For true link tables, they typically do not exist as object entities in my object models. Thus the surrogate key is not ever used. The removable of an item from a collection results in a removal of an item from a link relationship where both foreign keys are known (Person.Siblings.Remove(Sibling) or Person.RemoveSibling(Sibling) which is appropriately translated at the data access layer as usp_Person_RemoveSibling(PersonID, SiblingID)).

As Mike mentioned, if it does become an actual entity in your object model, then it may merit an ID. However, even with addition of temporal factors like effective start and end dates of the relationship and things like that, it s not always clear. For instance, the collection may have an effective date associated at the aggregate level, so the relationship itself may still not become an entity with any exposed properties.

I d like to add that you might very well need the table indexed both ways on the two foreign key columns.

I would use composite key, and no extra meaningless key.

I would not use a ORM system that enforces such rules on my db structure.

If this is a true many-to-many join table, then dump unecessary id column (unless your ORM requires one. in that case you ve got to decide whether your intellect is going to trump your practicality).

But I find that true join tables are pretty rare. It usually isn t long before I start wanting to put some other data in that table. Because of that I almost always model these join tables as entities from the beginning and stick an id in there.

Having a single column pk can help out alot in disaster recovery situation. So though while correct in theory that you only need the 2 foreign keys. In practice when the shit hits the fan you may want the single column key. I have never been in a situation where i was screwed because I had a single column identifier but I have been in ones where I was screwed because I didn t.

Composite PK and turn off clustering.

I have used composite key to prevent duplicate entry and let the database handle the exception. With a single key, you are rely on the front-end application to check the database for duplicate before adding a new record.

There is something called identifying and non-identifying relationship. With identifying relationships the FK is a part of the PK in the many-to-many table. For example, say we have tables Person, Company and a many-to-many table Employment. In an identifying relationship both fk PersonID and CompanyID are part of the pk, so we can not repeat PersonID, CompanyID combination.

TABLE Employment(PersonID int (PK,FK), CompanyID int (PK,FK))

Now, suppose we want to capture history of employment, so a person can leave a company, work somewhere else and return to the same company later. The relationship is non-identifying here, combination of PersonID, CompanyID can now repeat, so the table would look something like:

TABLE Employment(EmploymentID int (PK), PersonID int (FK), CompanyID int (FK), 
                     FromDate datetime, ToDate datetime)

If you are using an ORM to get to/alter the data, some of them require a single-column primary key (Thank you Tom H for pointing this out) in order to function correctly (I believe Subsonic 2.x was this way, not sure about 3.x).

In my mind, having the primary key doesn t impact performance to any measurable degree, so I usually use it.

If you need to traverse the join table in both directions , that is starting with a table1fk or a table2fk key only, you might consider adding a second, reversed, composite index.

ADD KEY ( table2fk ,  table1fk )

The correct answer is:

• Primary key is ( table1fk , table2fk )
• Another index on ( table2fk , table1fk )

Because:

• You don t need an index on table1fk or table2fk alone: the optimiser will use the PK
• You ll most likely use the table "both" ways
• Adding a surrogate key is only needed because of braindead ORMs

i ve used both, the only benefit of using the first model (with uid) is that you can transport the identifier around as a number, whereas in some cases you would have to do some string concatenation with the composite key to transport it around.

i agree that not indexing the foreign keys is a bad idea whichever way you go.

I (almost) always use the additional single-column primary key. This generally makes it easier to build user interfaces, because when a user selects that particular linking entity I can identify with a single integer value rather than having to create and then parse compound identifiers.