----- Original Message -----
From: "Martijn Tonies" <[EMAIL PROTECTED]>
To: <mysql@lists.mysql.com>
Sent: Wednesday, March 15, 2006 10:47 AM
Subject: Re: Accountability with MySQL
Hello,
(please read all before replying)
>> >> > Basics of database design: store what you know.
>> >> >
>> >> > Given that NULLs basically means the absence of a value
>> >> > (eg: unknown), you shouldn't be storing NULLs.
>> >> >
>> >> Nonsense!!
>> >
>> > That's a bold statement ...
>> >
>> >> That's simply wrong. A null means "unknown or not applicable" and
>> >> is
a
>> >
>> > Ah, so now things become clear, a NULL actually can mean
>> > two things? So much for clarity then...
>> >
>> Come on; that's not fair. "Unknown" and "not applicable" are more like
>> different senses of the same thing, not two opposite things.
>
> Read the literature on how to design databases. What you do
> is storing "true propositions". That is, each attribute defines a
> certain "true proposition". For example:
>
> Employee "Martijn" has Employee# 14.
>
> You should be able to derive these sentences from every row.
> Putting a NULL or N/A in there fails to meet this requirement.
>
Perhaps my theory is a bit rusty but I have never heard of this "true
propositions" business before. Do you have a citation where I can read
about
this?
I would certainly recommend this one:
http://www.amazon.com/exec/obidos/ASIN/0321197844/databasede095-20/103-0880452-0854264?creative=327641&camp=14573&adid=0VQ1KDVJ56MV7389RPEQ&link_code=as1
Hmm, I had hoped for a citation to a free online document that I could read.
I don't fancy the thought of spending $100+ US on a book just to argue the
fine points of database design....
Then again, I didn't ask for a link to a free source of this information so
I can't fault you for your link, it (presumably) answers the question.
Frankly, I'm skeptical about your interpretation of this idea. While you
clearly don't want false information in a database, it isn't false to say
that student such-and-such's exam mark or such-and-such an employee's
termination date is unknown or not applicable. It _would_ be false to
store
a grade of 0 for a student who didn't take the exam because it implies
that
he got every question wrong when in fact he didn't write the test at all;
Agreed.
that would be an example of 0 having two contradicatory meanings. Given a
phobia about nulls, it is a reasonable design to put students who don't
take
the test in a separate table but it's not the ONLY way to handle the
situation: a null to indicate a student who didn't take the test is
perfectly reasonable too.
Disagreed. Let me explain further:
In this case, a "null" means (according to you and in your design)
that the student did not take the test.
Yes.
It could also mean that we lots it's grade for now, and that we
want to fill it in later, as in: "unknown", which is what NULL also
means (your own words).
In another case, null would be the employee isn't terminated yet,
but could be in the future.
That's the problems with nulls -> there's no actual defined
meaning.
I agree that the meaning is not a single, very narrow thing like 'zero' or
'unknown' in every single case. You have to put a null into some kind of
context. But that context is usually fairly obvious, at least in my opinion.
You can never say:
"This table holds test results, a row will mean:
Student with StudentID 123 made test #12 and scored a 4"
Cause you can have rows that can mean this:
Student with StudentID 123 made test #12 and scored a
yet unknown score
or
Student with StudentID 123 made test #12 and did not make
the test
or
Student with StudentID 123 and test #12 doesn't apply to him
If the exact reason that a null is in a particular row is particularly
important, you can always define an extra column that describes the reason
for the null. "A" could mean the score hadn't been entered yet because the
exam hadn't taken place yet, "B" could mean the test has already taken place
but the student didn't write the exam, "C" could mean that student has no
score because he was excused from the test for some reason. I would agree
that this is cumbersome but I think you'd have a very similar problem with
your design: you would have put some people in a separate table if they
didn't write the test but you'd have to identify the reason that the test
wasn't written too, if you cared about that, or have separate tables for
students who failed to attend the test and students who were excused from
the test. So I don't see that you're a lot farther ahead with your design.
In fact, your design could be _more_ complicated. In your design a program
couldn't simply look up the student by his ID in a single table and then
react to a null; the program would have to look in the main Grades table
first, then, if it didn't find the student there, would have to look in the
table that records people who didn't write the test because they were sick.
