI or name::

    [app:myapp]
    use = config:another_config_file.ini#app_name

    # or any URI:
    [app:myotherapp]
    use = egg:MyApp

    # or even another section:
    [app:mylastapp]
    use = myotherapp

It would seem at first that this was pointless; just a way to point to
another location.  However, in addition to loading the application
from that location, you can also add or change the configuration.

The other way to define an application is to point exactly to some
Python code::

    [app:myapp]
    paste.app_factory = myapp.modulename:app_factory

You must give an explicit *protocol* (in this case
``paste.app_factory``), and the value is something to import.  In
this case the module ``myapp.modulename`` is loaded, and the
``app_factory`` object retrieved from it.

See `Defining Factories`_ for more about the protocols.

Configuration
~~~~~~~~~~~~~

Configuration is done through keys besides ``use`` (or the protocol
names).  Any other keys found in the section will be passed as keyword
arguments to the factory.  This might look like::

    [app:blog]
    use = egg:MyBlog
    database = mysql://localhost/blogdb
    blogname = This Is My Blog!

You can override these in other sections, like::

    [app:otherblog]
    use = blog
    blogname = The other face of my blog

This way some settings could be defined in a generic configuration
file (if you have ``use = config:other_config_file``) or you can
publish multiple (more specialized) applications just by adding a
section.

Global Configuration
~~~~~~~~~~~~~~~~~~~~

Often many applications share the same configuration.  While you can
do that a bit by using other config sections and overriding values,
often you want that done for a bunch of disparate configuration
values.  And typically applications can't take "extra" configuration
parameters; with global configuration you do something equivalent to
"if this application wants to know the admin email, this is it".

Applications are passed the global configuration separately, so they
must specifically pull values out of it; typically the global
configuration serves as the basis for defaults when no local
configuration is passed in.

Global configuration to apply to every application defined in a file
should go in a special section named ``[DEFAULT]``.  You can override
global configuration locally like::

    [DEFAULT]
    admin_email = webmaster@example.com

    [app:main]
    use = ...
    set admin_email = bob@example.com

That is, by using ``set`` in front of the key.

Composit Applications
~~~~~~~~~~~~~~~~~~~~~

"Composit" applications are things that act like applications, but
are made up of other applications.  One example would be a URL mapper,
where you mount applications at different URL paths.  This might look
like::

    [composit:main]
    / = mainapp
    /files = staticapp

    [app:mainapp]
    use = egg:MyApp

    [app:staticapp]
    use = egg:Paste#static
    document_root = /path/to/docroot

The composit application "main" is just like any other application
from the outside (you load it with ``loadapp`` for instance), but it
has access to other applications defined in the configuration file.

Other Objects
~~~~~~~~~~~~~

In addition to sections with ``app:``, you can define filters and
servers in a configuration file, with ``server:`` and ``filter:``
prefixes.  You load these with ``loadserver`` and ``loadfilter``.  The
configuration works just the same; you just get back different kinds
of objects.

Filter Composition
~~~~~~~~~~~~~~~~~~

There are several ways to apply filters to applications.  It mostly
depends on how many filters, and in what order you want to apply them.

The first way is to use the ``filter-with`` setting, like::

    [app:main]
    use = egg:MyEgg
    filter-with = printdebug

    [filter:printdebug]
    use = egg:Paste#printdebug

Also, two special section types exist to apply filters to your
applications: ``[filter-app:...]`` and ``[pipeline:...]``.  Both of
these sections define applications, and so can be used wherever an
application is needed.

``filter-app`` defines a filter (just like you would in a
``[filter:...]`` section), and then a special key ``next`` which
points to the application to apply the filter to.

``pipeline:`` is used when you need apply a number of filters.  It
takes *one* configuration key ``pipeline`` (plus any global
configuration overrides you want).  ``pipeline`` is a list of filters
ended by an application, like::

    [pipeline:main]
    pipeline = filter1 egg:FilterEgg#filter2 filter3 app

    [filter:filter1]
    ...

``egg:`` URIs
-------------

`Python Eggs <http://peak.telecommunity.com/DevCenter/PythonEggs>`_
are a distribution and installation format produced by `setuptools
<http://peak.telecommunity.com/DevCenter/setuptools>`_ that adds
metadata to a normal Python package (among other things).

You don't need to understand a whole lot about Eggs to use them.  If
you have a `distutils
<http://python.org/doc/current/lib/module-distutils.html>`_
``setup.py`` script, just change::

    from distutils.core import setup

to::

    from setuptools import setup

Now when you install the package it will be installed as an egg.

The first important part about an Egg is that it has a
*specification*.  This is formed from the name of your distribution
(the ``name`` keyword argument to ``setup()``), and you can specify a
specific version.  So you can have an egg named ``MyApp``, or
``MyApp==0.1`` to specify a specific version.

The second is *entry points*.  These are references to Python objects
in your packages that are named and have a specific protocol.
"Protocol" here is just a way of saying that we will call them with
certain arguments, and expect a specific return value.  We'll talk
more about the protocols later_.

.. _later: `Defining Factories`_

The important part here is how we define entry points.  You'll add an
argument to ``setup()`` like::

    setup(
        name='MyApp',
        ...
        entry_points={
            'paste.app_factory': [
                'main=myapp.mymodule:app_factory',
                'ob2=myapp.mymodule:ob_factory'],
            },
        )

This defines two applications named ``main`` and ``ob2``.  You can
then refer to these by ``egg:MyApp#main`` (or just ``egg:MyApp``,
since ``main`` is the default) and ``egg:MyApp#ob2``.

The values are instructions for importing the objects.  ``main`` is
located in the ``myapp.mymodule`` module, in an object named
``app_factory``.

There's no way to add configuration to objects imported as Eggs.

Defining Factories
------------------

This lets you point to factories (that obey the specific protocols we
mentioned).  But that's not much use unless you can create factories
for your applications.

There's a few protocols: ``paste.app_factory``,
``paste.composit_factory``, ``paste.filter_factory``, and lastly
``paste.server_factory``.  Each of these expects a callable (like a
function, method, or class).

``paste.app_factory``
~~~~~~~~~~~~~~~~~~~~~~

