I've been reading some posts about large-scale, or "enterprise", applications involving Javascript.
As a brief aside I'll note that to me, to call software "enterprise" is
to pronounce a terrible condemnation upon it,
to evoke my memories of awful intranet applications,
sprawling old distributed client-server systems,
horrible user interface designs obsolete at their debut,
and tremendous, sometimes astronomical costs of purchase and deployment.
Though I'm still basically doing "enterprise software development"
I never say those words to myself. But that's just a personal problem.
It's funny how words get poisoned.
Anyway,
here's
an interesting article I just read this morning.
I'm really impressed at some of the apparent capabilities of those
tools the article discusses.
However, the "standard folder structure" described therein
evinced a distinctly different approach to the management
of client-side code (well, client-side everything, by extension)
than what I've come to think of as the One True Way.
It's always interesting to read something like that and
have one's own notions illuminated by glaring contrast with someone else's.
I'm not going to say my ideas are better of course, but I think they're
fundamentally different.
It's going to be a little hard to describe why, because I don't
think I'm equipped with the terminology necessary to discuss the
subject; I'm not even sure I know exactly what the subject is.
Nevertheless, as I've been preparing to do a quick presentation
about the little
wallflower jQuery plugin
that I wrote a few weeks ago, I've been forced to think not only
about the way I actually go about designing and arranging and managing
my application as it grows, but also about the
way I think I should be doing it.
Back a really long time ago, in the early '90s, one of the
trendy must-read business books of the time was Peter Senge's
The Fifth Discipline.
I guess it says something about how business/management improvement
trends go that you can pick up copies of that book now for the cost
of shipping. In other words, it's almost literally worthless.
But it's a pretty good book, as things like that go.
Senge's an interesting guy, an engineer by training who got all fired
up about the topic of "organizational learning", and that's what he's
been doing at MIT's Sloane School of Management for a couple decades.
The book is basically about that: what it means for an organization to
learn.
Much of the material in The Fifth Discipline is sourced from
the work of two other Sloane School researchers, Chris Argyris and
Donald Schön.
Since the 1950's, those two have been doing some really interesting
work on the theory and mechanics of learning, both individual and
organizational.
Check out
Dr. Argyris's Wikipedia page
for a brief taste of what he's written about.
One of the lasting concepts I took away from
one of his books
(which, you'll note, is not considered worthless)
is an idea that Senge also latched onto, that of
recognizing and contemplating the distinction between theory in use
and espoused theory.
An espoused theory is something you type into a blog post about how
everyone should architect their enterprise application.
A theory in use is how a development team actually does it.
(Seriously, that book is one of those eye-opening, mind-opening
books that you just can't forget, providing enormous insight into
all the stupid things that go on in organizations large and small.
If everybody read and believed in the stuff Argyris and Schön have
written about, the world would be a better place.)
Now I'm not going to pontificate about how you should architect your
web applications.
My point in bringing up those books and those ideas is that
writing about how one does things can be pretty humbling.
I know that the way I actually work isn't as close as I'd like it to
be to the grandiose fantasies I entertain about The Perfect Application.
More than that, however, what I learned from those books is that
I should not only be aware of the ways my actual work doesn't live
up to my ideals, but also to be aware of the reasons I think those
ideals are ideal.
What do I want from my application design, and why do I want it?
If I were able to fully realize my ambitions for perfect design,
would I perform some sort of test to see if I actually was getting
what I wanted?
That's something that's generally left out of most process
improvement activities — and I mean process improvement
activities as actually done, not as they're described in books
and blog posts.
How many teams have you worked on that, after taking up some
trendy activity like stand-up meetings, have actually applied
an even vaguely scientific testing method to see whether the
new activity was helping?
How many really analyze the goals for starting their process
improvement in the first place, and subsequently test to
see if achieving the goals helps out their business bottom line?
I think the same rigor should apply to technical
processes like application construction and code management,
but in my experience that's not so common.
One of the difficulties I'm having with this subject is that
I don't know what to call the central activity.
That activity is the construction of the application code
behind an application, and the design for how that
construction should proceed.
I use the term "construction" because I've been driving
through a bit of it for several months now on my way into
and out of town.