If the row wasn't found there, it would have to look in the table that
records people who were excused from the exam in advance. Frankly, I think
most programmers would rather look in one table than have to check two or
three or more for a given student.
In short: NULL is ambigious.
It _can_ be. But the context frequently makes its meaning obvious.
>> >> perfectly valid value to use in many, many situations. Two classic
>> > examples:
>> >
>> > Let me first state that there's a difference between storing NULLs
>> > and handling NULLs in your result.
>> >
>> Yes, of course there is: so what? You said NULLs should never be
> _stored_;
>> that's what I'm responding to. Using NULLs is not very _difficult_
>> anyway,
>> just a bit tedious because it is another case to handle.
>
> There's no point in storing what you don't know.
>
But you _do_ know something: that the student didn't take the exam. That
is
a piece of true knowledge all by itself.
But that's not what the rows in the table would mean - see above.
In the relational model, the idea is to store true propositions - that is,
each row means the same thing, but with different attributes. See above.
>> > So, let's debunk these two classic examples...
>> >
>> Please note that I did not say that you MUST use NULLs, just that they
> were
>> perfectly valid to use in a design.
>
> Depends on who you're asking ...
>
> A few years ago, I was working at a company that had developed
> an application with an Oracle database with about 400 tables, nothing
> too large.
>
> Plenty of tables had NULLs, not because of missing business data,
> but rather used internally in the system.
>
I'm not sure what you mean by that phrase "used internally in the
system".
Missing business data, eg: no middle name. Some people would
put in a NULL in there as well, while NULL can also mean: I do
not KNOW this persons middle name. Ambiguity, see above.
Agreed, a null might mean "this guy doesn't have a middle name" or "I didn't
ask him for his middle name so I don't know if he has one". That may or may
not be important; it could be very important if we were trying to
distinguish between George Herbert Walker Bush (the 41st president of the
US) and George Walker Bush (the current president of the US). But null isn't
really a problem in this situation so much as a symptom of the real problem:
if the middle name is important, why aren't I determining the middle name(s)
of the person when I add him to the database? The null is not dishonest in
this case; it correctly states that the first and last names of both men are
"George" and "Bush". But our input process is weak because we didn't realize
that such cases could occur so we failed to ensure that our programs were
capturing middle names.
> The number of problems we had with them, I cannot count on my
> two hands, nor can I count them on the hands of all other developers
> (5) for that project.
>
How about enumerating a few of them here? I've dealt with relational
databases in several capacities for over 20 years and I've never found
nulls
to be a big problem.
Not big, but a problem... :-)
Do you often go around looking for (little) problems while you
can easily avoid them?
In most of the cases you've described, you don't really avoid problems so
much as trade them for different problems.
If you store students who don't take the test in one table and students who
are excused in advance in another table and students who took the test in
yet another table, you avoid having to deal with nulls but you introduce new
problems: you have to manage three tables now - more backups, probably more
space consumed, etc. - and your programs have to have more sophisticated
logic to find a given student. You may find that the problems you've avoided
were smaller problems than the ones you traded them for. Or not. As usual in
database design, "it depends".
> Just last week, I spoke to a guy who used to be my collegue there
> and we discussed the design a bit (he still does database design and
> development for a large company, on Sybase with over 4000
> deployed systems) and we both agreed that using NULLs in most
> cases as part of the initial design was causing us more problems
> then helping us.
>
Again, how about listing some of these problems?
Well, for one, we had a table that stored all kinds of different
actions that should take place. Each action could be part of
a list of actions.
Someone decided that actions with a NULL for the "belongs
to list of actions #" would be an action "by itself".
Business wise, one could create an order for a product and that
product had to be created by a list of certain actions. The order
creator would then use any pre-defined list of actions, or create
such a list himself by selecting the different actions.
Asking for stand alone actions would then (obviously simplified):
select * from bewerking where list is null
This is a slow query. Period. One that could have easily been
avoided by not storing NULLs, but using a separate table.
Why a table? Because "standalone actions" do not equal
"actions that belong to a list".
Sorry, your example wasn't sufficiently detailed for me to really follow it.
I really don't understand the scenario you are describing.
>> Okay, that might be acceptable, if it doesn't cause you to lose track
of
> the
>> student altogether. But if this table was the only one that even
recorded
>> the _existence_ of the student, you'd have a problem;
>
> Indeed, then I would have a problem, cause you cannot derive from
> the "test results" table that a student exists. This is a rather silly
> statement.
