Can this ORDER BY on a CASE clause be made faster? - sql

Question

I'm selecting results from a table of ~350 million records, and it's running extremely slowly - around 10 minutes. The culprit seems to be the ORDER BY, as if I remove it the query only takes a moment. Here's the gist:
SELECT TOP 100
(columns snipped)
FROM (
SELECT
CASE WHEN (e2.ID IS NULL) THEN
CAST(0 AS BIT) ELSE CAST(1 AS BIT) END AS RecordExists,
(columns snipped)
FROM dbo.Files AS e1
LEFT OUTER JOIN dbo.Records AS e2 ON e1.FID = e2.FID
) AS p1
ORDER BY p1.RecordExists
Basically, I'm ordering the results by whether Files have a corresponding Record, as those without need to be handled first. I could run two queries with WHERE clauses, but I'd rather do it in a single query if possible.
Is there any way to speed this up?

Answer 1

The ultimate issue is that the use of CASE in the sub-query introduces an ORDER BY over something that is not being used in a sargable manner. Thus the entire intermediate result-set must first be ordered to find the TOP 100 - this is all 350+ million records!2
In this particular case, moving the CASE to the outside SELECT and use a DESC ordering (to put NULL values, which means "0" in the current RecordExists, first) should do the trick1. It's not a generic approach, though .. but the ordering should be much, much faster iff Files.ID is indexed. (If the query is still slow, consult the query plan to find out why ORDER BY is not using an index.)
Another alternative might be to include a persisted computed column for RecordExists (that is also indexed) that can be used as an index in the ORDER BY.
Once again, the idea is that the ORDER BY works over something sargable, which only requires reading sequentially inside the index (up to the desired number of records to match the outside limit) and not ordering 350+ million records on-the-fly :)
SQL Server is then able to push this ordering (and limit) down into the sub-query, instead of waiting for the intermediate result-set of the sub-query to come up. Look at the query plan differences based on what is being ordered.
1 Example:
SELECT TOP 100
-- If needed
CASE WHEN (p1.ID IS NULL) THEN
CAST(0 AS BIT) ELSE CAST(1 AS BIT) END AS RecordExists,
(columns snipped)
FROM (
SELECT
(columns snipped)
FROM dbo.Files AS e1
LEFT OUTER JOIN dbo.Records AS e2 ON e1.FID = e2.FID
) AS p1
-- Hopefully ID is indexed, DESC makes NULLs (!RecordExists) go first
ORDER BY p1.ID DESC
2 Actually, it seems like it could hypothetically just stop after the first 100 0's without a full-sort .. at least under some extreme query planner optimization under a closed function range, but that depends on when the 0's are encountered in the intermediate result set (in the first few thousand or not until the hundreds of millions or never?). I highly doubt SQL Server accounts for this extreme case anyway; that is, don't count on this still non-sargable behavior.

Answer 2

Give this form a try
SELECT TOP(100) *
FROM (
SELECT TOP(100)
0 AS RecordExists
--,(columns snipped)
FROM dbo.Files AS e1
WHERE NOT EXISTS (SELECT * FROM dbo.Records e2 WHERE e1.FID = e2.FID)
ORDER BY SecondaryOrderColumn
) X
UNION ALL
SELECT * FROM (
SELECT TOP(100)
1 AS RecordExists
--,(columns snipped)
FROM dbo.Files AS e1
INNER JOIN dbo.Records AS e2 ON e1.FID = e2.FID
ORDER BY SecondaryOrderColumn
) X
ORDER BY SecondaryOrderColumn
Key indexes:
Records (FID)
Files (FID, SecondaryOrdercolumn)

Answer 3

Well the reason it is much slower is because it is really a very different query without the order by clause.
With the order by clause:
Find all matching records out of the entire 350 million rows. Then sort them.
Without the order by clause:
Find the first 100 matching records. Stop.

Answer 4

Q: If you say the only difference is "with/outout" the "order by", then could you somehow move the "top 100" into the inner select?
EXAMPLE:
SELECT
(columns snipped)
FROM (
SELECT TOP 100
CASE WHEN (e2.ID IS NULL) THEN
CAST(0 AS BIT) ELSE CAST(1 AS BIT) END AS RecordExists,
(columns snipped)
FROM dbo.Files AS e1
LEFT OUTER JOIN dbo.Records AS e2 ON e1.FID = e2.FID
) AS p1
ORDER BY p1.RecordExists

Answer 5

In SQL Server, null values collate lower than any value in the domain. Given these two tables:
create table dbo.foo
(
id int not null identity(1,1) primary key clustered ,
name varchar(32) not null unique nonclustered ,
)
insert dbo.foo ( name ) values ( 'alpha' )
insert dbo.foo ( name ) values ( 'bravo' )
insert dbo.foo ( name ) values ( 'charlie' )
insert dbo.foo ( name ) values ( 'delta' )
insert dbo.foo ( name ) values ( 'echo' )
insert dbo.foo ( name ) values ( 'foxtrot' )
go
create table dbo.bar
(
id int not null identity(1,1) primary key clustered ,
foo_id int null foreign key references dbo.foo(id) ,
name varchar(32) not null unique nonclustered ,
)
go
insert dbo.bar( foo_id , name ) values( 1 , 'golf' )
insert dbo.bar( foo_id , name ) values( 5 , 'hotel' )
insert dbo.bar( foo_id , name ) values( 3 , 'india' )
insert dbo.bar( foo_id , name ) values( 5 , 'juliet' )
insert dbo.bar( foo_id , name ) values( 6 , 'kilo' )
go
The query
select *
from dbo.foo foo
left join dbo.bar bar on bar.foo_id = foo.id
order by bar.foo_id, foo.id
yields the following result set:
id name id foo_id name
-- ------- ---- ------ -------
2 bravo NULL NULL NULL
4 delta NULL NULL NULL
1 alpha 1 1 golf
3 charlie 3 3 india
5 echo 2 5 hotel
5 echo 4 5 juliet
6 foxtrot 5 6 kilo
(7 row(s) affected)
This should allow the query optimizer to use a suitable index (if such exists); however, it does not guarantee than any such index would be used.

Answer 6

Can you try this?
SELECT TOP 100
(columns snipped)
FROM dbo.Files AS e1
LEFT OUTER JOIN dbo.Records AS e2 ON e1.FID = e2.FID
ORDER BY e2.ID ASC
This should give you where e2.ID is null first. Also, make sure Records.ID is indexed. This should give you the ordering you were wanting.