coleslaw/docs/hacking.md

Coleslaw: A Hacker's Guide

Here we'll provide an overview of key concepts and technical decisions in coleslaw and a few suggestions about future directions. Please keep in mind that coleslaw was written on a lark when 3 friends had the idea to each complete their half-dreamed wordpress replacement in a week. Though it has evolved considerably since it's inception, like any software some mess remains.

Core Concepts

Data and Deployment

Coleslaw is pretty fundamentally tied to the idea of git as both a backing data store and a deployment method (via git push). The consequence is that you need a bare repo somewhere with a post-recieve hook. That post-recieve hook (example) will checkout the repo to a $TMPDIR and call (coleslaw:main $TMPDIR).

It is then coleslaw's job to load all of your content, your config and templates, and render the content to disk. Deployment is done by moving the files to a location specified in the config and updating a symlink. It is assumed a web server is set up to serve from that symlink. However, there are plugins for deploying to Heroku, S3, and Github Pages.

Blogs vs Sites

Coleslaw is blogware. When I designed it, I only cared that it could replace my server's wordpress install. As a result, the code until very recently was structured in terms of POSTs and INDEXes. Roughly speaking, a POST is a blog entry and an INDEX is a collection of POSTs or other content. An INDEX really only serves to group a set of content objects on a page, it isn't content itself.

This isn't ideal if you're looking for a full-on static site generator. Content Types were added in 0.8 as a step towards making coleslaw suitable for more use cases but still have some limitations. Any subclass of CONTENT that implements the document protocol counts as a content type. However, only POSTs are currently included on INDEXes since their isn't yet a formal relationship to determine what content types should be included on which indexes.

The Document Protocol

The document protocol was born during a giant refactoring in 0.9.3. Any object that will be rendered to HTML should adhere to the protocol. Subclasses of CONTENT (content types) that implement the protocol will be seamlessly picked up by coleslaw and included on the rendered site.

All current Content Types and Indexes implement the protocol faithfully. It consists of 2 "class" methods, 2 instance methods, and an invariant.

Class Methods:

Since Common Lisp doesn't have explicit support for class methods, we implement them by eql-specializing on the class, e.g.

(defmethod foo ((doc-type (eql (find-class 'bar))))
  ... )

• discover: Create instances for documents of the class and put them in in-memory database with add-document. If your class is a subclass of CONTENT, there is a default method for this.
• publish: Iterate over all objects of the class

Instance Methods:

• page-url: Generate a unique, relative path for the object on the site sans file extension. An :around method adds that later. The slug slot on the object is generally used to hold a portion of the unique identifier. i.e. (format nil "posts/~a" (content-slug object)).
• render: A method that calls the appropriate template with theme-fn, passing it any needed arguments and returning rendered HTML.

Invariants:

• Any Content Types (subclasses of CONTENT) are expected to be stored in the site's git repo with the lowercased class-name as a file extension, i.e. (".post" for POST files).

Current Content Types & Indexes

There are 5 INDEX subclasses at present: TAG-INDEX, MONTH-INDEX, NUMERIC-INDEX, FEED, and TAG-FEED. Respectively, they support grouping content by tags, publishing date, and reverse chronological order. Feeds exist to special case RSS and ATOM generation. Currently, there is only 1 content type: POST, for blog entries.

I'm planning to add a content type PAGE, for static pages. It should be a pretty straightforward subclass of CONTENT with the necessary methods: render, page-url and publish. It could have a url slot with page-url as a reader to allow arbitrary layout on the site. The big question is how to handle templating and how indexes or other content should link to it.

Templates and Theming

User configs are allowed to specify a theme, otherwise the default is used. A theme consists of a directory under "themes/" containing css, images, and at least 3 templates: Base, Index, and Post.

Coleslaw uses cl-closure-template exclusively for templating. cl-closure-template is a well documented CL implementation of Google's Closure Templates. Each template file should contain a namespace like coleslaw.theme.theme-name.

Each template creates a lisp function in the theme's package when loaded. These functions take a property list (or plist) as an argument and return rendered HTML. Coleslaw defines a helper called theme-fn for easy access to the template functions. Additionally, there are RSS, ATOM, and sitemap templates coleslaw uses automatically. No need for individual themes to reimplement a standard, after all!