Civil engineering projects involve an outward architecture:
somebody designs what a bridge should end up looking like,
the shape of the arch, the look of embankments, the decorations
on railings and lane dividers, and all those other things
that'll be mostly ignored by the public when the project is
complete.
However, I suspect (knowing essentially nothing at all about
what civil engineers do) that a whole lot of work goes into
the design of the construction effort itself.
I'm sure that there's a tremendous amalgamation of
experience that instructs thousands of planning decisions
for a big roadway project.
Those must be planned with a rigor that I've never
seen even distantly approached in any software project I've
experienced. Maybe I've just been unlucky, but I don't
really think so.
What do I call, then, the design and planning of software
implementation,
and by extension the theory of that design? The
term "application design" seems to be understood as
being more about what an interface looks like and
what application users should experience.
"Coding practices" to me means "tabs or spaces"
and details like that.
I'm going to use the term "application implementation
design" because I can't think of anything else.
I'm absolutely not an expert, of course, as should be clear
by this admission that I literally don't know what it is
that I'm talking about, or at least what to call it.
Now back to Javascript.
What is it that I want from the way I type in Javascript code?
To put it another way, what is it that I want to be true
at some point in the future when the surrounding application
is, if not "finished", something large and functional and
under heavy use?
Of course it's clear that I want it to work and to work well.
That means it doesn't break in small browser windows,
it doesn't slow down pages, it doesn't annoy users,
and so on, satisfying all those little de rigueur
stipulations we twitter about constantly.
But I also think there are aspects to the application that
are of great importance to the developer
experience.
What should it be like to add a new feature?
To revise an old feature?
To fix bugs?
To cope with off-the-shelf framework and tool upgrades?
And once I've answered those questions, why do I
think those are good goals?
In other words, why would it be good for the process
of fixing a bug to play out in some particular way?
Thought of that way, it's a pretty big topic, and though
I've been doing it for almost 30 years now it's challenging
to tease apart good experiences I recall from the bad ones.
Please Either Shut Up Or Talk About Javascript
Here's the point I'm trying to make.
My own goals for the shape of my application code
— and not just where the files are in folders, and what the folders
look like, and how they're arranged, but the underlying ways in
which the code all fits together —
are mostly summed up like this:
I want to minimize the number of dimensions of code expansion
that result from a given increment in functionality.
In other words, I want it to be the case that when new functionality
is called for, a developer should have to type in as little
new code as possible.
OK well that's pretty obvious; that's what
everybody wants.
The tough part, however, is being sensitive about what sort
of new code we consider to be "matter of course" new code —
code that we think of as having to be written due to laws of nature.
That's tough because it's easy to grow numb to some things that
really aren't laws of nature, but consequences of application
implementation choices.
For example, in my current application, when I need a new
database table or changes to an existing one, I type in a
little SQL.
To a lot of people nowadays, that seems horrible (or at least that's
an impression I've formed).
The framework should do that!
Well, maybe; I'm aware of how much that part of my world deviates
from some current fashion, but in that particular case I actually
prefer it that way.
In other words, I'm aware of the fact that I've
chosen
a scheme that requires me to type in some SQL manually,
but I know what my reasons are and I'm OK with that.
In a former position, I worked on a really
large web application: a huge pile of Java on the server
side, almost a thousand separate .jsp files, and lots
and lots of Javascript.
The application implementation design in that situation
was such that implementing new pages meant, more-or-less
as a matter of course, writing Javascript code for all
the widgets on each page.
Though there was some re-use of some components,
like confirmation dialogs for example,
the Javascript required to integrate a confirmation
dialog with a particular new page was pretty much
unique to the page.
With relatively little experience dealing with
an application of that magnitude (and at the time
the development effort was really ramping up,
back in 2003 or so, there probably weren't a lot of
people around anywhere who had much perspective on what a
huge AJAX-heavy application would be like for the
developers involved),
all of that per-page plumbing seemed pretty
natural to me.
I got pretty good at sort-of automatically typing
in the stuff I needed, knowing how to think ahead for
what a particular sort of page was likely to need.
Lots of the script code was directly in the pages
because it was often reliant on bits of information
that'd be dropped into the code by the server-side
template engine (that is, via JSP JSTL/EL expressions).
How else could such code possibly be written? Well
I didn't actually ponder that question very often,
at least at first.
