A while back I was chatting to Arthur and bitching about all the various blog systems. To date, I've tried a wide cross section of them and found them all to be inadequate. Arthur's proposed solution was to write his own. I shied away from that notion until a few days ago. Then I finally caved and started thinking about doing my own.
It's only in the past year or so that I've viewed blogging as a public activity. Previously, I used it more as a personal journal even if others peeked in. My first exposure was the trashy, but extremely functional Xanga. I didn't care very much about web standards or design early in high school, so Xanga was quite good to me. It was free, customizable, and all my friends used it. My relationship with Xanga went sour when I tried to screen-scrape all my posts with WWW::Mechanize. It worked ok, but it made me realize how shitty the HTML was and that there could be better.
Sometime after that I had a brief stint with LiveJournal. I lacked motivation to keep it up, but I liked the external clients, and open source nature of LJ.
I really dislike writing markup, and I also despise looking at markup. Neither was avoidable with other systems. Writing markup by hand was tedious and often resulted in erroroneous markup. Once you've written the markup, it's nice to know that the data is transformable with XSLT.
Writing a blog with just plain text and not blog management system is also a lost cause. Once you get over a certain number of posts, it's hard to manage the permalinks by hand. There's also lots of duplicated HTML that would best be abstracted into templates. On top of that, it's important to have a way for users to give feedback about your articles.
With these points in mind, I looked to create my ideal blog management system. Plain text will be used to manage storage for maximum future-proofing, and minimal software requirements. The implementation language of choice is a hybrid of Perl, Bash, and anything that fits in the future. For the sake of simplicity and at the expense of customization, I've opted for enforcing a hardcoded directory hierarchy:
articles/ YYYY/ MM/ DD/ title/ index.draft index.text index.html images stylesheets .config/ override default options .meta/ tags, word count, etc
This keeps each individual article self-contained. Rather than model a data structure for what an article is, I've made the structure implicit in the way the data is stored. This idea separates the system from the data so that the data makes sense standalone. You could argue that I duplicate this into a schema for some database, but why would I need one?
The .text and .html is carried over from my earlier text-based blog system. The best web word processor is Markdown. It lets you write in any editor you want and has syntax rules that become natural almost instantly. This freed my eyes from staring at markup and also gave me very clean XHTML and other formats as needed. The editor and external client was solved free of charge by Emacs and ssh. Intermdiate articles can be saved with a .draft extension.
As far as comments go, I plan to write it as a separate system with an public API. There's no reason to make a blog-specific comment system when I can create one that can be used anywhere. This also removes a huge chunk of complexity involved with fighting spam that I really don't feel belongs in a blog system. This guy will come later.
As I'm writing this, one annoying piece of the system has cropped up: the templating system. This puts me in a sticky spot because using an existing templating system would introduce more bloat than I'd like, but writing my own simplistic templating system sounds tricky. At the moment, all I need is a declarative way of piecing together snippets of HTML. My favorite option so far is to match the path of a file to be published against a set of rules to determine which template to use. The actual template will be a simple perl script that takes the content to be embedded as input and spit out a laid out version. The downside of this is the assumption that there will be only be one snippet to be inserted within a template. This is limiting, but can be fixed by adding options so that the layout script knows what is given. For example:
perl article.layout --content=articles/2008/01/29/foo/index.html --header=includes/head.html --footer=includes/foot.html
it'd be nice to have a generator script to scaffold a generic template:
script/generate layout article [header,content,footer]
Update: I've been working with TemplateToolkit for work, and I've found it to be very easy to use, but not bloaty.
This one's a bit of a bummer because none all the solutions I come up with are tradeoffs. The slowest and most accurate solution is to have the author of articles fully specify the headline and blurb. There could be separate markup at the top of a file, or put into separate files. Basically, this involves adding more structure to the writing and more logic to handle this structure. This idea doesn't mix well with my goal of uber-simplicity.
