How do I create a unique constraint that also allows nulls?

SQL Server 2008 +

You can create a unique index that accept multiple NULLs with a WHERE clause. See the answer below.

Prior to SQL Server 2008

You cannot create a UNIQUE constraint and allow NULLs. You need set a default value of NEWID().

Update the existing values to NEWID() where NULL before creating the UNIQUE constraint.

SQL Server 2008 And Up

Just filter a unique index:

CREATE UNIQUE NONCLUSTERED INDEX UQ_Party_SamAccountName
ON dbo.Party(SamAccountName)
WHERE SamAccountName IS NOT NULL;

In Lower Versions, A Materialized View Is Still Not Required

For SQL Server 2005 and earlier, you can do it without a view. I just added a unique constraint like you're asking for to one of my tables. Given that I want uniqueness in column SamAccountName, but I want to allow multiple NULLs, I used a materialized column rather than a materialized view:

ALTER TABLE dbo.Party ADD SamAccountNameUnique
   AS (Coalesce(SamAccountName, Convert(varchar(11), PartyID)))
ALTER TABLE dbo.Party ADD CONSTRAINT UQ_Party_SamAccountName
   UNIQUE (SamAccountNameUnique)

You simply have to put something in the computed column that will be guaranteed unique across the whole table when the actual desired unique column is NULL. In this case, PartyID is an identity column and being numeric will never match any SamAccountName, so it worked for me. You can try your own method—be sure you understand the domain of your data so that there is no possibility of intersection with real data. That could be as simple as prepending a differentiator character like this:

Coalesce('n' + SamAccountName, 'p' + Convert(varchar(11), PartyID))

Even if PartyID became non-numeric someday and could coincide with a SamAccountName, now it won't matter.

Note that the presence of an index including the computed column implicitly causes each expression result to be saved to disk with the other data in the table, which DOES take additional disk space.

Note that if you don't want an index, you can still save CPU by making the expression be precalculated to disk by adding the keyword PERSISTED to the end of the column expression definition.

In SQL Server 2008 and up, definitely use the filtered solution instead if you possibly can!

Controversy

Please note that some database professionals will see this as a case of "surrogate NULLs", which definitely have problems (mostly due to issues around trying to determine when something is a real value or a surrogate value for missing data; there can also be issues with the number of non-NULL surrogate values multiplying like crazy).

However, I believe this case is different. The computed column I'm adding will never be used to determine anything. It has no meaning of itself, and encodes no information that isn't already found separately in other, properly defined columns. It should never be selected or used.

So, my story is that this is not a surrogate NULL, and I'm sticking to it! Since we don't actually want the non-NULL value for any purpose other than to trick the UNIQUE index to ignore NULLs, our use case has none of the problems that arise with normal surrogate NULL creation.

All that said, I have no problem with using an indexed view instead—but it brings some issues with it such as the requirement of using SCHEMABINDING. Have fun adding a new column to your base table (you'll at minimum have to drop the index, and then drop the view or alter the view to not be schema bound). See the full (long) list of requirements for creating an indexed view in SQL Server (2005) (also later versions), (2000).

Update

If your column is numeric, there may be the challenge of ensuring that the unique constraint using Coalesce does not result in collisions. In that case, there are some options. One might be to use a negative number, to put the "surrogate NULLs" only in the negative range, and the "real values" only in the positive range. Alternately, the following pattern could be used. In table Issue (where IssueID is the PRIMARY KEY), there may or may not be a TicketID, but if there is one, it must be unique.

ALTER TABLE dbo.Issue ADD TicketUnique
   AS (CASE WHEN TicketID IS NULL THEN IssueID END);
ALTER TABLE dbo.Issue ADD CONSTRAINT UQ_Issue_Ticket_AllowNull
   UNIQUE (TicketID, TicketUnique);

If IssueID 1 has ticket 123, the UNIQUE constraint will be on values (123, NULL). If IssueID 2 has no ticket, it will be on (NULL, 2). Some thought will show that this constraint cannot be duplicated for any row in the table, and still allows multiple NULLs.

What you're looking for is indeed part of the ANSI standards SQL:92, SQL:1999 and SQL:2003, ie a UNIQUE constraint must disallow duplicate non-NULL values but accept multiple NULL values.

In the Microsoft world of SQL Server however, a single NULL is allowed but multiple NULLs are not...

In SQL Server 2008, you can define a unique filtered index based on a predicate that excludes NULLs:

CREATE UNIQUE NONCLUSTERED INDEX idx_yourcolumn_notnull
ON YourTable(yourcolumn)
WHERE yourcolumn IS NOT NULL;

In earlier versions, you can resort to VIEWS with a NOT NULL predicate to enforce the constraint.