A couple of days ago I did some significant work in authorization in one of my apps, involving creating a Groups class with an HABTM relationship to Person, so I could assign roles to people a group at a time. It all worked out great, and I pushed the product to GitHub. The next day, my collaborator wrote in that my recent contribution broke 119 specs.

I pulled and retested the code, and everything worked perfectly. WTF? After a bit of investigation, I discovered that the specs worked great when I ran 'autotest' or 'spec spec', but that 119 specs broke when I ran the exact same spec suite with 'rake spec'! Double WTF.

Setting constants at class loading

Ultimately, I tracked the problem down to this line and method, in Person.rb:
class Person << ActiveRecord::Base
ADMIN_GROUP = Group.find_by_name('Admin')
def admin?
groups.include? ADMIN_GROUP
end
end


I consider a person an administrator if they are a member of this group, and I was loading it as a constant at the class level in order to avoid having to query the database again every time Person#admin? is called. This worked just fine for me, both in the application, and every time I ran Person#admin?.

But, remarkably, ADMIN_GROUP does not get initialized correctly when I run the tests via rake. I found this via the ruby debugger, running in this particular spec in spec/models/person_spec.rb:


describe Person do
it "should load an admin user from fixture" do
debugger
people(:admin).should be_admin
end
end


When I run the specs and evaluate Person::ADMIN_USER, I get very different results depending on which spec runner I'm using:

Running 'spec spec/models/person_spec.rb':


[11:17:54 CITAlumni]$ spec spec/models/person_spec.rb
spec/models/person_spec.rb:64
people(:admin).should be_admin
(rdb:1) eval Person::ADMIN_GROUP
#


Running 'rake spec SPEC=spec/models/person_spec.rb':


[11:17:13 CITAlumni (48c51f1...)]$ rake spec SPEC=spec/models/person_spec.rb
(in /Users/evan/Development/Ruby/CITAlumni)
FF.....spec/models/person_spec.rb:64
people(:admin).should be_admin
(rdb:1) eval Person::ADMIN_GROUP
nil


How very interesting ... when I use rake, that constant initializes to nil. At some point, I'll actually get around to figuring out why this is so different when the specs are run via rake. In the meantime, the fix was easy:

The Solution

The fix was just to refactor ADMIN_GROUP as a class method with a memoized instance variable. This will at least limit DB queries for the admin group to one per page load; not quite as good as a single DB hit when the class is first loaded, but still a major improvement over querying for the admin group every time Person#admin? is called. I moved it to the Group class at the same time, which was probably the right place for it in the first place:


#app/models/group.rb:
class Group < ActiveRecord::Base
def self.admin_group
@admin_group ||= self.find_by_name('Admin')
end
end

#app/models/person.rb
class Group < ActiveRecord::Base
def admin?
groups.include? Group.admin_group
end
end


And this worked just fine in all environments, solving the problem with 'rake spec'.

The Moral

Be careful with depending on behavior that occurs only during the loading of classes, as it can be environment-dependent!

If anyone out there with uber Ruby skills knows exactly why running specs via rake prevents that class variable from loading correctly, please enlighten us in comments!

The problem YAML file looked like this:

spec/fixtures/people.yml (broken)


one:
funky_database_id: 0000012345

two:
funky_database_id: 0000012346


The trouble with this is that yaml interprets those values as integers, not strings, and Person#funky_database_id is a string column. So Ruby conveniently loads the value as an integer and runs to_s on it before inserting. Worse, because these start with 0, they get translated from octal. So people(:one).funky_database_id comes out "5349". Definitely not what I wanted.

This works as expected:

spec/fixtures/people.yml (fixed)


one:
funky_database_id: "0000012345"

two:
funky_database_id: "0000012346"


An excerpt:

In the photo, the young person’s eyes are brown and kind-looking. She is in need of financial help. A new Web site that brings together the charitable minded and those in need has posted the details of her request.

This is not one of those arrangements where donors can sponsor a needy child or a sorghum farmer in the developing world. The person asking for help is a 21-year-old neurobiology major at Harvard, and she is requesting a loan from Harvard alumni.

Specifically, I needed to put annotations in the description of a work order object that said, for example "this work order was escalated from Problem Report 293. This was done in a create action that redirected at the end and never rendered a view, so I really did need to generate that link in the controller. And for consistency with the rest of the application, I wanted to generate the link with link_to(@task).

Now, ActionView::Helpers::UrlHelper is not loaded in a Rails controller, even if you've put helper :all in application.rb (application_controller.rb in newer versions). So, when I tried to use link_to in the controller, I got an error:

NoMethodError: undefined method `link_to' for #
/Users/evan/Development/Ruby/eclipticdb/app/helpers/tasks_helper.rb:64:in `task_link'
/Users/evan/Development/Ruby/eclipticdb/app/controllers/tasks_controller.rb:103:in `escalate'
... etc ...


The first fix - but with a problem

A year ago, I fixed this just by adding include ActionView::Helpers::UrlHelper at the top of that controller. This worked great ... for a while.