Probably sometime around the hundredth new page,
however, I started to get this nagging feeling that
something was wrong.
At that point, I started to see the value (now so
glaringly apparent) in common modular Javascript components.
And as I started writing those, I started to also
see the need for some sort of mechanical way that
the chains of component inter-dependencies could
be dealt with automatically.
If there were some way that the component
dependencies could be declared, that'd
factor out a big headache from page construction
and maintenance.
We never got there, but it sure seemed like
one of those "ideals".
When I started work on my current application,
I spent a lot of time looking around at various
dependency management tools and modularization
techniques. Many of them looked great, but I
didn't really dive in because I was also immersed
in the problems of getting an application concept
started from scratch.
I'd never really done that,
so I had to spend a lot of time thinking both
about the outward-facing design as well as
the application implementation design; the
first problem generally got the lion's share
of my anxiety.
I was also spending a large amount of time thinking
about server-side application implementation, so
the Javascript module problem didn't get a lot of attention.
All this time, another "ideal" banging around in my
head was a desire to really get away from in-page
Javascript, and to get much closer to a universally
unobtrusive way of creating pages.
My server-side investigations had led me to
the Stripes framework
which seemed like some magical heaven-sent gift
written by people who shared most of my opinions about that problem space, and
were a million times smarter than me too.
One of the nice things Stripes provides on the
page construction side of things is a simple but
powerful layout templating system.
Once I started using that, it became clear that if
we were careful the pages we'd be writing for
application features could be really succinct.
The layout templates could handle just about all the
boilerplate page code that, in my previous position,
had definitely seemed like a law of nature.
Well stuffing all the boilerplate into layout templates
brought an implication to the domain of managing the
Javascript code that supported fancy page behaviors.
If I were going to have an implementation design
based on the idea that every page would somehow
grab the components it needs, then dependency management
or not there'd have to be some communication of those
needs between pure page code and the layout templates;
otherwise, how would the Javascript end up
at the browser?
For all their convenience, the mechanisms available for
pages to "talk to" the Stripes layout template code are a little
crude.
On top of that, of course, the templates themselves are JSP
code, and I really don't like writing code that has to "think"
very much in JSP; it's hideous.
(Is it a law of nature that such page-template
communication should be hideous? Probably not ...)
This situation drove me down the path I'm on now.
The more I tinkered with it, and the more that some early
deadlines loomed in getting the application put together,
the more it looked like the simplest thing to do would
be to just stick a common block of script tags in the
layout template, and have every single page just pull
in essentially all the Javascript I had.
I had concerns about performance of course,
naturally imagining the worst about what that could lead to
in the future.
However, I had to get some pages working and make it
possible for other people on the team to write pages too,
so it was the simplest place to start.
The implementation of my unobtrusive behaviors
proceeded, generally on an as-needed basis.
A new page would need something, and I'd cobble
together some way for the HTML to announce that
need (a new class name, or grouping under a
particular element id, or whatever) and write
the Javascript to deal with it.
The code evolved along the lines of little
stanzas like this:
$where.find('.dialogLink').each(function() {
var $link = $(this);
// ...
});
Each page feature, in other words, had a chunk
of code that would look for the tell-tale mark
in the HTML indicating the need for some
activity. The "$where" variable was the DOM
root under which I needed the work done, and
was generally either the page body as a whole,
or else a piece of the page into which new content
had just been dropped.
I was lumping all this code into a single routine
that would be applied upon page load, and also
when page fragments were loaded dynamically.
As you can imagine, this function became
ridiculously big after only a short while.
I knew it was awful, but it did work after all
and I had other things to worry about.
To address the problem of the insane size of my
"do all the behaviors" script, I eventually broke
up the code into lots of separate files, basically
organized around the page features they supported.
That helped a lot with keeping the code base
manageable and maintainable; fixing a behavior meant
a change only to a small file, not touching any
other Javascript in the repository.
However, I couldn't possibly stand the idea of extending
my layout template to include thirty or forty
separate script tags, even if they were small.
To solve the script tag problem, I leveraged the fact
that as a Java application I had a build system anyway.
I added an Ant task to the build that leveraged another
tool we were also incidentally using, the Freemarker
template engine.
