A simple self join would seem to perform much better than a row referencing subquery
Generate 10k rows of test data:
drop table test10k
create table test10k (Id int, Number int, constraint test10k_cpk primary key clustered (id))
;WITH digits AS (
SELECT 0 as Number
UNION SELECT 1
UNION SELECT 2
UNION SELECT 3
UNION SELECT 4
UNION SELECT 5
UNION SELECT 6
UNION SELECT 7
UNION SELECT 8
UNION SELECT 9
)
,numbers as (
SELECT
(thousands.Number * 1000)
+ (hundreds.Number * 100)
+ (tens.Number * 10)
+ ones.Number AS Number
FROM digits AS ones
CROSS JOIN digits AS tens
CROSS JOIN digits AS hundreds
CROSS JOIN digits AS thousands
)
insert test10k (Id, Number)
select Number, Number
from numbers
I would pull the special case of the first 3 rows out of the main query, you can UNION ALL those back in if you really want it in the row set. Self join query:
;WITH NumberedRows
AS
(
SELECT rta.*, row_number() OVER (ORDER BY rta.ID ASC) AS RowNumber
FROM test10k rta
)
SELECT nr.ID, nr.Number,
avg(trailing.Number) as MovingAverage
FROM NumberedRows nr
join NumberedRows as trailing on trailing.RowNumber between nr.RowNumber-3 and nr.RowNumber-1
where nr.Number > 3
group by nr.id, nr.Number
On my machine this takes about 10 seconds, the subquery approach that Aaron Alton demonstrated takes about 45 seconds (after I changed it to reflect my test source table) :
;WITH NumberedRows
AS
(
SELECT rta.*, row_number() OVER (ORDER BY rta.ID ASC) AS RowNumber
FROM test10k rta
)
SELECT nr.ID, nr.Number,
CASE
WHEN nr.RowNumber <=3 THEN NULL
ELSE ( SELECT avg(Number)
FROM NumberedRows
WHERE RowNumber < nr.RowNumber
AND RowNumber >= nr.RowNumber - 3
)
END AS MovingAverage
FROM NumberedRows nr
If you do a SET STATISTICS PROFILE ON, you can see the self join has 10k executes on the table spool. The subquery has 10k executes on the filter, aggregate, and other steps.