From: Garthy D <garthy_lmkltybr@...>
Date: 2011-12-22T21:03:08+09:00
Subject: Re: Accessing class instance variables from an instance?


Hi Robert,

On 22/12/11 21:11, Robert Klemme wrote:
> On Thu, Dec 22, 2011 at 2:09 AM, Garthy D
> <garthy_lmkltybr@entropicsoftware.com>  wrote:
>> I've run into this kind of thing before, and it does impact me regularly.
>> Having said that, I know my C++ background (which allows this) biases me
>> somewhat on the issue. ;)
>>
>> I tend to steer away from "Foo.bar" style calls inside class Foo for the
>> following reason- it complicates refactoring. If you determine that a class
>> would benefit from being split (perhaps because another class has common
>> functionality and you want it to share a common base), then you have to
>> selectively rename all of the calls you have made to point to the correct
>> class. Sure, it's not hard, and doesn't take too long, but when you're
>> refactoring something for the sake of something else you have developed, not
>> having to deal with additional chores to make it work makes it less likely
>> that you'll stuff something up in the process. A similar problem arises with
>> renaming the class, but admittedly the solution tends to be far easier.
>>
>> For that reason, I'd instead use "class.bar" calls, if it were possible,
>> which it is not, which brings us to "self.class.bar". I do this as you can't
>> always tell ahead of time when you'll need to refactor something to pull out
>> a base class, and so I'll assume that (almost) all of the classes I work on
>> could be subject to such changes later in its lifetime. If you could tell
>> beforehand, you would have designed it that way in the first place. :) Now,
>> if you have a class method that is used both inside and out, but mostly
>> inside (and regularly), it's a pain to have to write "self.class.bar" each
>> time you want to call "bar". Incidentally, for "bar", I have a specific
>> method in mind for something I have been working on (ie. a real-world
>> problem), that makes the most sense as a class method, since it works
>> entirely off the arguments and carries no state. The call in question has
>> very specific structured input requirements, and is called extremely
>> frequently inside the class, and occasionally outside.
>>
>> In my case, I ended up solving it by adding an "abar" method that just
>> called "self.class.bar", although using it that way does feel a bit silly,
>> and makes the code less readable (an observer might ask what the difference
>> between "bar" and "abar" is, for example).
>
> A better alternative would be to do this:
>
> class Object
>    def my_class; self.class end
> end
>
> Then you do not need to do
>
> Foo.bar
> self.class.bar
>
> but instead you can do
>
> my_class.bar
>
> which seems a lot clearer to me while avoiding the nasty "self.class".
>   Maybe someone else can think of a better (presumably: shorter) name
> than #my_class.

I like that. I hadn't thought of it myself- it's one of those "of 
course, it's so obvious!" type things, which leaves you wondering why 
you didn't think of it yourself. ;)

I have to admit a guilty pleasure of mine is adding the occasional 
method to "Object" in Ruby, and I could see myself adding such a thing 
to "Object". But I shall have to keep such impure thoughts away from the 
list. ;)

>> Anyway, I just thought I'd share my particular experience. It's not hard to
>> work around, but there are definitely good reasons to want to make it work
>> that way. It falls under the area of a "quirk" for me, but that is of course
>> just my personal experience based on my personal experience.
>
> I guess my main point is that making a method call appear to be on the
> local instance which in fact is done on another instance is
> misleading.  For the reader it is better to at least signal that #foo
> is not invoked on self but rather on another instance.  For that you
> need a method or a variable reference.

There are definitely cases where the additional information (ie. that it 
is not being performed on the local instance) is vitally important. I am 
just saying that there are also cases where it is not. It is important 
to emphasise that I'm not saying only one case exists or one is more 
important than the other. Anyway, the question becomes: Whose 
responsibility is it to ensure that in these cases that it is properly 
emphasised? Is it the responsibility of the language to ensure these 
cases are always handled correctly by enforcing it through the language, 
or should the language step back, and leave it to the user to write 
either good or bad code?

I suspect that when designing a language you would come up across a lot 
of these sorts of choices. C++ requires impressive template gymnastics 
to pass a single generic parameter- type safety through fear of daunting 
pain. ;) Python makes whitespace syntactically significant; pretty code 
whether you like it or not. ;) C++ uses operator overloading for ease of 
expression, Java disallows it to avoid difficult to follow code. A good 
case could be made to either allow, or forbid, the exact same feature in 
this case. I think the problem we are talking about is that kind of thing.