The Ant task processes a small Freemarker template
file that is a Javascript skeleton.
Freemarker directives in the template call out to
routines in the Ant task that find those little script
files in the source tree and supply them for inclusion
in the Freemarker output.
The result is a horrible-looking yet syntactically correct
monolithic Javascript file that gets cranked through
YUI compressor (giving the unanticipated but welcome
benefit of a Javascript syntax check at build time)
and dropped into the final .war file.
I even concatenate the jQuery and jQueryUI libraries
into my script, so ultimately what resulted was something
that I would have considered completely insane just
a few months previous: every single one of the pages
of this application import exactly one script, that
being the sum total of all the Javascript
I've got.
That's obviously a disaster, right?
One huge (about 250K at present) script applied to
every single page?
All those jQuery selector searches eating up
time when 90% of them aren't going to actually do
anything?
Well, yes, early on we did have some performance
problems.
Part of that was foolish ignorance, mostly around
my failure to use smart event handling (bubble-based
handlers with jQuery "live" and "delegate").
Part of it was badly-written selectors, like that
".dialogLink" above.
So we did spend time worrying about that and tuning
the script, and I experienced a lot of anxiety at
the prospect of having to break it all up.
Eventually, however, we crossed a line due to
good ideas from my co-workers and good ideas from
blogs and forums, and I realized that my gigantic
monolith of code really wasn't posing a performance
problem after all, at least not simply due to its
bulk.
I've now got good selectors for everything, and
dealing with all that Javascript takes only a
very short time as my pages are painted by the
browser.
One of the benefits of having just one file, of course,
is that it's the only script a client browser has to
cache.
The mechanism supporting all my unobtrusive page features
was still a mess, however, and that's what led me to type
in the wallflower code.
I wanted to see what a more organized approach might look
like.
If and when I actually try and convert my real code to
use wallflower, I'll probably stick with the idea of
the build-time "assembly" step, so that my individual
wallflower features will live on as separate source
files.
But it's been a really long time since I've spent any
time at all thinking about a component dependency
manager. Why should I? I've "solved" the
dependency problem (without explicitly planning to do
so, and in fact without really being aware that I was
doing it) by having exactly one common dependency for
every page on the site.
That's just hard-coded into the template, and I'm done.
New page with new needs? Sure, I'll just add a
new little behavior script file, and it works.
Given how fast things currently run (and while it's truly
breathtaking how fast modern Javascript interpreters
are, this stuff works fine even in IE7), I doubt I'd
see problems if I had 3 times as many behavioral
features to support.
I doubt I'll get there, of course, mostly because a
site with that many whiz-bang page features would
be terribly annoying (I'm looking at you, mint.com).
Another unanticipated benefit — at least, I think it's
a benefit — of all this is that it's
led me to having a much more disciplined approach
to adding fancy behaviors to the site.
In my previous job, because the way code was
written was almost certain to involve so much
specific new code to be created for everything,
having unique behaviors on individual pages for
no really good reason was pretty common.
There was nothing in the process standing in the
way of that, after all.
Now, when thinking of new application features,
I tend to think in terms of things I can already
make pages do.
Yes, there are two or three "special" parts of my
app that do things no other pages do, and I confess
to a few pages with on-page Javascript (I'm working
on it, really), but keeping things under control
I think really helps with maintaining consistency,
and that's good not just for the development team
but also for the end user.
So that brings me to this morning.
I've been typing in this ridiculously long
ramble for a while, but not until I saw a couple
of posts about "enterprise" components did a
unifying theme really click in my head.
I'm delighted that people are finding success with
smart dependency management schemes, and I'm
sure I've got a lot to learn about large-scale
web applications from other people approaching
the problem in more well-informed ways than I did.
(That'd probably be just about anybody doing it.)
However, there's a lot I like about the world
I've made for myself.
What are the "laws of nature" that I'm assuming?
Where are all my unnoticed wastes of time?
Would one giant script file really work out in
an application of the scale that my previous
position involved?
I'm not really sure at the moment.
I also don't know whether what I'm doing is something
that lots of people do, that's been written about
elsewhere, possibly extensively, and possibly with
well-reasoned heavy criticism.
If you know, and you're not too exhausted after
reading all this stuff, leave me a comment or two.