Plugins

Coleslaw also encourages extending functionality via plugins. The Plugin API is well-documented and flexible enough for many use cases. Do check the API docs when contemplating a new feature and see if a plugin would be appropriate.

The Lifecycle of a Page

• (load-content)

A page starts, obviously, with a file. When coleslaw loads your content, it iterates over a list of content types (i.e. subclasses of CONTENT). For each content type, it iterates over all files in the repo with a matching extension, e.g. ".post" for POSTs. Objects of the appropriate class are created from each matching file and inserted into the an in-memory data store. Then the INDEXes are created by iterating over the POSTs and inserted into the data store.

• (compile-blog dir)

Compilation starts by ensuring the staging directory (/tmp/coleslaw/ by default) exists, cd'ing there, and copying over any necessary theme assets. Then coleslaw iterates over the content types and index classes, calling the publish method on each one. Publish iterates over the class instances, rendering each one and writing the result out to disk with write-page (which should probably just be renamed to write-file). After this, an 'index.html' symlink is created to point to the first index.

• (deploy dir)

Finally, we move the staging directory to a timestamped path under the the config's :deploy-dir, delete the directory pointed to by the old '.prev' symlink, point '.curr' at '.prev', and point '.curr' at our freshly built site.

Areas for Improvement

Render Function Cleanup

There are currently 3 render-foo* functions and 3 implementations of the render method. Only the render-foo* functions call write-page so there should be some room for cleanup here. The render method implementations are probably necessary unless we want to start storing their arguments on the models. There may be a different way to abstract the data flow.

User-Defined Routing

There is no reason coleslaw should be in charge of the site layout or should care. If all objects only used the slug slot in their page-url methods, there could be a :routing argument in the config containing a plist of (:class "~{format string~}") pairs. A default method could check the :class key under (routing *config*) if no specialized page-url was defined. This would have the additional benefit of localizing all the site routing in one place. New Content Types would probably pushnew a plist onto the config key in their enable function. This has been implemented on the branch user-defined-routing.

Better Content Types

Creating a new content type is both straightforward and doable as a plugin. All that is really required is a subclass of CONTENT with any needed slots, a template, a render method to call the template with any needed options, a page-url method for layout, and a publish method.

Unfortunately, this does not solve:

1. The issue of compiling the template at load-time and making sure it was installed in the theme package. The plugin would need to do this itself or the template would need to be included in 'core'. Thankfully, this should be easy with cl-closure-template.
2. More seriously, there is no formal relationship between content types and indexes. Consequentially, INDEXes include only POST objects at the moment. Whether the INDEX should specify what Content Types it includes or the CONTENT which indexes it appears on is not yet clear.

New Content Type: Shouts!

I've also toyed with the idea of a content type called a SHOUT, which would be used primarily to reference or embed other content, sort of a mix between a retweet and a del.icio.us bookmark. We encounter plenty of great things on the web. Most of mine winds up forgotten in browser tabs or stored on twitter's servers. It would be cool to see SHOUTs as a plugin, probably with a dedicated SHOUT-INDEX, and some sort of oEmbed/embed.ly/noembed support.

Incremental Compilation

Incremental compilation is doable, even straightforward if you ignore indexes. It is also preferable to building the site in parallel as avoiding work is better than using more workers. Moreover, being able to determine (and expose) what files just changed enables new functionality such as plugins that cross-post to tumblr.

Git's post-receieve hook is supposed to get a list of refs on $STDIN. A brave soul could update our post-receive script to figure out the original hash and pass that along to coleslaw:main. We could then use it to run git diff --name-status $HASH HEAD to find changed files and act accordingly.

This is a cool project and the effects are far reaching. Among other things the existing deployment model would not work as it involves rebuilding the entire site. In all likelihood we would want to update the site 'in-place'. Atomicity of filesystem operations would be a reasonable concern. Also, every numbered INDEX would have to be regenerated along with any tag or month indexes matching the modified files. If incremental compilation is a goal, simply disabling the indexes may be appropriate for certain users.