I follow a few rules:

1. Primary keys should be as small as necessary. Prefer a numeric type because numeric types are stored in a much more compact format than character formats. This is because most primary keys will be foreign keys in another table as well as used in multiple indexes. The smaller your key, the smaller the index, the less pages in the cache you will use.
2. Primary keys should never change. Updating a primary key should always be out of the question. This is because it is most likely to be used in multiple indexes and used as a foreign key. Updating a single primary key could cause of ripple effect of changes.
3. Do NOT use "your problem primary key" as your logic model primary key. For example passport number, social security number, or employee contract number as these "natural keys" can change in real world situations. Make sure to add UNIQUE constraints for these where necessary to enforce consistency.

On surrogate vs natural key, I refer to the rules above. If the natural key is small and will never change it can be used as a primary key. If the natural key is large or likely to change I use surrogate keys. If there is no primary key I still make a surrogate key because experience shows you will always add tables to your schema and wish you'd put a primary key in place.

Natural verses artifical keys is a kind of religious debate among the database community - see this article and others it links to. I'm neither in favour of always having artifical keys, nor of never having them. I would decide on a case-by-case basis, for example:

• US States: I'd go for state_code ('TX' for Texas etc.), rather than state_id=1 for Texas
• Employees: I'd usually create an artifical employee_id, because it's hard to find anything else that works. SSN or equivalent may work, but there could be issues like a new joiner who hasn't supplied his/her SSN yet.
• Employee Salary History: (employee_id, start_date). I would not create an artifical employee_salary_history_id. What point would it serve (other than "foolish consistency")

Wherever artificial keys are used, you should always also declare unique constraints on the natural keys. For example, use state_id if you must, but then you'd better declare a unique constraint on state_code, otherwise you are sure to eventually end up with:

state_id    state_code   state_name
137         TX           Texas
...         ...          ...
249         TX           Texas

Just an extra comment on something that is often overlooked. Sometimes not using a single surrogate key as primary has benefits in the child tables. Let's say we have a design that allows you to run multiple companies within the one database (maybe it's a hosted solution, or whatever).

Let's say we have these tables and columns:

Company:
  CompanyId   (primary key)

CostCenter:
  CompanyId   (primary key, foreign key to Company)
  CostCentre  (primary key)

CostElement
  CompanyId   (primary key, foreign key to Company)
  CostElement (primary key)

Invoice:
  InvoiceId    (primary key)
  CompanyId    (primary key, in foreign key to CostCentre, in foreign key to CostElement)
  CostCentre   (in foreign key to CostCentre)
  CostElement  (in foreign key to CostElement)

In case that last bit doesn't make sense, Invoice.CompanyId is part of two foreign keys, one to the CostCentre table and one to the CostElement table. The primary key is (InvoiceId, CompanyId).

In this model, it's not possible to screw-up and reference a CostElement from one company and a CostCentre from another company. If a single surrogate key was used as primary on the CostElement and CostCentre tables, and without the foreign key relations in the Invoice table, it would be.

The fewer chances to screw up, the better.

There´s no problem in making your primary key from various fields, that's a Natural Key.

You can use a Identity column (associated with a unique index on the candidate fields) to make a Surrogate Key.

That´s an old discussion. I prefer surrogate keys in most situations.

But there´s no excuse for the lack of a key.

RE: EDIT

Yeah, there´s a lot of controversy about that :D

I don´t see any obvious advantage on natural keys, besides the fact that they are the natural choice. You will always think in Name, SocialNumber - or something like that - instead of idPerson.

Surrogate keys are the answer to some of the problems that natural keys have (propagating changes for example).

As you get used to surrogates, it seems more clean, and manageable.

But in the end, you´ll find out that it's just a matter of taste - or mindset -. People "think better" with natural keys, and others don´t.

Tables should have a primary key all the time. When it doesn't it should have been an AutoIncrement fields.

Sometime people omit primary key because they transfer a lot of data and it might slow down (depend of the database) the process. BUT, it should be added after it.

Some one comment about link table, this is right, it's an exception BUT fields should be FK to keep the integrity, and is some case those fields can be primary keys too if duplicate in links is not authorized... but to keep in a simple form because exception is something often in programming, primary key should be present to keep the integrity of your data.

