W3C home > Mailing lists > Public > public-webapps@w3.org > July to September 2012

[selectors-api] Kudos on find/findAll, feedback on spec readability

From: David Greenspan <david@meteor.com>
Date: Sun, 16 Sep 2012 14:40:28 -0700
Message-ID: <CABJCAJ_B9nj0jvnvC_Pt4uMyb3kz=RVzCePm_0Z6x1feauSvwg@mail.gmail.com>
To: public-webapps@w3.org
I've been reading the CSS/selector specs lately.  I'm interested both as a
framework implementer, designing and implementing jQuery-like functionality
for Meteor, and as an app developer who looks forward to the days when
browsers provide the APIs we want.

Coming into the new Selectors API, I was really hoping to discover that the
proposed find and findAll methods behaved essentially like the jQuery
calls.  Put simply, in running elem.find("div p"), both the "div" and the
"p" in question would have to be descendants of elem (as well as each
other) in order to match.  In running elem.find("> div p"), there would be
the additional constraint that only immediate descendants of elem would be
considered as candidates for the "div".  Unfortunately, this didn't seem to
be the case... until I read the spec about a half dozen more times, and I
realized it was!  If I'm reading it right, the spec does work this way and
is additionally much more flexible about the relationship between the
element, the scoping element (or elements), and the selector.  From reading
the list archives, I get the feeling that few web developers who come
across the spec understand this much.

First of all, I want to offer some encouragement.  The names "find" and
"findAll" are great, and their functionality is exactly what web developers
will expect and appreciate.  It's familiar and practical.  And as long as
in basic cases like element.find("a b c + d"), it's required that all of a,
b, c, and d descend from the element, I am happy with whatever scope
selectors and edge cases are necessary or desirable.

However, the spec is very hard to understand.  I've read several specs
cover-to-cover (Java VM, ECMAScript, JPEG) and I think there is a lot of
low-hanging fruit in this short document (Selectors API Level 2) -- ways to
improve its clarity and enlist your readers more in support of this spec.

Some feedback from a reader, for what it's worth.

6.2 -  "... first matching Element node within the subtrees..."

Don't define a new term "subtrees" which just means "descendants."  Say,
"... first matching Element node that is a descendant of the context

The main take-away from section 6.2 should be that find/findAll use a
*different algorithm* from querySelector/querySelectorAll, or the same
algorithm with a different setting (relative=true).  This point is almost
entirely lost.  I think the intent is for the reader to perform a sort of
"dispatch on type" when the algorithm in 6.5 is invoked, but it's confusing
because you can't really look at a selector and a relative selector and
tell them apart, and certainly not before parsing them.  Isn't a relative
selector (string) just a selector (string) optionally preceded by a
combinator?  It seems to me that find() and querySelector() *do different
things*, conceptually; for example, the former will add an implicit :scope
to your selector and the latter won't.  It's nothing to do with their
arguments.  Given the same arguments, they will do different things.

To make it concrete by analogy, how would you interpret this spec (or short
of that, how does it feel to read and try to understand)?

Definitions: "Chairs" are made of wood.  "Generalized Chairs" are made of
wood, metal or plastic.

The VARNISH function accepts a chair.
The PAINT function accepts a generalized chair.

VARNISH must coat its input with transparent, Acme brand varnish.
PAINT must coat its input with liquid.

When either method is invoked, the implementation must perform Algorithm

Algorithm Coat:
1) If the input is a wooden chair, coat it with varnish.
2) Otherwise, if the input is a generalized chair, coat it with paint whose
color matches the chair's referential color context.

Here are the stumbling blocks:

- If you try to simulate the algorithm, it's a head-scratcher.  Aren't
wooden chairs also generalized chairs?  How do we get to step 2 to paint a
wooden chair?  Are we really supposed to varnish or paint based on whether
the chair is known-to-be-wood or merely incidentally-wood, or is painting
wooden chairs not possible?

- Presumably, painting was invented to support new chair materials and
achieve color-coordinated finishes, yet there is barely a functional
description of what PAINT does, let alone the intent.  "PAINT must coat its
input with liquid" is the closest thing we have to a definition.

- We're describing these two functions along three axes, but the
information is so spread out that it is hard to integrate.

The ideal in clarity, which admittedly is not always possible with the need
to specify algorithms and share them between sections, is something like:
The VARNISH function accepts a chair and coats it with transparent, Acme
brand varnish.

