Hi,

I have a question regarding best practices in database design. In a
relational database, is it wise/necessary to sometimes create tables
that are not related to other tables through a foreign Key
relationship or does this always indicate some sort of underlying
design flaw. Something that requires a re evaluation of the problem
domain?

The reason I ask is because in our application, the user can perform x
number of high level operations (creating/updating projects, creating/
answering surveys etc. etc.). Different users can perform different
operations and each operation can manipulate one or more table. This
part of the system is done and working. Now there is a requirement to
have some sort of audit logging inside the database (separate from the
text based log file that the application generates anyway). This
"audit logging" table will contain high level events that occur inside
the application (which may or may not relate to a particular
operation). This table is in some sense related to every other table
in the database, as well as data that is not in the database itself
(exceptions, external events etc.). For example : it might have
entries that specify that at time x user created project y, at time A
user filled out survey B, at time C LDAP server was down, At time D an
unauthorized login attempt occurred etc.

As I said, these seems to suggest a stand alone, floating table with a
few fields that store entries regarding whats going on the system
without any direct relationship to other tables in the database. But I
just feel uneasy about creating such an isolated table. Another option
is to store the "logging" information in another schema/database, but
that doubles the maintainance work load. Not really looking forward to
maintaining/designing two different schemas.

I had a look at the microsoft adventureworks database schema diagram
and they seem to have 3 standalong tables - AWBuildVersion, ErrorLog
and DatabaseLog (unless i am reading it wrong!)

Any advice, Information or resources are much appreciated.

On 25 Jun, 08:08, nyathan... (AT) hotmail (DOT) com wrote:
Hi Nyathan,

In general terms it is quite acceptable to have a standalone table
with no FK relationships instansiated.
Indeed, in times gone by whole databases were created in this manner
as the overhead for OLTP with all the index data manipulation behind
the scences could bring a system to its knees. (PK & FK are backed by
'hidden' indexes).

As regards your specific task, (and here you'll realise I'm not from a
SQL Server background), check to see if there is an existing audit
ability in the rdbms.
Some rdbms allow for tracking all the activity on a table or all the
activity of a certain user, some allow both and some changed from one
to the other, (rats).

If you need to create your own system then think generic.
A single table with datetime stamp, username, table effected, action
taken, and the say 100 columns of varchar 100.
Each table you are seeking to audit will need triggers, ( insert,
update, delete), that calls a genric stored procedure that populates
the table you have created.
If your going to be doing this for a lot of tables or repeatedly I'd
advise writting a little noddy program to get the source table column
information from the systables and then generate the triggers.
The triggers will need to cause the stored procedure to write away
both the before and after image of the data.

BEWARE of bulk updates or deletes, ( best to disable the triggers
before you do them).
Also archiving / purging of your generic table becomes interesting,
depending on activity levels.

I would not recommend the above for busy tables.

Hope that helps, Tim

<nyathancha (AT) hotmail (DOT) com> wrote

An isolated table that is not logically connected to the rest of the system
by foreign key references suggests something that, in concept, is really a
separate schema. Ordinarily the Universe of Discourse (the subject matter
that the data describes) is composed of items that are all related to each
other in one way or another. The schema that is derived from a model of the
Universe of discourse will likewise be interrelated.

The situation you describe may be just such a situation, a separate schema.
An audit logging table may be storing data not for what it says about the
"real world" subject matter that the rest of the data describes, but for
what it says about the series of events that the application processed.

If so, it could be legitimate design. Ultimately, the question boils down
to this: how do you intend to use the data captured in the logging table?
If it's going to be used in an isolated fashion (not combined with data in
other tables by the DBMS), then your design may well be legitimate.

(nyathancha (AT) hotmail (DOT) com) writes:
I have a couple of tables in the database that I work with that for
one reason or another have neither in- or outgoing foreign keys. And
I can't say that it keeps me awake at night.

In the end what matters a lot is how this data is going to be used. I've
been floating the idea for our app, to have triggers that captures the
after image in an xml column together with table name and keys. Then
it would be simple to write a generic application that reads the log
entries for a certain entity (an account, a customer, whatever) and
displays only the changed columns.

The table would essentially be a database in itself about the rest of
the database.

(Our system also have audit and history tables that do have FK relationships
to the mother tables.)


--
Erland Sommarskog, SQL Server MVP, esquel (AT) sommarskog (DOT) se

Books Online for SQL Server 2005 at
http://www.microsoft.com/technet/pro...ads/books.mspx
Books Online for SQL Server 2000 at
http://www.microsoft.com/sql/prodinf...ons/books.mspx

On Mon, 25 Jun 2007 00:08:45 -0700, nyathancha (AT) hotmail (DOT) com wrote:

Hi nyathancha,

It can happen, but it's definitely not common. For me, it would be a
reason to look again, but not to dismiss the design right away.

I have once encountered a situation where I needed unrelated tables.
This had to do with auditing, but not at all like the method you are
proposing - in fact, I don't really like what I think you're trying to
do. Having one table to log "everything" shares many of the problems of
the EAV design - you'll be creating a very generic table with a few very
generic columns. They can hold everything, making it virtually
impossible to constrain or query the data in the table. If this is the
kind of audit tale that should normally never be used but is only kept
for the 1 in a million chance of a completely unforeseen disaster, and
wasting countless man hours to sift manually through the collected data
is an acceptable price to pay in that situation, than this design MIGHT
be considered. In all other cases, I'd steer away from it and go for a
more constrained design.