Lately, I've been rewriting this application into a RESTful style - it had previously been a controller/action style application. In the process, I started linking things with resource paths and polymorphic paths ... a lot of link_to @task and edit_polymorphic_path(@task) sorts of bits. And these started breaking. I began seeing this mysterious error:

Error:

You have a nil object when you didn't expect it!
The error occurred while evaluating nil.url_for

vendor/rails/actionpack/lib/action_view/helpers/url_helper.rb:71:in `send'
vendor/rails/actionpack/lib/action_view/helpers/url_helper.rb:71:in `url_for'

This one was a real bitch to debug, I have to say. The line in question that was failing in url_helper.rb said this: url = @controller.send(:url_for, options). Clearly, @controller was nil ... which was very bizarre, because I never interact with that instance variable anywhere.

I thrashed around trying to find the cause of this error for quite some time. Eventually I realized that the link_to method was only failing when called from a view in TasksController, and not from any other controller. And then I realized that TasksController was the one where, a year ago, I'd put include ActionView::Helpers::UrlHelper at the top. Somehow, including that helper in the controller was nullifying @controller when those helper method we called from within the view. I removed the include and my polymorphic and resource links all started working again.

Now back to the original problem!

Of course, that then left me back with the problem I'd had a year ago ... needing to use link_to from within the controller and having no way to do it. After a fair bit of googling around I found this post from Neeraj, which had an interesting approach -- but a commenter had suggested a much easier solution:

[sourcecode language='ror']self.class.helpers.link_to[/sourcecode]

I'm not certain where one would find this in the docs, but it does seem to have solved my problem for now. Onward and upward!

Busy building and launching UniThrive.org!

UniThrive.org is a fantastic new non-profit startup that seeks to help reduce the cost of higher education by networking college students with alumni, and facilitating direct, zero-interest loans between alumni and students to defray tuition costs.

Technologically, UniThrive is a Rails application that began as a fork of the open-source social networking application Insoshi. Since forking Insoshi, we've nearly doubled the size of the code.

Today, UniThrive is in a live beta test available to students and alumni of Harvard University. Take a look, and check out the UniThrive Blog!

So, I ran into this little challenge: all three STI subclasses use the same controller, "tasks", because they all have essentially the same behavior and differ only in minor details. But, when I do a resources map:

map.resources :tasks

Then I get errors in much of my code when I say things like redirect_to @task, because if that task happens to be an ActionItem, it's trying to call action_item_path(@task), which doesn't exist.

I googled around a bit to no result. Striking out on my own, it turns out the answer is as simple as mapping each resource independently, and just overriding the controller in map.resources:

In config/routes.rb

map.resources :tasks
map.resources :action_items, :controller => 'tasks'
map.resources :work_orders, :controller => 'tasks'
map.resources :problem_reports, :controller => 'tasks'


Now, redirect_to @task works just fine regardless of which subclass @task happens to be.

I'm following Chapweske's approach of blocking mass assignment by default in all models, by putting this line in an initializer:

ActiveRecord::Base.send(:attr_accessible, nil)

This had the expected side effect of breaking several zillion tests, because tests frequently use things like Model.build() and Model.create!() to generate on-demand fixtures during testing. Hartl has a great bit of code that creates unsafe_build() and unsafe_create() methods in ActiveRecord. You can use these methods instead of build() and create() to function as expected in your tests.

This works great, except that I also use the mass-assignment method update_attributes! in my tests and specs frequently, particularly when I want to spec the effect a change on one model has on an associated models' methods. So, I expanded on Hartl's helper code a bit, to give myself the necessary methods. In case it helps anyone else:

/lib/initializers/unsafe_build_and_create.rb

class ActiveRecord::Base

def unsafe_attributes=(attrs)
attrs.each do |k, v|
send("#{k}=", v)
end
end
end

NoMethodError in 'LoansController POST 'create' with valid parameters should succeed'
undefined method `env' for

This happens on the line where I call post :create in a controller spec.   Undefined method 'env'?   What's that about? I'm certainly not trying to call a method named "env".

It took me a little bit to figure out what was going on.  See, this series of tests needed access to a particular LoanRequest object I was pulling out of fixtures.   So I'd put above the tests:

before(:each) do
... some other stuff ...
@request = loan_requests(:johns_loan_request) # fetch fixture
end

Well, kids, it just so turns out that it's a bad idea to overwrite the @request instance variable in any rails context. Who knew?

Come to think of it, it would be nice to change the accessibility and/or mutability of Rails' basic instance variables and classes to prevent this kind of accidental overwrite by the programmer. Because when you make that mistake, it's invariably a bit of a pain to figure out because the error it causes is obscure.

Maybe I'll have to dig into the code one of these days to see if anything can be done about it.

While I understand the guy's complaints, I have to say I disagree. I think he's complaining about a code convention that he shouldn't be following in the first place, rather than the underlying language, and he's ignoring some of the other most useful things Sass brings to the table.

Nesting in Sass

One of the most common problems I've faced over the last eight years of writing stylesheets is interfering selectors. When you have a complex cascading selector, it's often not obvious exactly where it will apply, because of the way priorities work. So a hundred times I've set some styling on UL's and LI's (thinking of ones in my #content block), only to have them accidentally interfere with the layout of my suckerfish dropdowns back in #nav.

