Do you guys still use the model when you make your database for your websites?

My professor used to remind us to do that though others seem don't bother.

Absolutely. You can't have a database as a single point of knowledge that holds billions of rows that isn't normalized according to the relational model. The only way you could try to mimic that outside relational space is to federate the data out across millions of servers using some proprietary data storage and access methodology. Or, a single box with a great hard disc sub system and decent CPU, RAM, etc, will easily manage the same data and respond quickly.

Well in fact sometimes your should denormalize to achieve better performance, but in absolutely most cases it is only required under extreme highload.

My professor used to remind us to do that though others seem don't bother.

Yes, Usually the I read how to do this on the Internet "DBAs".

If you have a few hundred records that are being accessed by a handful of people then probably not.

When you scale up on on users and records, then it becomes pretty much essential.

What happens in a monolithic system when Miss P Jones marries Mr D Smith and wants all her data changing AND there are 6 different Miss P Jones's in the system?

Originally Posted by mtishetsky

Well in fact sometimes your should denormalize to achieve better performance, but in absolutely most cases it is only required under extreme highload.

Half right. You should denormalize to get better performance when

1. You have 10x more reads than writes.
2. Most of the queries coming in to the system have the same joins.
3. The joins are expensive.

That really only describes data warehouse reporting databases. Anything transactional, and the more load your system comes under, the more you need normalized structure. Instead of writing to one giant table to reflect a small change, under high load, you need to write to very narrow tables.

One important thing normalization gets you is less disc IO. And this is one of the best optimizations possible.

you SHOULD use 3rd normal form if possible. This way it will prevent update/delete anomilies.

Originally Posted by lindaadams

you SHOULD use 3rd normal form if possible. This way it will prevent update/delete anomilies.

Until you get 10M select requests daily.

Originally Posted by mtishetsky

Until you get 10M select requests daily.

Generally, yes, but it depends what kind of selects, or, more precisely, what's being selected. If you have an application with many users, that's constantly hitting the database to populate combo boxes, highly normalized schemas will give the best performance. If it's mostly* management reports, denormalized is best from a perf standpoint.

The best way to measure "mostly" is the total amount of time for SQL commands to execute. Counting the number of executions is meaningless, and resource utilization is tricky (logical vs physical, etc), but as long as you don't have an overwhelming problem with locks, duration is the best overall measure, and time is the real cost the organization pays.

OLTP tends to do best when normalized, and OLAP tends to benefit somewhat from denormilization.