The situation where I had to use unrelated tables was at a firm that had
to keep a full record of changes for some tables - so for each of those
tables, a history table was made with all the same columns, plus a
datetime (as part of the primary key), userid of who made the change,
etc. We then added triggers to the main tables to ensure that each
modification in those tables was properly recorded in the corresponding
history table. But we did NOT define any foreign keys, for the simple
reason that after e.g. a deletion of a customer, the change history of
that customer still had to be kept on file; we couldn't remove the
customer from the history table, and a foreign key to the customers
table would have prevented the DELETE.

Another example of a design with an unrelated table that I never used in
practice but can imagine easily enough, would be a single-row table to
hold an application's "processing date" (so that a batch that runs past
midnight can all be processed as if all was on the same date, and that
actions can be "redone" [or tested] on a simulated dy - I have worked
with such systems back in my mainframe PL/I programming days, but they
used flat files rather than databases <g>).

For examples of good design, please don't look at any Microsoft supplied
samples. Even though AdventureWorks is miles ahead of Northwind and
pubs, it's still filled to the brim with bad practices.

--
Hugo Kornelis, SQL Server MVP
My SQL Server blog: http://sqlblog.com/blogs/hugo_kornelis

In addition to the other replies, I would add that foreign key constraints
are just one of many tools the database designer can use to help ensure that
bad data does not get placed in your database. Other tools include check
constraints, using the right datatypes (eg, store dates in a datetime
column, not a varchar column), sometimes triggers, etc.

So an important question is what the consequences will be if (when!, my
experience is if bad data can be put into a database, sooner or later, it
will be) invalid data is put into your audit table(s). That might range
from nobody really cares, to it's going to be a lot of work to fix it, to
somebody (you?) gets fired, to your company would be subject to a
significant fine, to somebody might go to prison (if, for example, your
audit trail is being used to prove compliance with SOX). So ask yourself
questions like what will happen if your boss comes to you and says the audit
trail says that user x created project y at time z, but there is no project
y in the system.

I certainly have tables in databases I have designed that do not have any
foreign key relationships to other tables, but before implementing one, I
would always think carefully about it.

Tom

<nyathancha (AT) hotmail (DOT) com> wrote

On Jun 26, 10:29 am, "Tom Cooper"
<tomcoo... (AT) comcast (DOT) no.spam.please.net> wrote:
Thanks for the prompt replies everyone.


One good point everyone seems to raise is "what is it used for?" ...
To be perfectly honest I am not entirely sure myself. Its one of those
requirements that filtered down from the management cloud. I think
the view is to use it mainly for "reporting" kind of functionality and
maybe only on some rare occasion for some sort of postmortem
debugging. Although in the latter situation, the application logs and
the sql server logs will probably end up being more helpful. I think
there is a system table somewhere in sql server that logs all the
transactions and changes that happen in the table right?

Crystal reports were being considered at some stage for more
sophisticated reports, but for now they want some sort of entries in
there to see whats happening (not necessarily at the database level,
but at the application level). The resolution of the reporting and
entries hasn't been decided yet ... as in, do we want to know
everytime someone retrieves a list of customers or only when someone
adds/removes customers. I have a feeling that if I chase this up, the
answer is going to be "both", "we may not want to start logging very
detailed stuff into the database right away, but if at some stage we
want to do it, the design should allow for it."

So just thinking in terms of some sort of "reporting" solution, in
abstract a sort of condensed data for easier consumption, does it make
sense to store an isolated table(s)/schemas along with the actual
data?

As to the consequences of a bad audit trail/log entry, I don't think
it would be catastrophic (fines, people going to prison etc.). Its an
internal application used to streamline inhouse processes. But of
course, we still don't want bad, inconsistent data in there and it
would lead to a lot of headaches, finger pointings, late nights etc.

<nyathancha (AT) hotmail (DOT) com> wrote


In architecture, form follows function. If the question "what is this data
being captured for" is unanswered, then the question of whether the design
is appropriate becomes moot.

<nyathancha (AT) hotmail (DOT) com> wrote

No. It doesn't. And depending on database options, the transaction log may
not contain a persistent record of changes either, not to mention that
special tools are required to read those logs.

It is sometimes important to determine who did what, when, where (from which
workstation), and how (with which application). It's easier to track down a
bug if you don't have to rely on what the user claims they did. It can also
be used to determine which users need additional training. A DBMS can only
guarantee that the result of a modification is consistent with respect to
the schema, it cannot determine whether or not a consistent modification is
also correct.

I think it all depends on the requirements. Logging consumes resources and
therefore impacts all database operations.

Actually, you DO want bad information in there. Nobody's perfect, and it's
easier to find and correct mistakes if you have a record of how they got
into the database in the first place.

On Jun 26, 11:08 am, nyathan... (AT) hotmail (DOT) com wrote:
Actually, another best practices question now that I am here. Does it
make sense for a table to have two (or more different foreign keys)
both (or all) of which can be nullable and then tie them to different
tables for different records? For example I have a survey table. It
has all the fields are relations for describing various survey data
(survey questions, participants, start, finish dates etc. ). Now a
survey can be related to a project or a supplier. Of course, the same
thing can be done with two different junction tables. Which is the
better method? Add the junction tables and increasing the number of
tables, complexity of the system (and the number of joins required for
a query) or just adding extra nullable foreign key field(s) to the
table? Is there a rule of thumb I should be following here?