You know the answer already: It depends. But I often see some percentage value quoted and the point of this post is to show that there is no such percentage value.
To get the most out of this blog post, you should understand the basic structure for an index, i.e. how the b+ tree look like. You should also understand the difference between a clustered and a non-clustered index. In essence, you should be able to visualize these structures and searches through them as you read the text. If you find that difficult, draw a few versions on a piece of paper and “navigate” through them by tracing through them with a pen or your finger. After a while, you will do this in your mind. For instance, check out the sections under this.
I’m not saying that we shouldn’t consider selectivity when designing indexes – of course we should! I’m not saying that one shouldn’t have some vague “feeling” about how much data to be return when making such decisions. What I will prove is that there is in reality no set percentage that the optimizer uses. The comment we usually see is something like:
“If we return more than 5% of the rows, then an index will not be used.”
Where did that 5% number came from? I can assure you that this is not some hard-wired number in the optimizer (except for an edge-case, see below). The optimizer aims at running the query with as low cost as possible. Considering the data access part (think WHERE clause and the condition), this is pretty much about reading as few pages as possible (few page-accesses).
Just to cut down a bit on the thread that might follow these types of blogs (“Hey, when I do this, your observations doesn’t match, your blog post is incorrect!”), let us first consider some special cases:
The clustered index *is* the data. If the search condition (SARG) is SEEKable, then SQL Server will obviously seek through a clustered index instead of scan it. Anything else would be stupid. There can be *other* search conditions that are more efficient, but we are considering one search condition at-a-time.
Non-clustered index that covers the query
This is pretty much the same argument as for above. Since all data is in the index (“covers the query”), not seeking it would be stupid. Again, there can be cheaper alternatives for any of the other search conditions, but we are considering one condition at-a-time.
The value is not known to the optimizer
This is what happens when you have a TSQL variable you compare against. Something like “colname = @v”. The optimizer has no knowledge of the contents of this variable. Either it uses density (where applicable, like “=”), as stored in the statistics information of the index. Where not applicable (like “>”, “<“, “BETWEEN” etc), then the optimizer actually do use some hard-wired percentage value. This value can change between versions so give it a spin of you want to know what value you have for your version/build number. Note that a variable is not the same thing as a parameter. SQL Server sniffs parameters (parameter sniffing). Read this for elaboration: http://msdn.microsoft.com/en-us/library/ee343986.aspx.
The search expression is not seek-able
I hope you know this already, but just to point it out. In most cases, having some calculation at the column side will void the ability to seek through the index. This should ideally be known to all T-SQL developers: Never do calculations at the column side! So, things to avoid are like “colname * 23 > 45” or “SOMEFUNCTION(colname) = 44”.
Hopefully by now we all understand that there are always special cases and exceptions. The more of Kalen’s books you have read, the more you understand this. What we are discussing here is the typical situation. OK? Fine. So, “Why is there no percentage value that the optimizer uses?”, you ask. Because the value will differ. In short, SQL Server wants to read as few pages as possible. In the most simple example, the alternative to an index seek is a table scan. So we will use this as basis for your discussion. There can be other alternatives to the table scan (using some other index for some other condition), but that doesn’t change the principal “it depends” concept.
In essence, it is all about the alternative. As I said, our example wil use a table scan as alternative. A table scan (or clustered index scan if it is a clustered table) means that SQL Server will look at every page and see what rows satisfies the search condition on each page.
My example has two different tables, both with 100,000 rows. These tables both have an integer column with consecutive increasing unique values, which also has a non-clustered index. I will see how selective I need to be when searching on this column in order for an index search to be done, compared to a table scan. I.e, find this percentage cut-off value.
The fewrows table only fit one row per data page. This means 100,000 data pages. My tests show that the cutoff for fewrows is about 31,000 rows. I.e., we need to be more selective than 31% for the index to be used.
The manyrows table fit 384 rows per page. This means 260 pages. My tests show that the cutoff for fewrows is about 155 rows. I.e., we need to be more selective than 0.16% for the index to be used.
You might end up with different exact numbers, depending on what you have in the statistics, the build number of your SQL Server etc. But what you will see that a similar pattern. A huge difference between the two.
It is all about the alternative
If I look at my indexes using sys.dm_db_index_physical_stats, I will see that the non-clustered index on the int column for the two tables are exactly the same (same number of pages in the index, etc). So, two indexes with the same characteristics have very different cut-off values. How can that be? It is because the alternative differs. The alternative for this example is a table scan. For the bigrows table, the alternative means reading 100,000 pages. But for the smallrows table, the alternative means reading only 260 pages. There can of course be other alternatives, like using some other index for some other search condition. This is, in the end, why we don’t have a set percentage value: it is all about the alternative!
The typical cases will of course fall somewhere between my more extreme examples. But my examples show that there is no set percentage value used by the optimizer. I showed that for my test, the percentage value can be as low as 0.15% or as high as 31%. What matter is the alternative!
IF OBJECT_ID('manyrows') IS NOT NULL DROP TABLE manyrows
IF OBJECT_ID('fewrows') IS NOT NULL DROP TABLE fewrows
CREATE TABLE manyrows(c1 INT IDENTITY PRIMARY KEY, c2 INT, c3 INT)
CREATE TABLE fewrows(c1 INT IDENTITY PRIMARY KEY, c2 INT, c3 CHAR(4500))
INSERT INTO manyrows
SELECT TOP 100000 ROW_NUMBER() OVER (ORDER BY a.OBJECT_ID) AS c2, 5 AS c3
FROM sys.columns AS a, sys.columns AS b
INSERT INTO fewrows
SELECT TOP 100000 ROW_NUMBER() OVER (ORDER BY a.OBJECT_ID) AS c2, 'hi' AS c3
FROM sys.columns AS a, sys.columns AS b
CREATE INDEX x ON manyrows (c2)
CREATE INDEX x ON fewrows (c2)
--Number of pages etc:
EXEC sp_indexinfo 'manyrows'
-- Data: 265 pages (2.07 MB)
-- Index x: 187 pages (1.46 MB)
EXEC sp_indexinfo 'fewrows'
-- Data: 100385 pages (784 MB)
-- Index x: 187 pages (1.46 MB)
SELECT OBJECT_NAME(OBJECT_ID), *, OBJECT_NAME(OBJECT_ID)
FROM sys.dm_db_index_physical_stats(DB_ID(), NULL, NULL, NULL, 'DETAILED')
--Run below with showplan:
SELECT * FROM manyrows
WHERE c2 BETWEEN 1 AND 155
--155 = ix search, 156 = ts
--Cut-off is 0.16%
SELECT * FROM fewrows
WHERE c2 BETWEEN 1 AND 31000
--31000 = ix search, 32000 = ts
--Cut-off is 31%