The application is the most common.  You define one like::

    def app_factory(global_config, **local_conf):
        return wsgi_app

The ``global_config`` is a dictionary, and local configuration is
passed as keyword arguments.  The function returns a WSGI application.

``paste.composit_factory``
~~~~~~~~~~~~~~~~~~~~~~~~~~~

Composits are just slightly more complex::

    def composit_factory(loader, global_config, **local_conf):
        return wsgi_app

The ``loader`` argument is an object that has a couple interesting
methods.  ``get_app(name_or_uri, global_conf=None)`` return a WSGI
application with the given name.  ``get_filter`` and ``get_server``
work the same way.

A more interesting example might be a composit factory that does
something.  For instance, consider a "pipeline" application::

    def pipeline_factory(loader, global_config, pipeline):
        # space-separated list of filter and app names:
        pipeline = pipeline.split()
        filters = [loader.get_filter(n) for n in pipeline[:-1]]
        app = loader.get_app(pipeline[-1])
        filters.reverse() # apply in reverse order!
        for filter in filters:
            app = filter(app)
        return app

Then we use it like::

    [composit:main]
    use = <pipeline_factory_uri>
    pipeline = egg:Paste#printdebug session myapp

    [filter:session]
    use = egg:Paste#session
    store = memory

    [app:myapp]
    use = egg:MyApp

``paste.filter_factory``
~~~~~~~~~~~~~~~~~~~~~~~~~

Filter factories are just like app factories (same signature), except
they return filters.  Filters are callables that take a WSGI
application as the only argument, and return a "filtered" version of
that application.

Here's an example of a filter that checks that the ``REMOTE_USER`` CGI
variable is set, creating a really simple authentication filter::

    def auth_filter_factory(global_conf, req_usernames):
        # space-separated list of usernames:
        req_usernames = req_usernames.split()
        def filter(app):
            return AuthFilter(app, req_usernames)
        return filter

    class AuthFilter(object):
        def __init__(self, app, req_usernames):
            self.app = app
            self.req_usernames = req_usernames

        def __call__(self, environ, start_response):
            if environ.get('REMOTE_USER') in self.req_usernames:
                return self.app(environ, start_response)
            start_response(
                '403 Forbidden', [('Content-type', 'text/html')])
            return ['You are forbidden to view this resource']

``paste.filter_app_factory``
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

This is very similar to ``paste.filter_factory``, except that it also
takes a ``wsgi_app`` argument, and returns a WSGI application.  So if
you changed the above example to::

    class AuthFilter(object):
        def __init__(self, app, global_conf, req_usernames):
            ....

Then ``AuthFilter`` would serve as a filter_app_factory
(``req_usernames`` is a required local configuration key in this
case).

``paste.server_factory``
~~~~~~~~~~~~~~~~~~~~~~~~~

This takes the same signature as applications and filters, but returns
a server.

A server is a callable that takes a single argument, a WSGI
application.  It then serves the application.

An example might look like::

    def server_factory(global_conf, host, port):
        port = int(port)
        def serve(app):
            s = Server(app, host=host, port=port)
            s.serve_forever()
        return serve

An implementation of ``Server`` is left to the user.

``paste.server_runner``
~~~~~~~~~~~~~~~~~~~~~~~~

Like ``paste.server_factory``, except ``wsgi_app`` is passed as the
first argument, and the server should run immediately.

Outstanding Issues
------------------

* Should add a ``python:`` scheme for loading objects out of modules
  directly.  It has to include the protocol somehow...?

* Should there be a "default" protocol for each type of object?  Since
  there's currently only one protocol, it seems like it makes sense
  (in the future there could be multiple).  Except that
  ``paste.app_factory`` and ``paste.composit_factory`` overlap
  considerably.

* ConfigParser's INI parsing is kind of annoying.  I'd like it both
  more constrained and less constrained.  Some parts are sloppy (like
  the way it interprets ``[DEFAULT]``).

* ``config:`` URLs should be potentially relative to other locations,
  e.g., ``config:$docroot/...``.  Maybe using variables from
  ``global_conf``?

* Should other variables have access to ``global_conf``?

* Should objects be Python-syntax, instead of always strings?  Lots of
  code isn't usable with Python strings without a thin wrapper to
  translate objects into their proper types.