> If the "test results" table should ALSO store the existence of a
student,
> you're design is wrong :-)
>
>>if someone tried to
>> verify that the student had attended this school, you wouldn't know
that
>> they had.
>
> In your own words: nonsense.
>
> Each table should store what is was designed to store. If I would
> have to know if a student was enlisted in a certain course or would
> be attending this school, I would not be using the "test results"
> table.
>
Look, I agree that this scenario is not particularly likely to happen but
it
COULD happen.
It's not me who made up the example :-)
Well, I shouldn't have lead off with a weak example but it was the first one
I thought of....
Nevertheless, even an improbable scenario should be a concern if it is
possible that it can happen, even if the likelihood is fairly low. If you
can rule out the possibility altogether, then you can dismiss it as a
concern.
I think it is POSSIBLE that an exam table might be the only record of a
student's existence, even if it is unlikely. I've certainly seen some pretty
weak designs proposed on this and other newsgroups/lists over the years.
Newbies in particular are prone to be very weak in design. Just yesterday, I
saw a post where someone wanted to store the purchase of up to 5 stocks in a
single table. That, of course, is terrible design since it causes a
repeating group, which violates First Normal Form. But the poster had no
idea of the bad consequences and was quite prepared to build such a table.
Suppose you run a small company that runs certification exams:
your only contact with the students is that they show up to take an exam
and
you record the mark and send it to whoever awards the certification. In
that
case, you likely wouldn't have an elaborate set of tables containing a
variety of student information the way you would if you were a
university.
In that case, it's quite reasonable to believe that the only table in the
system is the one that records the test marks. That table may include the
student name, address, test name, score, etc. etc. That table may only be
populated as the student comes in the door: the 'greeter' welcomes the
student to the test center, asks his name and other details, points him
to
the exam room, and records the test result when the student hands in the
exam. In that scenario, if student Joe Blow does not show up for the test
because his car broke down, he may never be entered in the system at all
and
you have therefore lost sight of the fact that he ever existed. Again,
this
is not a very likely scenario but it is not an impossible one and that's
the
key point: if it COULD happen, you will have a problem.
See above. There's no point in taking exams from random people.
You are trying to come up with all sorts of unlikely scenarios to
justify a solution that's, IMO, faulty.
Unfortunately, I'm having trouble thinking of better scenarios at the
moment. Still, the scenarios I've given are certainly possible,
particularly in a small shop where the database guy is very inexperienced.
>>Or in a more probable case, if that was the only test for that
>> course and the student missed it and then had no row in the table, you
> might
>> not have any way of knowing that they took the course!
>
> See above.
>
Dismissing the less probable case does not dismiss this more probable
scenario.
Wrong problem description for task at hand. See above (my previous
reply, not this one). Attack each problem with the right solution, don't
try to find problems in solutions that aren't meant to solve the problem
you just found.
>>And if they later
>> wanted to write the exam, having recovered from their illness, your
query
>> might have the effect of keeping them from taking the second exam:
>> your
>> query would report that they had never taken the first exam so an
>> adminstrator might refuse to let him/her take the second exam because
>> they
>> (apparently) had never been scheduled to take the first one.
>
> Different problem, see above.
>
Again, you can't dismiss the second scenario by dismissing the first one;
the scenarios are different and the second one is more probable than the
first.
See above. The "Different problem" remark still holds.
>> Now, you could have a second table to record people who were scheduled
to
>> take tests but failed to take them to cover that situation but I think
it
>
> Sounds like a decent design to me.
>
Not surprisingly since it is the one you suggested as the "correct" way
to
handle the situation. But the design that uses nulls is NOT invalid and
you
have not demonstrated that it is.
What I DID demonstrate, however, is that it's _very_ easy to
create a design without NULLs, that is much clearer to read,
doesn't have to handle NULLs, avoids NULLs in general and
make it easy to understand what problems are solved.
Considering maintainability, ability to refactor and without having
to think up missing data, I prefer my design.
Fine, you PREFER your design: I can live with that. I just took exception to
the idea that a design using nulls was inherently bad, or wrong, or
unworkable.