I guess that personally I err on the side of giving the user the means 
to shoot themselves in the foot, if they really want to, provided that 
you don't set things up so that they can accidentally do it. Perhaps in 
this case it would be something like this:

class Foo

def self.bar *args
   whatever *args
end

attr_class_method_from_instance :bar

def baz
   one
   bar 2,4
   two
   bar 6 if three
   four
   bar 7
end

Working on the assumption for now that allowing class methods to be 
called implicitly from an instance is bad (it's probably easiest just to 
assume this for now), then we only allow a user to go for the foot shot 
(calling a class method for an instance) if they've explicitly requested it.

I would hate to be forced to, for example, write:

def baz
   one
   self.class.bar 2,4
   two
   self.class.bar 6 if three
   four
   self.class.bar 7
end

Because it draws attention toward the instance versus class method 
distinction, which simply may not be the most important thing in this 
context.

(Of course, the "self.class.bar" call could be wrapped in an "abar" 
wrapper, which might help to make it more understandable, or use the 
my_class shortcut you suggested above)

>>> That has the added advantage that
>>> you make the call explicit in instance methods which use it and it is
>>> immediately clear that this method does not manipulate instance state.
>>
>> There are definitely cases where this could be advantageous, but there are
>> also cases where this is not necessary or desirable. In general, I would
>> personally say that if a particular mechanism is sometimes beneficial, but
>> sometimes not, it is better to leave it up to the person writing the code to
>> emphasise the mechanism where it is appropriate, and de-emphasise where it
>> is not. For example, sometimes it is beneficial to be specific about the
>> particular class of an object (or its capabilities), but would that justify
>> requiring every method call to be prefixed by the class name, every time? I
>> would say not. There are times where being specific about it would be
>> beneficial, but in the remaining cases it would clutter up the code
>> unnecessarily.
>
> We differ in the estimation about the distribution of useful and
> harmful.  I'd say obfuscating a method is invoked on another instance
> does more harm than good so the default should be to not make it too
> easy.  This forces people to know what they are doing when they do it.

I think this touches on the safety aspect I mentioned above. I think we 
would agree that powerful but dangerous techniques should require 
deliberate actions to trigger, to avoid them being used inadvertently. 
We might differ on some of the details but I think we have similar 
motivations here.

>> I think in the situation I outlined above, and possibly also in Josh's case,
>> it falls in the realm of clutter. This does not, of course, take away from
>> the fact that in some situations it would be important, or absolutely
>> essential, to explicitly convey that the instance state will not be
>> manipulated. I know that I have certainly run into this case as well- I
>> won't bother with a personal anecdote as I suspect you also already know of
>> many cases where it is extremely beneficial already. :)
>
> That's the point: OO is mostly about the state of instances and its
> changes.  Everything that helps keep clear what's happening is good
> IMHO.

Personally, I'd adjust that to say that everything that causes the 
behaviour to become clearer is good to include, but I would exclude 
unnecessary detail that does not add to the understanding, and 
*especially* leave out unnecessary verbiage that draws attention away 
from what is most important in that context. I would say that sometimes 
being explicit is beneficial, but sometimes it is not, and sometimes it 
has the opposite effect. Again, all IMHO. :)

>> I guess what I am saying is that if something is sometimes beneficial, and
>> sometimes detrimental, that enforcing it is not always the best idea, but
>> providing an optional means to make it explicit is beneficial. I do not
>> believe that is a good idea in general to either force a mechanism that is
>> only sometimes useful, or argue that as the mechanism is sometimes
>> important, that having a means to conveniently avoid the mechanism when it
>> is detrimental is not justified. There are of course always all sorts of
>> tradeoffs when designing a language- perhaps the particular feature Josh and
>> I would find beneficial is not readily possible for some technical reason.
>
> There are no technical obstacles here defining instance methods which
> access class state.

Ah, sorry, I did not explain properly. I mean, if there are any 
technical reasons for which a class method of a certain name (say, 
"bar") could not be easily triggered when an instance method of the same 
name (eg. "bar") is invoked. Perhaps, for example, it isn't easy to 
ensure without complicating calls or adding some sort of instance method 
proxy, which might not be desirable? I'm basically just saying that 
there might be a tech blocker here that I'm unaware of which could 
actually be the main reason behind the behaviour.

>  I just happen to believe that the tradeoff fall
> on the side that we do not want to have such a feature because the
> convenience of the writer is the pain of the reader of the code.

I would say that if you get the wrong person writing the code, no 
framework of restrictions can prevent them from writing terrible code. 
Such people always find a way, somehow. ;) The right developer instead 
could emphasise such a thing, if it is important in the context. And 
anyway, I would argue, why hold back the better developers for the sake 
of the ones that cannot write easy-to-understand code?

Anyway, these are just a few more rambling evening thoughts from someone 
who should probably be getting a bit of sleep soon. ;) I hope they are 
useful, even if it's nothing more than just a different perspective on 
the problem.

Garth