Finish repository conversion docs, move re_define to app.php.

[wizard.git] / doc / repository-conversion.rst

Repository conversion
=====================

.. highlight:: sh

One of Wizard's goals is to replace the previous autoinstaller infrastructure.
Pre-wizard autoinstalls live in :file:`/mit/scripts/deploy` and consist of a
tarball from upstream, possibly a scripts patch, and possibly some post-install
munging (such as the creation of a :file:`php.ini` file and appropriate
symlinks).

Conversion to use Wizard involves placing :term:`pristine` versions of the source
code (from the upstream tarballs) and appropriately patched scripts
versions into a Git repository, as well as writing a :mod:`wizard.app`
module for the application that implements application specific logic, such
as how to install, upgrade or backup the installation.

Here is a tutorial for performing a conversion, using Wordpress as
an example.

Setup
-----

Probably the easiest way to do development is entirely on :term:`AFS`:  all
of your source code should live in publically readable (i.e.
``system:anyuser`` as read permissions) directories, so that if you
SSH into a scripts server to perform testing, you will be able
to invoke your tools and read your development repository.  In order
to be able to run the test scripts in the tests directory, this
is preferably in :file:`web_scripts`. In that
case, setup is as simple as::

    git clone /mit/scripts/git/wizard.git /mit/$USER/web_scripts/wizard
    athrun consult fsr /mit/$USER/web_scripts/wizard system:anyuser read
    # for any application you're going to do development on, also:
    git clone /mit/scripts/git/autoinstalls/$APP.git /mit/$USER/web_scripts/wizard/srv/$APP

If you'd like to be able to develop offline, just note that you will
have to push your changes to AFS once you start doing testing on
scripts servers, but before your changes get incorporated into
canonical upstream.  Git doesn't exactly make this easy, but you
can follow this recipe::

    git clone /mit/scripts/git/wizard.git ~/wizard
    cd /mit/$USER
    mkdir wizard.git
    cd wizard.git
    git init --bare
    cd ~/wizard
    git remote add afs /mit/$USER/wizard.git
    git push -f afs master
    git clone /mit/$USER/wizard.git /mit/$USER/wizard

And then you can perform updates from your local copy with::

    git push afs master
    cd /mit/$USER/wizard
    git pull

If :file:`/mit/$USER/wizard.git` has write permissions for
``system:scripts-security-upd``, this is especially useful if you were hacking
on a copy living on ``not-backward.mit.edu``, and now need to transfer the
changes back to the canonical repository (please don't give ``not-backward.mit.edu``
your root tickets!)  You can also setup a wizard directory similar to the
first set of directions for on-server testing.

From this point on, we will assume you are doing development from an AFS directory
named ``$WIZARD``; note that application repositories live in ``$WIZARD/srv``.

Pristine
--------

This is a tutorial centered around migrating `Wordpress <http://wordpress.org/>`_
into our Git repository.  For the sake of demonstration,
we shall assume that this repository hasn't been created yet.
The repository then doesn't exist, we should create it::

    cd "$WIZARD/srv"
    mkdir wordpress
    cd wordpress
    git init

We also have to create a module for the application, so we
create ``$WIZARD/wizard/app/wordpress.py`` and fill it in with a bare bones template:

.. code-block:: python

    import os
    import re
    import logging

    from wizard import app, install, resolve, sql, util
    from wizard.app import php

    class Application(app.Application):
        pass

Now we are ready to put some code in our repository: the first thing we will
add is the :term:`pristine` branch, which contains verbatim the code from upstream.
If we were starting a new autoinstaller, we'd pop off and use the latest version,
but since we're dealing with legacy we want to start our repository history
with the **oldest** version still extant on our servers.  To find this out run::

    wizard summary version APP

You'll need to be in the ``scripts-team`` list in order to access the data to
run this command.