Do note I don't claim to have the key to every problem, but I
would like to know people that design WITH NULLs doesn't
have to be the standard although SQL provides them.
Agreed; you don't have to use nulls. In fact, some shops put a "NOT NULL"
beside every column definition in every CREATE TABLE statement. But many
shops _don't_ do that and have perfectly workable database designs.
>> would be easier to record all students in one table and then simply
store
> a
>> null for any test that they fail to take and a zero for every case
where
>> a
>> student got every answer wrong. When you compute the class average,
>> the
>> avg() function would ensure that the students who got every question
>> wrong
>> would pull down the class average but that students who failed to
>> write
> the
>> test at all would NOT skew the average because the avg() function
ignores
>> nulls.
>
> Ah, an excellent example of why NULLs are tedious: they are ignored
> by some functions, but not by others.
>
_ALL_ of the column functions ignore nulls: it is NOT inconsistent. In
fact,
it is the only logical thing that these functions can do, which you'd
realize if you gave it a bit of open-minded thought.
Don't accuse me of not being open minded. I did develop systems
with NULLs before. I became a database developer back then
from practicing. However, I'm beginning to understand the theory
more and more and dislike NULLs more and more because of it.
Suppose you had a table with a primary key of employee number and a
column
containing bonuses. Some of the rows contain nulls in the bonus column
because the employee didn't earn a bonus yet. Now, you need to answer
these
questions: what is the largest bonus amount? What is the smallest bonus
amount? What is the sum of the bonuses? What is the average bonus? Since
a
null, by definition, is unknown, it isn't bigger or smaller or equal to
any
other value, including another null so it can't qualify as the biggest
value
or the smallest value. A number plus (or minus, or times, or divided by)
a
null is undefined so it makes no sense to add it to the sum of the
bonuses.
A null is NOT a zero so treating it as a zero in an average would unduly
skew the result. So, in each case, the function does the logical thing
and
IGNORES the null.
> Without having the record there in the first place, there would have
> to be no rule of AVG ignoring NULLs. Problem solved.
>
That is the other way of the solving the problem. But it is NOT the ONLY
way.
>> > Why store a date column if you don't know?
>> >
>> > Why not use:
>> >
>> > EMPLOYEES
>> > EmployeeID int,
>> > StartingDate Date,
>> > ...
>> >
>> > TERMINATED_EMPLOYEES (albeit a bit agressive ;) )
>> > EmployeeID
>> > TerminationDate
>> >
>> > Once more: perfectly valid design.
>> >
>> Yes, that is also a valid design but it means you have to have yet
>> another
>> table that you could have avoided simply by permitting a null in the
>> employee table. Hey, if you really want your tables to proliferate
>> like
>> this, that's up to you.
>
> It seems you have a fear of creating new tables ;-)
>
No, I don't. It seems to me that you have an undue fear of nulls.
> This is what database systems are designed to do...
>
Database systems are designed to accomodate large quantities of tables
but
any designer with any real world experience will advise you not to create
tables simply for the sake of creating tables: that is an unnecessary
proliferation of tables. Look at normalization. As you probably know,
each
phase of normalization tends to increase the number of tables in the
database. I know for a fact that the theorists have identified at least
up
to 17NF (17th Normal Form) but how many levels of normalization do
businesses typically use in their normalizations? Answer: Three.
(Occasionally four). And I think you'll find that the main reason is that
the benefits of normalizing beyond 3NF are outweighed by the increasing
number of tables to manage.
>> The key point is that NULLs _do_ work and are a legitimate design
> decision.
>> You don't _have_ to use them but they can save you some work and
>> reduce
> the
>> number of tables you need. (They can also be a bit more work on the
>> programming side.)
>>
>> I took your remarks to mean that NULLs were always a bad idea and were
a
>> symptom of bad design and I strongly disagree with that.
>
> So I've noticed.
>
I'm trying to be gracious here and accept that I may just have
misunderstood
the emphasis in what you are saying.
See above.
> The message I'm trying to get across is that:
> 1 - according to (proper) design literature, you should not use NULLs
> if you don't have to
Perhaps you can cite some of this "proper" literature. I'm interested in
seeing their definition of what an appropriate time to use nulls is.
> 2 - NULLs can cause you more problems than you can think of
I'd like to hear what these problems are. I've never had any big problems
because of nulls.
