Boost logo

Boost Users :

From: Paul Mensonides (pmenso57_at_[hidden])
Date: 2006-02-05 09:16:23


> -----Original Message-----
> From: boost-users-bounces_at_[hidden]
> [mailto:boost-users-bounces_at_[hidden]] On Behalf Of Jim Douglas

> I may have used the words concrete->abstract axis. It might
> be a basis
> on which to rate the libraries with a "fuzzy" classification that
> indicates position on the axis and spread of coverage. E.g. Boost.Crc
> even contains some pre-defined ready-to-use concrete types,
> whereas at
> the other extreme Boost.MPL is an abstract framework.

And the pp-lib is a framework that is more general still.

> From a purely practical POV here is the communications problem:
>
> I have a piece of software to construct and I have sufficient
> confidence
> that I can write it in C/C++ in N weeks from the ground up.
>
> How can I extract enough information from the Boost docs to identify
> those Boost libs that are applicable to my problem domain?

For a library approaching full generality, it is applicable to just about
*every* domain, but that doesn't mean that it should always be used. Whether or
not it should be has to be determined on a case by case basis in comparison with
other possibilities.

> How can I then evaluate those libraries so identified, enough to be
> convinced that the full *learning* time plus deployment time
> will result
> in the project lasting N weeks or less?

You can't because it doesn't work that way. For libraries like these, you learn
them to increase your general toolset which can later be applied to any number
of projects. Note that this is a lot like learning a language and the
accompanying techniques outside of the context of a real-world project with
real-world deadlines.

> > Examples (as opposed to use cases) in the
> > documentation exist mainly as an attempt to induce the
> necessary lateral
> > thought, not to show what problems the library solves.
>
> Examples - yes, I can't get enough of them. A couple of small caveats:
>
> a) There is a danger that you might not relate to an example if you
> can't see that by changing a few things you can put it to good use in
> your particular context. OK so you can't please all of the
> people all of
> the time.

The danger is not that you might not relate to an example. The danger is that
you might not "get it". Meaning not that you don't understand what the example
does, but that you might not achieve the perspective (or the beginnings of the
perspective) that the example hints at.

Does this mean that less people use such a library? Yes, it does, and, in its
way, that is a good thing. Because such a library can be used just about
anywhere, it takes both a reasonable amount of understanding and a reasonably
correct perspective to apply wisely. It shouldn't be thrown at any possible
problem where it could be used. Instead, a smaller number of people (that know
what they're doing) use libraries like this to implement other things (often
still fairly general) that can be used effectively by more people.

As a practical example of this effect, the pp-lib can be used to generate code
for a template metaprogramming framework (e.g. the MPL). The MPL can, in turn,
be used to implement a binary literal utility (see Dave and Aleksey's book).
The binary literal utility can be used to implement some sort of message
cracking function. Each step down this path represents a decrease in
generality. With each decrease, the number of people that are likely to use it
correctly increases (sometimes dramatically). This occurs because the higher
the generality, the higher the number of possible malconceived uses (as well as
the higher the number of well-conceived uses).

The person that uses the binary literal utility (which is fairly easy to use) is
also ultimately using the pp-lib, though indirectly. What I'm getting at here
is that although the number of direct users of something as general as the
pp-lib or the MPL may be fewer, the number of indirect users may be much
greater. All of this isn't about documentation per se, but rather that the goal
of a library is not always about maximum number of direct users.

> b) Examples need to be grounded in the real world - which tenuously
> links to use cases. 'foo/bar' examples should be banned :-)

Presumably you mean "contrived" examples. I think that depends heavily on what
a piece of code is supposed to be an example of. If you make examples too
complex, the point that it is trying to make is lost in the particular details
of the example. Then the example itself becomes a cookbook interface when it is
really supposed to be an example of something else. Using an example to bring a
very general interface or group of interfaces all the way down the generality
axis to a particular use case virtually requires that the example be complex.
Even for something as simple as a binary literals example requires dealing with
domain-specific issues (like octal literals) that have nothing to do with the
library being used to implement it.

The problem that I have with your viewpoint is that you seem to be expecting
even very general libraries to show where they're useful in concrete, specific
examples. By definition, a finite set of use case examples doesn't even come
close to being sufficent. A different tack is needed. Instead, examples for a
highly general library should focus on the perspective and creativity required
to apply the highly general library to an unbounded number of specific
situations. So, the purpose of such examples aren't about use cases, they're
about a perspective-shift that, if attained, allows users to _identify_ use
cases. If that is what it is about, whether or not an example is practical or
contrived is a non-issue.

I don't, OTOH, have a problem with your viewpoint when it is applied to less
general libraries where the number of use cases is relatively small or falls
into a relatively confined scope. I just don't think that the whole concept of
"how to do documentation" can fit into a unilateral mold for all types of
libraries. I don't think that the documentation for highly general libraries
like the MPL or the pp-lib, for example, should be "dumbed down" (i.e. made
concrete enough) to the point of being understandable by all users. I actually
think that that is dangerous. Such tools can create gigantic messes if not used
wisely (e.g. if they get used mainly for the sake of being clever). For those
that are actually interested in expanding their perspective (as opposed to just
getting a job done), then the learning of these kinds of tools should fall
outside of a "project that is supposed to take N weeks or less" because such
learning implies not just learning what each interface does, but how to identify
use cases when you come across them. Documentation that induces that ability in
users is a lot harder to write than one might think based on a casual first
glance.

I think it's worth pointing out that for such libraries, the reference
documentation alone (that includes only contrived examples) is enough for
someone that is actually motivated to learn these things. I'm not the original
author of the pp-lib, for example, yet I understand it completely and know how
to identify use cases when I come across them--simply because I invested the
time and effort to shift the way that I view things (and I'm not the only one).

If more C or C++ programmers were familiar with functional programming, with
metalevel programming (such as the macro systems in Lisp or Scheme), and with
lazy evaluation (such as in Haskell) rather than just imperative runtime-level
strict-evaluation programming, this stuff would be a lot easier for users to
understand and apply. It really is a pity that the imperative paradigm is, by
definition, a superset of the functional paradigm, yet so much of the functional
paradigm is unused or forgotten. Don't misunderstand me. I am not one who
thinks that the functional paradigm is superior. In fact, I think that the
functional paradigm does not adequately represent reality and, as often as not,
that a functional solution requires a great deal of logical contortion. [Some
models in software are naturally imperative while others are naturally
functional. Any time that you use an imperative design to solve a functional
model or vice versa, it causes logical contortion that complicates the design.
Sometimes that complication is worth it for other reasons (e.g. efficiency or
provability)--there are always tradeoffs.] Nevertheless, there is a lot of good
stuff that is common in functional designs that shouldn't be ignored but largely
has been.

Regards,
Paul Mensonides


Boost-users list run by williamkempf at hotmail.com, kalb at libertysoft.com, bjorn.karlsson at readsoft.com, gregod at cs.rpi.edu, wekempf at cox.net