A natural key, if available, is usually best. So, if datetime/char uniquely identifies the row and both parts are meaningful to the row, that's great.

If just the datetime is meaningful, and the char is just tacked on to make it unique, then you might as well just go with an identify field.

I suspect Steven A. Lowe's rolled up newspaper therapy is required for the designer of the original data structure.

As an aside, GUIDs as a primary key can be a performance hog. I wouldn't recommend it.

Natural versus artificial keys to me is a matter of how much of the business logic you want in your database. Social Security number (SSN) is a great example.

"Each client in my database will, and must, have an SSN." Bam, done, make it the primary key and be done with it. Just remember when your business rule changes you're burned.

I don't like natural keys myself, due to my experience with changing business rules. But if your sure it won't change, it might prevent a few critical joins.

I too always use a numeric ID column. In oracle I use number(18,0) for no real reason above number(12,0) (or whatever is an int rather than a long), maybe I just don't want to ever worry about getting a few billion rows in the db!

I also include a created and modified column (type timestamp) for basic tracking, where it seems useful.

I don't mind setting up unique constraints on other combinations of columns, but I really like my id, created, modified baseline requirements.

I look for natural primary keys and use them where I can.

If no natural keys can be found, I prefer a GUID to a INT++ because SQL Server use trees, and it is bad to always add keys to the end in trees.

On tables that are many-to-many couplings I use a compound primary key of the foreign keys.

Because I'm lucky enough to use SQL Server I can study execution plans and statistics with the profiler and the query analyzer and find out how my keys are performing very easily.

You should use a 'composite' or 'compound' primary key that comprises of multiple fields.

This is a perfectly acceptable solution, go here for more info :)

I always use an autonumber or identity field.

I worked for a client who had used SSN as a primary key and then because of HIPAA regulations was forced to change to a "MemberID" and it caused a ton of problems when updating the foreign keys in related tables. Sticking to a consistent standard of an identity column has helped me avoid a similar problem in all of my projects.

GUIDs can be used as a primary key, but you need to create the right type of GUID so that it performs well.

You need to generate COMB GUIDs. A good article about it and performance statistics is The Cost of GUIDs as Primary Keys.

Also some code on building COMB GUIDs in SQL is in Uniqueidentifier vs identity(archive).

All tables should have a primary key. Otherwise, what you have is a HEAP - this, in some situations, might be what you want (heavy insert load when the data is then replicated via a service broker to another database or table for instance).

For lookup tables with a low volume of rows, you can use a 3 CHAR code as the primary key as this takes less room than an INT, but the performance difference is negligible. Other than that, I would always use an INT unless you have a reference table that perhaps has a composite primary key made up from foreign keys from associated tables.

If you really want to read through all of the back and forth on this age-old debate, do a search for "natural key" on Stack Overflow. You should get back pages of results.

We do a lot of joins and composite primary keys have just become a performance hog. A simple int or long takes care of many problems even though you are introducing a second candidate key, but it's a lot easier and more understandable to join on one field versus three.

I'll be up-front about my preference for natural keys - use them where possible, as they'll make your life of database administration a lot easier. I established a standard in our company that all tables have the following columns:

• Row ID (GUID)
• Creator (string; has a default of the current user's name (SUSER_SNAME() in T-SQL))
• Created (DateTime)
• Timestamp

Row ID has a unique key on it per table, and in any case is auto-generated per row (and permissions prevent anyone editing it), and is reasonably guaranteed to be unique across all tables and databases. If any ORM systems need a single ID key, this is the one to use.

Meanwhile, the actual PK is, if possible, a natural key. My internal rules are something like:

• People - use surrogate key, e.g. INT. If it's internal, the Active Directory user GUID is an acceptable choice
• Lookup tables (e.g. StatusCodes) - use a short CHAR code; it's easier to remember than INTs, and in many cases the paper forms and users will also use it for brevity (e.g. Status = "E" for "Expired", "A" for "Approved", "NADIS" for "No Asbestos Detected In Sample")
• Linking tables - combination of FKs (e.g. EventId, AttendeeId)

So ideally you end up with a natural, human-readable and memorable PK, and an ORM-friendly one-ID-per-table GUID.

Caveat: the databases I maintain tend to the 100,000s of records rather than millions or billions, so if you have experience of larger systems which contraindicates my advice, feel free to ignore me!