Try running the following command in :file:`$WIZARD/srv/wordpress`::

    wizard prepare-pristine wordpress-2.0.2

You should get an error complaining about :meth:`wizard.app.Application.download`
not being implemented yet. Let's fix that:

.. code-block:: python

    class Application(app.Application):
        def download(self, version):
            return "http://wordpress.org/wordpress-%s.tar.gz" % version

We determined this by finding `Wordpress's Release Archive <http://wordpress.org/download/release-archive/>`_
and inferring the naming scheme by inspecting various links.  You should now
be able to run the prepare-pristine command successfully: when it is
done, you'll now have a bunch of files in your repository, and they
will be ready to be committed.  Inspect the files and commit (note that the
format of the commit message is a plain Appname Version.Number)::

    git status
    git commit -asm "Wordpress 2.0.2"
    git tag wordpress-2.0.2

.. note::

    Sometimes, ``http://wordpress.org/wordpress-2.0.2.tar.gz`` won't
    actually exist anymore (it didn't exist when we did it).  In this case,
    you'll probably be able to find the original tarball in
    :file:`/mit/scripts/deploy/wordpress-2.0.2`, and you can feed it
    manually to prepare pristine with
    ``wizard prepare-pristine /mit/scripts/deploy/wordpress-2.0.2/wordpress-2.0.2.tar.gz``

Some last house-keeping bits:  now that you have a commit in a repository, you
can also create a pristine branch::

    git branch pristine

From the point of view of the pristine branch pointer, history should be a
straight-forward progression of versions.  Once you have more versions,
it will look something like:

.. digraph:: pristine_dag

    rankdir=LR
    node [shape=square]
    subgraph cluster_pristine {
        a -> b -> c
        a [label="1.0"]
        b [label="1.1"]
        c [label="2.0"]
        label = "pristine"
        color = white
        labeljust = r
    }

Scriptsify
----------

In a perfect world, the pristine version would be equivalent to the scriptsified
version that would actually get deployed.  However, we have historically needed
to apply patches and add extra configuration files to get applications to
work correctly.  Due to the way Git's merge algorithm works, the closer we are
able to reconstruct a version of the application that was actually used, the
better off we will be when we try to subsequently upgrade those applications.

To give you an idea of what the Git history graph will look like when we
are done, here is the graph from before, but augmented with the scripts versions:

.. digraph:: master_dag

    rankdir=LR
    node [shape=square]
    subgraph cluster_pristine {
        a -> b -> c
        a [label="1.0"]
        b [label="1.1"]
        c [label="2.0"]
        label = "pristine"
        color = white
        labeljust = r
    }
    subgraph cluster_master {
        as -> bs -> cs
        as [label="1.0-scripts"]
        bs [label="1.1-scripts"]
        cs [label="2.0-scripts"]
        label = "master"
        color = white
        labeljust = r
    }
    a -> as
    b -> bs
    c -> cs
    a []

First things first: verify that we are on the master branch::

    git checkout master

Check for pre-existing patches in the old application directory,
:file:`/mit/scripts/deploy/wordpress-2.0.2` in the case of Wordpress,
and apply them::

    patch -n0 < /mit/scripts/deploy/wordpress-2.0.2/wordpress.patch

Then, run the following command to setup a :file:`.scripts` directory::

    wizard prepare-new

This directory holds Wizard related files, and is also used by
:command:`parallel-find.pl` to determine if a directory is an autoinstall.

Finally, if you are running a PHP application, you'll need to setup
a :file:`php.ini` and symlinks to it in all subdirectories::

    cp /mit/scripts/deploy/php.ini/wordpress php.ini
    athrun scripts fix-php-ini

.. note::

    As of November 2009, all PHP applications load the same :file:`php.ini` file.

Now commit, but don't get too attached to your commit; we're going
to be heavily modifying it soon::

    git commit -asm "Wordpress 2.0.2-scripts"

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

We now need to make it possible for a user to install the application.
Most web applications have a number of web scripts for generating a
configuration file, so creating the install script involves:

    1. Determining what input values you will need from the user, such
       as a title for the new application or database credentials; more
       on this shortly.

    2. Determining what POST values need to be sent to what URLs.
       Since you're converting a repository, this job is even simpler: you just
       need to port the Perl script that was originally used into Python.

There's an in-depth explanation of named input values in
:mod:`wizard.install`.  The short version is that your application
contains a class-wide :data:`~wizard.app.Application.install_schema`
attribute that encodes this information.  Instantiate it with
:class:`wizard.install.ArgSchema` (passing in arguments to get
some pre-canned input values), and then add application specific
arguments by passing instances of :class:`wizard.install.Arg`
to the method :meth:`~wizard.install.ArgSchema.add`.  Usually you should
be able to get away with pre-canned attributes.  You can access
these arguments inside :meth:`~wizard.app.Application.install` via
the ``options`` value.