Instead I opted for an approach that sacrifices some generality. I notice that all articles that I like usually involve an engaging headline followed by an abstract. So I figure it's fair for me to look through either the source .text or markdowned .html for an h1 headline, and the first paragraph. The exit strategy is straightforward because it's a simple map of a function over a list or articles. No biggie.
Pre-generating and Flattening out dyanmic content will only get me so far. Hopefully it'll be far enough :)
Performance-wise, I want this system to be fast. How fast? Here are my design goals:
|View an article||O(1) - as fast as it takes to serve a static page||O(2n + m) - where n is the number of words in an article. Each article will be kept in both Markdown and XHTML format. The extra m factor stands for the extra templating markup, but is pretty insignificant. I'm trading off a bit of storage for best-case speed.|
|Write an article||O(1) - Emacs ftw!||N/A|
|Publish articles||O(n) - where n is the number of articles to publish. The script uses 'GNU find' and command line options to only publish and generate needed content||Some meta information about the articles are created, so a bit of extra space is needed. Again, pretty insignificant.|
For now, I have no ambition of making this system scale, but I'm smug with the knowledge that it will scale. ZFS immediately comes to mind. I can imagine publishing as a bottleneck, but clever uses of 'find', 'locate' and meta information can minimize these operations. Another storage consideration is how wasteful it would be to duplicate common HTML snippets across articles. Headers, footers, and sidebars are good examples of this. My take on this issue is: Care when it matters. My own personal use case is to have meaty articles with minimal fluff snippets. For me, the benefits of having blazing fast page views outweighs the negative of wasted storage. I call it "Lazy Pre-Evaluated Caching", or LPEC for short. On the other hand, I can see how saving space can lead to more articles being stored in memory. My exit strategy, should it ever be needed, is to LPEC the most popular pages, and leave rarely viewed pages in un-markdown'ed source .texts.
For maintenance, I want it all. I don't want the code to be obsolete, and I definitely don't want my data to suffer bitrot. By using a hybrid of languages and frameworks, I get the best of each and guarantee easy replace of any subsystem if a better solution arises.
Update: That last bit was too vague for my tastes. Most of the system will be written in Perl so I can beef up some of my Perl-fu. I doubt anything will feel slow, but I might use XS just to play with it.
Common tasks can be described with Rake and to automate these tasks, I'll periodically run them from cron. Here's what I have in mind for a basic publish_all task
# month hour day-of-month month day-of-week command 0 4 * * * ~/portfolio/publish_all
Update: There will be options in the 'publish' command that allows users to filter what they want to publish.
I plan to use the file hierarchy and file names as an encoding of the data. However, from a programatic perspective, it makes more sense to have an API to manipulate this data. This also makes it easy to transform the data into any other storage backend as needed.
An article is a source un-markdowned entry. It keeps state for everything that describes an article and is implemented as a self-contained folder.
Object used to transform and layout objects. Each template type expects arguments of the things they need in order to lay out themselves. For example, an ArticleHTMLTemplate may require 'stylesheets', and 'articles', whereas an ArticleXMLTemplate may only require 'articles'.
I didn't want to call it controllers, but I'm thinking of it in a MVC kind of way.
I'm keeping some bullets here for updating later.
thinking about the different use cases definitely revealed a pattern for processing articles. A key observation I noticed is that there are two types of operations: taking an article as input, doing something, and spitting out a result. The other is to take a list of articles as input, do something, and spit out a single result.
using Moose as the object system to represent an article. Update - I've decided against this because it's too fancy.
markdown has an annoying feature that it will wrap multi-line 'li' content in a 'p' tag. I fixed this by styling 'li p' to be inline. This is still valid xhtml and still makes markup sense.
rebuilding the pages that are indicies for lists of articles requires rebuilding all of the indicies. This does scale, so I've been toying with the idea of keeping a B-tree of published entries. There needs to be meta files to keep track of how many entries are in what pages. When a page gets too full, it gets split into two separate pages; When a page gets too empty, it gets joined with it's neighbor. When publishing, it means that only specific pages are being rebuilt, rather than all index pages.