Ah well, SQL (these days) provides plenty of functions to avoid
the problems. Coalesce being one of them...
How often do people write:
(probably not valid MySQL)
select firstname || coalesce(middlename || ' ', ' ') || lastname
from ...
Your syntax isn't quite correct; coalesce doesn't use a concatenation
symbol, it uses a list of arguments like this:
coalesce(Salary, commission, bonus)
Actually, coalesce has been around for a long time. DB2 (on the mainframe)
had it starting in Version 1.3 if I recall correctly; it was called values()
then but it worked exactly the same as coalesce. The SQL standard decided to
call it coalesce so DB2 developers just created a new entry point for the
function and the now both function names can be used.
But MySQL is quite a bit later to get some of these functions because it is
a much younger product.
> 3 - its sometimes much easier to avoid storing NULLs AND to be
> able to refactor your database because of it
I agree that nulls can frequently be avoided and that avoiding them will
simplify _some_ situations.
> 4 - the meaning of NULL can change, so why store it in the first place
>
Simple: because it happens to be true when it is stored. In the case of
the
hypothetical employee, I store a null termination date when I hire him
because I don't know when he is going to leave.
Why store something that you -don't- know.
To acknowledge that the employee will leave eventually and is still a
perfectly valid employee even though you don't know the termination date.
If he laters gives his
notice, then I know when he is leaving and can store that date for his
termination date instead of a null; then his row of the table is true
again,
based on the new facts.
>> If you are simply saying that you don't like them and prefer to use
>> different designs to avoid them, then I don't have any problem with
that.
>
> I have seen that when I avoid storing NULLs, my applications
> became more clear and easier to understand.
>
Beauty is in the eye of the beholder as we all know. I have no problem
with
having nulls in my tables and consider that a better design than a
separate
table for special cases most of the time.
Again, if you are saying that you don't like nulls and prefer to avoid
them
in your designs, I have no problem with that; that's just your personal
preference. It's the same as if we sat down to eat a meal and you asked
for
chocolate ice cream for dessert and I had vanilla; neither choice is
"wrong", they are just personal preference.
Well, to that we agree and I would order sugared whipped cream with it ;-)
I guess you're not diabetic :-) Yet ;-)
But if you are stating categorically that nulls are always "bad" or
"wrong",
I disagree strongly. That's like saying only chocolate ice cream is
acceptable and that all other flavours are evil. That's just wrong.
We are not going to agree, that much is clear.
I think we raised some very interesting points though (which is exactly
why I started this in the first place).
I would avoid NULLs if I could. It makes - very often - more sense
to avoid them than to include them.
You say that it's not "bad" to have NULLs every now and then, a
statement to which I can agree up to a certain height.
Okay, that's really all I was trying to say. Your initial tone seemed to be
that nulls were always bad and should never be used; that's what I was
disagreeing with.
I would also like to state, that I've seen design where NULLs is
grossly overused for all sorts of situations. Your simple examples
can be among those. IF you're using NULLs, you'd better be aware
of what you're doing. Just don't use NULL for the sole purpose of
avoiding a table or relation.
Agreed.
I'd still say that in general, the "true proposition" remark holds and
is very very valid. With that as the basis, design your database.
I think I've also described why using NULLs in a "true proposition"
makes them invalid. I would advice against that.
As long as you accept that nulls are not always evil and are actually
sometimes valid and useful, I don't think we have a problem. Clearly, if you
and I were both assigned to design a database for the same specifications,
yours would have more tables and fewer nulls than mine. I'd still expect
both of our designs to work satisfactorily.
This topic will pop up every now and then. If there's one thing that
I would like, is that people stop using NULLs way too often, I think
we can agree on that.
I for one had many big and small problems with NULLs stored in
the database. That made me appreciate more how easy it is to avoid
them and how to handle situations differently.
I have no problem with that point of view. This is just a difference of
opinion or emphasis that two reasonable people can have.
--
Rhino
--
No virus found in this outgoing message.
Checked by AVG Free Edition.
Version: 7.1.385 / Virus Database: 268.2.3/281 - Release Date: 14/03/2006
--
MySQL General Mailing List
For list archives: http://lists.mysql.com/mysql
To unsubscribe: http://lists.mysql.com/[EMAIL PROTECTED]
