THE SQL Server Blog Spot on the Web

Welcome to SQLblog.com - The SQL Server blog spot on the web Sign in | |
in Search

Tibor Karaszi

How selective do we need to be for an index to be used?

This blog has moved! You can find this content at the following new location:

http://sqlblog.karaszi.com/how-selective-do-we-need-to-be-for-an-index-to-be-used/

Published Thursday, April 1, 2010 10:39 AM by TiborKaraszi

Comment Notification

If you would like to receive an email when updates are made to this post, please register here

Subscribe to this post's comments using RSS

Comments

 

rudy said:

you forgot to 'splain the difference between seek, scan, and just plain have-a-look

April 1, 2010 8:10 AM
 

TiborKaraszi said:

Rudy,

No, I didn't forget that. I assumed that knowledge. All blog posts need to assume a certain level of knowledge, otherwise it wouldn't be possible to blog in the first place. For instamce, I used below for search in my search engine and came out with plenty of hits:

index seek table scan

April 1, 2010 8:39 AM
 

rudy said:

you know what happens when you assume...  ;o)

it's called begging the question, also known as "petitio principii"

if someone is already familiar with b-trees, seeks, and scans, how useful is a blog post on clustered and covering indexes going to be?

April 1, 2010 10:48 AM
 

a.m. said:

Rudy,

No one is forcing you to read Tibor's material. If you feel it's too advanced for you, move along and read something else. Neither Tibor nor any other blogger can be expected to explain every basic point in every single post. Go take a course or something if you want to get up to speed.

April 1, 2010 11:11 PM
 

Klaus said:

Does he also need to explain what % means?

April 2, 2010 11:15 PM
 

AaronBertrand said:

Wow, what a ridiculous complaint about a thorough and insightful post.  From a Canadian, no less; shameful.

April 3, 2010 11:56 AM
 

David Walker said:

I just did some testing on an indexed bit field, with the distribution of 1's in the column at about half a percent.  It's a large table with half a million rows, and 8 GB of data in this table.

I would *not* have expected that an Index Seek is done regardless of whether I search for 0 or 1 in the table where the 1's make up half a percent of the values -- but that's what I saw (SQL 2008 SP1 (not R2)).

At least, the actual execution plan claims that it's doing an index seek on the bit column, regardless of whether I am searching for 1's or 0's.  Is this expected?

December 10, 2010 2:21 PM
 

David Walker said:

Followup: I meant to say 5 percent, not half a percent.  Arithmetic is complicated...

December 10, 2010 5:11 PM
 

TiborKaraszi said:

David, So we have two issues here.

1: The fact the index seek is selected when you return as much as 5%. Did you check execution plan and estimated rows vs actual rows? Also, did you compare number of pages (or some other cost measurement) when you compare to a table scan?

2: The fact that the index is used also when you search for the other value, of which you have 95%. This could either be explained by execution plan re-use or the index cover the query. Anything else would be weird.

Had to say anything more without actually playing with it. Of course, we also have things such as statistics, possible miscalculation etc.

December 11, 2010 5:18 AM
 

David Walker said:

I was looking at the actual execution plan, not the estimated erxecution plan.  From what I could tell, based on running tests with and without the index on the bitmap column, SQL was actually doing an index seek.  Very few actual reads were performed when the bitmap column was indexed.  This just seems different than all of the guidance available on the Web for index selectivity.

December 13, 2010 12:15 PM
 

TiborKaraszi said:

Looking at the actual plan is good, but you don't give us any hard numnbers to go on. Actual number of rows vs. estimated number of rows? Does the index cover the query? Etc. Basically, if you have few page accesses, and return many rows then either the index covers the query, or the index is a clustered index. If not, then you will have at least one page access per row.

December 13, 2010 12:50 PM
 

David Walker said:

I'll post some numbers here.  

December 13, 2010 4:00 PM
 

Kognjen said:

Can someone tell me why border (line) is exactly between 155-156 rows? Can this be calculate as ratio between number of pages, rows per page and number of rows which satisfy search condition...?

Thanks.

January 16, 2012 7:27 AM
 

TiborKaraszi said:

Kognjen,

Consider the execution plan, using the index. SQL Server will navigate the index, and for each row found, it will fetch that row. Each "fetch that row" in this case cost two pages (tables has a clutered index, this has depth 2, for each row navigate the clustered index = 2 pages). 155 rows means 155*2 pages = 310 pages. Plus traversing the index cost, which is about 20 pages. Sum is 330 pages. Remember that the root page for the clustered index will likely be in cache. So this will not be a physical I/O. So should we ignore the root page cost and instead of 330 pages calculate 175 pages? Remember that these are random I/O if physical I/O.

Now, scanning the data means looking at every data page which in this case is 262 pages. These are requential I/O when physical, meaning more efficient than random I/O.

As you probably understand, there is no one exact formula to determine which is more effective, and based on that where the cut-off should be. It has to be *somewhere* and this is where it is in current version. This has changed before and might change again. The principal is the important part here, "visualize the execution plan...". :-)

January 16, 2012 11:36 AM
 

Kognjen said:

To TiborKaraszi - thank you very much.

January 18, 2012 9:30 AM
 

rule30 said:

This was my take on this very subject:

http://rule30.wordpress.com/2014/03/30/sql-server-are-your-non-clustered-indexes-useless/

Thanks for a fantastic blog.

Best Regards

March 30, 2014 4:48 PM

Leave a Comment

(required) 
(required) 
Submit
Privacy Statement