Some tips and tricks for writing :meth:`wizard.app.Application.install`:

    * Some configuration file generators will get unhappy if the
      target directory is not chmod'ed to be writable; dropping
      in a ``os.chmod(dir, 0777)`` and then undoing the chmod
      when you're done is a decent workaround.

    * :func:`wizard.install.fetch` is the standard workhorse for making
      requests to applications.  It accepts three parameters; the first
      is ``options`` (which was the third argument to ``install`` itself),
      the second is the page to query, relative to the installation's
      web root, and ``post`` is a dictionary of keys to values to POST.

    * You should log any web page output using :func:`logging.debug`.

    * If you need to manually manipulate the database afterwards, you
      can use :func:`wizard.sql.mysql_connect` (passing it ``options``)
      to get a `SQLAlchemy metadata object
      <http://www.sqlalchemy.org/docs/05/sqlexpression.html>`_, which can
      consequently be queried.  For convenience, we've bound metadata
      to the connection, you can perform implicit execution.

.. todo::

    Our installer needs to also parametrize :file:`php.ini`, which we haven't
    done yet.

To test if your installation function works, it's probably convenient to
create a test script in :file:`tests`; :file:`tests/test-install-wordpress.sh`
in the case of Wordpress.  It will look something like::

    #!/bin/bash -e

    DEFAULT_HEAD=1
    TESTNAME="install_wordpress"
    source ./setup

    wizard install "wordpress-$VERSION-scripts" "$TESTDIR" --non-interactive -- --title="My Blog"

``DEFAULT_HEAD=1`` indicates that this script can perform a reasonable
operation without any version specified (since we haven't tagged any of our
commits yet, we can't use the specific version functionality; not that we'd want
to, though).  ``TESTNAME`` is simply the name of the file with the leading
``test-`` stripped and dashes converted to underscores.  Run the script with
verbose debugging information by using::

    env WIZARD_DEBUG=1 ./test-install-wordpress.sh

Versioning config
-----------------

A design decision that was made early on during Wizard's development was that
the scriptsified versions would contain generic copies of the configuration
files.  You're going to generate this generic copy now and in doing so,
overload your previous scripts commit.   Because some installers
exhibit different behavior depending on server configuration, you should run
the installation on a Scripts server::

    wizard install wordpress --no-commit

The installer will interactively prompt you for some values, and then conclude
the install.  ``--no-commit`` prevents the installer from generating a Git
commit after the install, and will make it easier for us to propagate the
change back to the parent repository.

Look at the changes the installer made::

    git status

There are probably now a few unversioned files lounging around; these are probably
the configuration files that the installer generated.

You will now need to implement the following data attributes and methods in your
:class:`~wizard.app.Application` class: :attr:`~wizard.app.Application.extractors`,
:attr:`~wizard.app.Application.substitutions`, :attr:`~wizard.app.Application.parametrized_files`,
:meth:`~wizard.app.Application.checkConfig` and :meth:`~wizard.app.Application.detectVersion`.
These are all closely related to the configuration files that the installer generated.

:meth:`~wizard.app.Application.checkConfig` is the most straightforward method to
write: you will usually only need to test for the existence of the configuration file.
Note that this function will always be called with the current working directory being
the deployment, so you can simplify your code accordingly:

.. code-block:: python

    class Application(app.Application):
        # ...
        def checkConfig(self, deployment):
            return os.path.isfile("wp-config.php")

:meth:`~wizard.app.Application.detectVersion` should detect the version of the application
by regexing it out of a source file.  We first have to figure out where the version number
is stored: a quick grep tells us that it's in :file:`wp-includes/version.php`:

.. code-block:: php

    <?php

    // This just holds the version number, in a separate file so we can bump it without cluttering the SVN

    $wp_version = '2.0.4';
    $wp_db_version = 3440;

    ?>

We could now grab the :mod:`re` module and start constructing a regex to grab ``2.0.4``, but it
turns out this isn't necessary: :meth:`wizard.app.php.re_var` does this for us already!
If such a module doesn't exist, you should create it and place an equivalent ``re_var()``
function for that language there.

With this function in hand, writing a version detection function is pretty straightforward.
There is one gotcha: the value that ``re_var`` returns as the second subpattern is quoted (the reasons for this
will become clear shortly), so you will need to trim off the last and first characters or
use :mod:`shlex`.  In the case of version numbers, there are probably no escape characters
in the string, so the former is relatively safe.

