Well, I didn't get to make this post last night. :-/ Things such as
work, food, and sleep really interfere with coding and posting.
I'll make a few comments now and flesh them out later.
As I said, I'm an Object Guy and Jamey is a Database Guy. What does
that mean?
I was working on an app and using KirbyBase. Because of the work Jamey
had done recently, creating objects from the database records was much
easier.
But I noticed something. My fields were all simple types. (After all,
that's how databases work.) But I was taking my smaller objects and
using them to build bigger objects. In most cases, I was storing a key
into another table. So I would look up a record, and four of its fields
would actually be keys into other tables. I'd then look up those four.
I'd then embed my four little objects inside an object corresponding to
the original row.
I had a feeling that I was doing too much work, that some of this could
be done for me.
My thoughts were like this: OK, we have primitive types such as integer
and string. From OOP, I have a habit of thinking of classes as abstract
types (see Bertrand Meyer -- I think).
So if I want an object to have a member which is (e.g.) a Person, why
can't I treat Person just like String and Integer?
In Ruby, we don't think of variables or attributes as being typed. I
make an exception in the case of retrieving objects from databases,
because I am very used to the idea of fields having types. (Although
someday I will comment on that also.)
After all, even in Ruby, we often think to ourselves, "This member
will be a string" -- although if we are not careful, we might assign
any other thing to it. But I won't go off on too much of a tangent
here.
So here I am. I'm thinking: Treat simple types (simple fields) the way
we already do. We store a type for each field. Let's store the complex
types the same way... but in the actual table, just store a key into
another table.
That requires the table name and key field name to be specified
somehow,
of course. My take: Let's get slightly railsy and assume the class name
lowercased is the table name (e.g. Person => person.tbl). That means we
only have to specify the field name explicitly (and I've been thinking
of a way to eliminate that also, by optionally establishing a known
"key field" for a table).
I was thinking "composite objects." I was thinking "embedding simpler
objects inside more complex ones." I was thinking "this is how I can
easily and transparently marshal my objects."
When I mentioned this to Jamey, he said in effect, "Oh, you want one-
to-one links between database tables." That's Mr. Database talking to
Mr. Object. He was quite right, of course, but it confused me to think
that way.
It was as if my mechanic said, "Oh, you're trying to use this
solenoid/starter assembly to initiate sequential explosions inside your
cylinders with the aid of these spark plugs." To which I would say,
"Uh, I just want to start my car." (As it happens, I really don't know
much about cars or engines. That's a hardware problem.)
So Mr. Database says, "Let's implement one-to-many also." (And he may
have done others also, I don't have a KB install in front of me.)
And my reaction is, well, OK, fine. But first of all, I don't
personally
see a need for it. (Of course, I might discover a need next week.)
Second of all, I am not sure what analogy that would have in object
terms. I suppose it would in effect be embedding an array where all the
objects are objects derived from the rows of the child class. This
bothers me a *little* because all the elements of the array would be
of the same type. If you ever assign something else to that array,
you'll
get an error when you try to insert.
Well, there's already some clash between Ruby and the database concept,
because db fields have types, and attributes (as variables) don't have
types. But this seemed like going a step farther to me -- heightening
the clash by creating an array that has to be homogeneous, like a
Pascal array.
Third of all, if you start to have "one-to-one" and "one-to-many" links
and such, you start having to distinguish between them. This makes
table
instantiation, even in the default case, just a tiny bit more
difficult.
It's creeping complexity. It's very small, but it adds up (and
sometimes
multiplies if you're not careful, the way probabilities multiply).
See, there is a sort of "conservation of complexity" in any system. If
I knew more information theory, I could express it better.
Take data compression as an example. I can write a simple, dumb
compression program that will compress most text by a factor of 2 or
better. But that's not very good. An extremely complex algorithm (a
much
better one, that is) might compress by a factor of 10 or more.