The PAINT function accepts a generalized chair and coats it with paint
whose color matches the chair's referential color context.
That's a breath of fresh air.

However, there are many other options, like:
The VARNISH function accepts a chair as its argument.
The PAINT function accepts a generalized chair as its argument.

Let the boolean variable "colorful" be true if we are PAINTing and false if
we are VARNISHing.

When either method is invoked, perform Algorithm Coat with two inputs, the
chair or generalized chair to coat and the variable "colorful".

Algorithm Coat:
1) If the "colorful" boolean variable from the caller is false, coat the
input chair with transparent, Acme brand varnish.
2) If "colorful" is true, coat the input generalized chair with paint whose
color matches the chair's referential color context.

Ok, enough about chairs.  The point is that a simple refactoring of the
explanation can beget huge readability gains.  Also, I'm still not 100%
sure if find("div p") will parse as a non-relative selector and run with no
implicit :scope, or if it is always a relative selector by virtue of being
find's argument.  I'm 90% sure it's the latter.

Other text that made me go "Wah??":

6.2 - "When either method is invoked..." - there are four methods, not 2

6.5 - "... begins with a combinator and that combinator is not ' ' (space)"
- how can a relative selector, in a comma-separated list with whitespace,
physically start with a space, i.e. a descendant combinator??  Supposing it
can (at the beginning of the string, perhaps), why would you special-case
it and prepend what's presumably a second descendant combinator in step 5?
 It's kind of hard to tell what the intent is here except to prepend :scope
if :scope doesn't already appear in the selector.  I can guess that's all
there is to it, but I'm not sure.

7 - "contextual reference element" - This term is revealed to mean simply a
potential :scope element.  I would call it something like a "scope
candidate" -- or basically anything besides what it's called.

I swear, the definition of a + b in this writing style would at some point
say: "Conforming implementations MUST perform the context operation on the
context number and the contextual reference number in the current context."
 Or else some of these terms would be given unusual names and defined in
external documents.

7 -  "The :scope pseudo-class MUST match any element that is in the
contextual reference element set."  Does it also have to not match elements
that aren't in the set, or are implementations allowed to match :scope with
anything?  It's strange to see RFC MUST paired with loose language.

7 - "Specifications intending for this pseudo-class to match specific
elements other than the document's root element must define a contextual
reference element set." - Whoa, there are deep, philosophical questions
here. The previous paragraphs already defined :scope, CRES, and their
relationship.  Is a spec stronger when it mandates that other specs conform
to it?  Is it enough to define an apple as red, or do I also have to say
that other specs that define apples must define them as red too, perhaps
contingent on their "intent"?  I'm going to take the easy way out here and
interpret this sentence as a helpful hint to other standards writers that
if you want to override the default specification of :scope in some context
defined in another standard, you have to do it by overriding the default
specification of what goes in the CRES in that context.  (But there is that
small-caps MUST again, inviting us to imagine what sort of non-conforming
W3C spec we might find in the wild violating this requirement.)

8 - Feature Strings - Change last paragraph to: "Conforming implementations
must return true ... if they are perfectly compliant, and false if they are

9 - Examples of find() and :scope MUST be added here.  Seriously.  This
section is the last chance to push that lucky reader over the final
threshold to feeling like they fully understand the spec.

In closing, after reading the spec and perusing the mailing list, I don't
really see any significant unresolved technical points, only lots of
confusion about what the spec says.  I can't find a single blog post,
article, or other basic explanation of this API on the entire Web.  I think
if I got 20 top web developers from the Valley together and managed to
explain the spec to them on a chalkboard as I understand it, they'd all
approve it in about 5 minutes and go home.  If I instead told them to sit
down individually and try to decode the spec, 19 of them would probably
give up.  That's terrible PR if nothing else.

Since computers unfortunately can't read specs, all standards have to be
written down in clear English at some point for implementers and software
authors to understand, whether the standard in question is simple and
airtight like JSON or long and exhaustively enumerated like chip
instruction sets.  That's what specs are for.  Yet even after dissecting
this document in such detail, I don't feel like it's intended to be read
and understood by me or any human I know.

Thanks for reading,
Received on Monday, 17 September 2012 10:16:33 UTC

This archive was generated by hypermail 2.4.0 : Friday, 17 January 2020 18:13:38 UTC