How can I remove duplicate rows?

Assuming no nulls, you GROUP BY the unique columns, and SELECT the MIN (or MAX) RowId as the row to keep. Then, just delete everything that didn't have a row id:

DELETE FROM MyTable
LEFT OUTER JOIN (
   SELECT MIN(RowId) as RowId, Col1, Col2, Col3 
   FROM MyTable 
   GROUP BY Col1, Col2, Col3
) as KeepRows ON
   MyTable.RowId = KeepRows.RowId
WHERE
   KeepRows.RowId IS NULL

In case you have a GUID instead of an integer, you can replace

MIN(RowId)

with

CONVERT(uniqueidentifier, MIN(CONVERT(char(36), MyGuidColumn)))

Another possible way of doing this is

; 

--Ensure that any immediately preceding statement is terminated with a semicolon above
WITH cte
     AS (SELECT ROW_NUMBER() OVER (PARTITION BY Col1, Col2, Col3 
                                       ORDER BY ( SELECT 0)) RN
         FROM   #MyTable)
DELETE FROM cte
WHERE  RN > 1;

I am using ORDER BY (SELECT 0) above as it is arbitrary which row to preserve in the event of a tie.

To preserve the latest one in RowID order for example you could use ORDER BY RowID DESC

Execution Plans

The execution plan for this is often simpler and more efficient than that in the accepted answer as it does not require the self join.

This is not always the case however. One place where the GROUP BY solution might be preferred is situations where a hash aggregate would be chosen in preference to a stream aggregate.

The ROW_NUMBER solution will always give pretty much the same plan whereas the GROUP BY strategy is more flexible.

Factors which might favour the hash aggregate approach would be

• No useful index on the partitioning columns
• relatively fewer groups with relatively more duplicates in each group

In extreme versions of this second case (if there are very few groups with many duplicates in each) one could also consider simply inserting the rows to keep into a new table then TRUNCATE-ing the original and copying them back to minimise logging compared to deleting a very high proportion of the rows.

There's a good article on removing duplicates on the Microsoft Support site. It's pretty conservative - they have you do everything in separate steps - but it should work well against large tables.

I've used self-joins to do this in the past, although it could probably be prettied up with a HAVING clause:

DELETE dupes
FROM MyTable dupes, MyTable fullTable
WHERE dupes.dupField = fullTable.dupField 
AND dupes.secondDupField = fullTable.secondDupField 
AND dupes.uniqueField > fullTable.uniqueField