That's an easy case, but sometimes with complex selectors it can be hard to figure out who's interfering with whom. However, once you've found the culprit, the solution is generally to go back and wrap all of the rules in an outer selector, changing all of my li {rule} selectors to #nav li selectors, or whatever. When you have twenty different rules in that section, doing this is a royal pain in the butt. Especially when you have multiple tag selectors on one line: it's seriously annoying to change the nice clean h1, h2, h3 to #content h1, #content h2, #content h3!

When you do need these wraps, Sass makes it super easy via auto-nesting:

#nav
li
:color #whatever
:float left

a
:whatever etc


will compile to:
#nav li {
color: #whatever
float: left
}
#nav a {
whatever: etc
}


Now, the author of fecklessmind is complaining about how this makes rules harder to find, and how it slows down parsing. Both of these can be true, if you overdo it. But I don't - Sass doesn't force you to wrap your rules this way, and I frequently don't when it doesn't provide any benefit or when it would cause me to write redundant rules. I can and frequently do write single-line cascading selectors, and rules without wraps at all - the very things fecklessmind is complaining that Sass takes away from him.

Nothing about Sass prevents me from writing things like this:
body #nav ul li a
:float left


or even
#content h1, #nav h2, .article h3, p h4
:font-weight bold


If that's really what I want to do. I learned how and when to wrap selectors with a near-decade of writing CSS, and I apply those same guidelines when I write Sass - Sass just makes it easier when I do want to do it.

The benefits of nesting early

While I don't use Sass nesting everywhere I possibly could, I do often use it slightly more than would be absolutely required.

The reason is that it heads off a lot of annoying bugs with interfering selectors. For example, say a rule I wrote for .article .body p, and it's not getting applied. After some sleuthwork (long, painful, frustrating sleuthwork if I'm on a platform without firebug, like IE), it turns out this is because there's a #content p rule 2000 lines earlier in the CSS file that's obscuring it. When I nest things in Sass to create a clean cascade hierarchy, this kind of interference is far less likely to occur in the first place.

Meanwhile, the other benefits of Sass

CSS is riddled with problems, and Sass solves two of the most egregious: magic numbers/constants, and compiled server-side imports.

Eliminating magic numbers in CSS

For constants, Sass lets me define commonly used tokens (like colors, for example), and reuse them throughout my stylesheets. This means if I want to adjust a color, I can change it in only one place and the result is reflected throughout my code. Very handy:

!main_link_color = #48950a

#content h1
:border-width 0 0 1px 0
:border-color= !main_link_color


Now, if the client says "make the links blue, not green", I can change that constant and it gets automatically reflected everywhere else. Brilliant.

Organizing my code

fecklessmind says this:

... imagine that the stylesheet is 5000-lines long and you’re looking for p selector, rather than #article. In classic CSS you could just search for #main p, but in Sass they are miles apart. Swell, isn’t it?

A five-thousand-line line file? You're doing it wrong. No code should ever look like that. CSS is the only major language that compels you to work that way and Sass fixes it.

Every good programming language lets me put my code across multiple files, in a nice, organized heirarchy. One class per file and all that: essential for readability and maintainability. But if I use CSS, I can't very well organize my stylesheets into multiple files. If I do, I have to import them client-side, which generates extra hits for the user's browser and extra load for my server. As a result, CSS files tend to be monolithic multiple-kiloline monstrosities.

Sass fixes this. If I use @import to import a Sass file into another Sass file, Sass automatically and transparently compiles that server-side and ships out a single file to the user.

The result is that, writing Sass, I often have 20-30 files containing only a page or so of code, each for a specific feature or layout section. The client still only sees screen.css (and maybe print.css, mobile.css, and ie6.css), but screen.css contains the compiled contents of layout.sass, nav.sass, links.sass, content.sass, footer.sass, etc. In case I need to scan through the compiled screen.css and figure out where a rule came from, I start each file with a single comment containing the name of the file; /*------ nav.sass and such.

If the rule is for paragraphs that could appear anywhere in #main, in my Sass code it would be a file called main.sass, which is usually a relatively short file; 50-60 lines. (All the things destined for other elements would appear in articles.sass, or calendar.sass, or data_tables.sass, keeping main.sass short for only the universal elements).

That logical grouping — the way every other programming language does it — helps me find my CSS rules much more quickly, I think, than fecklessmind's "search for one-line selectors" would. Because with his approach, I might think I'm searching for #main p, when in fact what I really want is #main .section p, and thus my search won't find it.

In reality, there's no way to make a single 5000-line file easily maintainable, period. fecklessmind's little tricks are just that: tricks built from years of experience working in a broken system. Better to use logical organization to solve the problem, and Sass lets me do this.

The bottom line is, badly-written Sass could be horrible to maintain, and maybe worse in some ways than badly-written CSS (but better in others, particularly in weird cross-reactions between unwrapped css selectors). But the same is true of badly-written code of any type. And in my experience Sass gives me much better tools to write maintainable stylesheets than CSS alone does.