Re: A real world example

Brian Selzer wrote:
[big snips for clarity]
> I agree that it is a violation of the information principle. I also agree
> that if there's no other way to distinguish things, that is, if there is no
> relevant natural key, such as for distinguishing individual quarters in a
> new roll, then they should definitely be exposed to end-users.

fair enough. We are on the same page there then.

> I don't agree that the lack of real world identification is the only reason to use
> surrogates;

It is the only logical reason though. Anything else is a kludge.

> I contend that the fact that that real world identification is
> not *guaranteed* to remain constant precipitates that need. If there IS a
> natural key, then as a best practice, the surrogate should not appear on any
> data entry screen or report. In that case, the surrogate is only needed to
> work around the limitations of the Relational Model, and since the users
> have another means of identify something, they don't have a need to know.

This still makes no sense to me. In the real world if you have a value that is insufficient to identify me over time, then it will be no good for the database either.

For instance if I change my name, then _in the real world_ this is of no use as identifying me when I was younger. It is an identifier for me now, but not at all for me as an entity that stretches over my lifetime. Same for my height, hair colour, age - they all change and hence do NOT identify me as a temporal concept. And thats in the real world, nothing to do with databases. Do you agree that?

Now, there are things that will identify me over my lifetime (in this case the universe of discourse is objects over a period of time), that will remain consistent. THESE are the correct identifying keys to choose if the lifetime of a changing entity is what one is concerned with. I do not see how one can argue against this.

Now sure, a surrogate can represent these perhaps unrecordable identifiers. But if I want to refer to myself at any point in time, I _need to know_ that surrogate (or something very intelligent in the middle has to deduce it from me giving it lots of other information). Either way the relational model must not hide it.

Hence I contend to you Brian that this is nothing to do with the relational model. Rather it is a mistake in the *universe of discourse* that the schema designer has made. They mistakenly believed they were identifying 'snapshot entities'. In reality they needed to identify 'an entity over its lifetime'. The domain they were considering was completely wrong, and hence the key they chose was completely incorrect.

This is exactly the same as the basic philosophical question of change. Are you the same person you were when you were 10? In one context the answer is absolutely not, so an appropriate identifier is not DNA. In another yes you are the same person, so DNA is perfect. It depends on the domain you are considering, and one must pick the correct key accordingly.

>
>
> Here's a simple example of what can happen:
>
> create table P
> (
> x int primary key,
> y int
> )
> create table F
> (
> x int primary key references P(x),
> z int
> )
> insert P (x, y) values (1, 1)
> insert P (x, y) values (2, 3)
> insert P (x, y) values (3, 2)
> insert F (x, z) values (1, 5)
> insert F (x, z) values (2, 4)
> select * from P join F on (p.x = F.x)
> update P
> set x = case x
> when 1 then 2
> when 2 then 1
> end
> where x in (1, 2)
> select * from P join F on (p.x = f.x)
>
> RESULTS:
>
> Before:
> x y x z
> ------ -------- ------- --------
> 1 1 1 5
> 2 3 2 4
>
> After:
> x y x z
> ------ ------- -------- --------
> 1 3 1 5
> 2 1 2 4
>
>
> Should the new values for x in P have cascaded into F? Assume that the
> system only has the before image and the after image (the first two columns
> in RESULTS) in order to complete the update. How can such a system
> differentiate between the above update and the following update?
>
> update P
> set y = case y
> when 1 then 3
> when 3 then 1
> end
> where x in (1, 2)
>
> Clearly first update affects a key, and consequently, the changes should
> cascade, but with the information available (both updates produce the exact
> same before and after images), the system cannot differentiate between the
> two updates; therefore, it cannot determine whether or not to cascade the
> changes. Updates within the Relational Model are are constrained in the
> same way as this hypothetical system. All that is available is the
> preceding instance and the succeeding instance, and if the only key can
> change, then there is no way to correlate tuples.

I'll consider the example in due course. RL work to do in the meantime, sigh. All best, Jim. Received on Tue Aug 15 2006 - 19:15:26 CEST