But the net information is the same, don't you see? We have moved the
information from the data into the program. We have a very high-entropy
resultant data file and a very low-entropy program. I have increased
the
data's entropy at the cost of increasing the entropy of the software.
(My freezer keeps things very cold, but there is hot air coming out the
back of the fridge.)
I can compress the Declaration of Independence into a single byte with
a
sophisticated enough program, like the one shown here in part:
if text == 'a'
puts "When in the course of human events, it becomes necessary"
# lines omitted
puts "our lives, our fortunes, and our sacred honor."
end
You think I'm kidding, and I am. But I'm also not kidding at all.
If I write a C program in 600 lines, I can probably write it in Ruby
in 100 lines. Where did the complexity go? It left my program and went
into the interpreter. That is where it belongs -- under the hood.
Information hiding is how humans manage complexity. The concept of the
black box is a greater human invention than the discovery of fire or
the wheel.
Now, suppose I specify a table with KirbyBase that has two simple
fields,
an integer and a string. We do it something like this (I probably am
forgetting the method name):
create_table(:mytablename,
:alpha, :Integer,
:beta, :String)
Fine and dandy, nice and simple.
Now suppose we add another field -- I'll abandon the Greek alphabet and
call it "boss" which will be a Person object.
In an absolutely perfect world, this would be "just another type." The
software would read my mind and do exactly the right thing, and all I
would type is:
create_table(:mytablename,
:alpha, :Integer,
:beta, :String,
:boss, :Person)
But this leaves some unanswered questions. Here are the questions and
my
answers -- pardon me for personifying the database software:
1. "What table will I get this 'person' stuff from?" (Just derive the
table name from the class name.)
2. "What field in the child table will I use as a key?" (Hmm, maybe
I'll
have to tell you this one. It might be cool to be able to
designate a
key field, though.)
3. "What type is that key field?" (Given the name, you can find it in
the
child table's information.)
4. "What if Person itself is another complex object?" (Relax, just
apply
the same algorithm recursively. Worst that can happen is there is
cyclic data, you'll go into an infinite loop, and I'll have to
kill you.)
Now, Jamey's first attempt at this had me writing code in the
MYtablename#kb_create method, calling a method named one_to_one_link or
some
such. It felt very manual to me, like I was hotwiring my car.
"What's the big deal?" says the mechanic. "All you do is, you grab
these two
wires, not these two, and touch them together for a little while, not
too long,
and..." And I reply: "I don't want to reach for the wires and grab
them, I
don't want to memorize the colors, I don't want to estimate the time
interval,
and I don't want to see wires dangling. Keep that under the hood."
In one iteration, perhaps not the present one, a one-to-many
relationship was
stored inside its parent object as a KBResultSet. Ugly to me. If we
must have
one-to-many, let it be just an array. I don't want my choice of
database to
intrude into my objects any more than necessary. Given the right glue,
*any*
database should be usable for my objects. KirbyBase is implementing
this glue
pretty well so far, INSIDE the db software where it belongs.
In the one-to-one case (present iteration), the type information is
specified
as an array. This was my suggestion, and in general I like it. But it
is too
complex. We may not reach the ideal shown above, but let's strive
toward it.
When I have Jamey's email in front of me, I'll tell you exactly what I
mean.
I was also disturbed a little to see that the class had to inherit from
a KB
class (I forget which one). It's hard to articulate why this bothers
me. It's
one more thing to remember, one more thing to do, and it's not a
totally clean
separation.
Also, a word or two about the kb_create method (called automatically
when a row
is retrieved, to turn it into an object). In a perfect world, we should
"normally" not have to define one at all. I'm thinking of ways to make
it
usually unnecessary. (In the case of "calculated fields," this might be
the
very best place to put them, however.)
Hope this helps clarify things a little.
And I'm glad to see this discussion happening in public, just in case I
give
Jamey some really stupid advice. 
I'd hate to persuade him into a bad
design.
Cheers,
Hal
I'd hate to persuade him into a bad
I'd hate to persuade him into a bad