.. code-block:: python

    class Application(app.Application):
        # ...
        def detectVersion(self, deployment):
            contents = open("wp-includes/version.php").read()
            match = php.re_var("wp_version").search(contents)
            if not match: return None
            return distutils.version.LooseVersion(match.group(2)[1:-1])

:attr:`~wizard.app.Application.parametrized_files` is a simple list of files that the
program's installer wrote or touched during the installation process.

.. code-block:: python

    class Application(app.Application):
        # ...
        parametrized_files = ['wp-config.php']

This is actually is a lie: we also need to include changes to :file:`php.ini` that
we made:

.. code-block:: python

    class Application(app.Application):
        # ...
        parametrized_files = ['wp-config.php'] + php.parametrized_files

And finally, we have :attr:`~wizard.app.Application.extractors` and
:attr:`~wizard.app.Application.substitutions`.  At the bare metal, these
are simply dictionaries of variable names to functions: when you call the
function, it performs either an extraction or a substitution.  However, we can
use higher-level constructs to generate these functions for us.

The magic sauce is a data structure we'll refer to as ``seed``.  Its form is
a dictionary of variable names to a tuple ``(filename, regular expression)``.
If we manually coded one out, it might look like:

.. code-block:: python

    seed = {
        'WIZARD_DBSERVER': ('wp-config.php', re.compile(r'''^(define\('DB_HOST', )(.*)(\))''', re.M)),
        'WIZARD_DBNAME': ('wp-config.php', re.compile(r'''^(define\('DB_NAME', )(.*)(\)), re.M)),
    }

There's a lot of duplication, though.  For one thing, the regular expressions are almost
identical, safe for a single substitution within the string.  We have a function
:meth:`wizard.app.php.re_define` that does this for us:

.. code-block:: python

    seed = {
        'WIZARD_DBSERVER': ('wp-config.php', php.re_define('DB_HOST')),
        'WIZARD_DBNAME': ('wp-config.php', php.re_define('DB_NAME')),
    }

.. note::

    If you find yourself needing to define a custom regular expression generation function,
    be sure to use :func:`wizard.app.expand_re`, which will escape an incoming variable
    to be safe for inclusion in a regular expression, and also let you pass a list,
    and have correct behavior.  Check out :mod:`wizard.app.php` for some examples.

We can shorten this even further: in most cases, all of the configuration values live in
one file, so let's make ourselves a function that generates the whole tuple:

.. code-block:: python

    def make_filename_regex_define(var):
        return 'wp-config.php', php.re_define(var)

Then we can use :func:`util.dictmap` to apply this:

.. code-block:: python

    seed = util.dictmap(make_filename_regex_define, {
        'WIZARD_DBSERVER': 'DB_HOST',
        'WIZARD_DBNAME': 'DB_NAME',
        'WIZARD_DBUSER': 'DB_USER',
        'WIZARD_DBPASSWORD': 'DB_PASSWORD',
    })

Short and sweet.  From there, setting up :attr:`~wizard.app.Application.extractors` and
:attr:`~wizard.app.Application.substitutions` is easy:

.. code-block:: python

    class Application(app.Application):
        # ...
        extractors = app.make_extractors(seed)
        extractors.update(php.extractors)
        substitutions = app.make_substitutions(seed)
        substitutions.update(php.substitutions)

Note how we combine our own dictionaries with the dictionaries of :mod:`wizard.app.php`, much like
we did for :attr:`~wizard.app.Application.parametrized_files`.

With all of these pieces in place, run the following command::

    wizard prepare-config

If everything is working, when you open up the configuration files, any user-specific
variables should have been replaced by ``WIZARD_FOOBAR`` variables.  If not, check
your regular expressions, and then try running the command again.

When you are satisfied with your changes, add your files, amend your previous
commit with these changes and force them back into the public repository::

    git status
    git add wp-config.php
    git commit --amend -a
    git push --force

Ending ceremonies
-----------------

Congratulations!  You have just implemented the installation code for a new install.
If you have other copies of the application checked out, you can pull the forced
change by doing::

    git reset --hard HEAD~
    git pull

One last thing to do: after you are sure that your commit is correct, tag the new
commit as ``appname-x.y.z-scripts``, or in this specific case::

    git tag wordpress-2.0.4-scripts
    git push --tags