5

I have a table (with 170M rows) that looks as follows:

CREATE TABLE [dbo].[Panel]
(
    [SubId] [varchar](15) NOT NULL,
    [LineageId] [int] NULL,
    [Buck] [varchar](20) NULL,
    [Lot] [varchar](20) NULL,
    [GlassType] [varchar](20) NULL,
    [ETA] [varchar](200) NULL,
 CONSTRAINT [PK_Panel] PRIMARY KEY CLUSTERED 
(
    [SubId] ASC
)

99% of the queries against this table reference SubId either in the Where clause or join. One of our DBAs told me he could make all those queries and joins perform better by creating the following index:

CREATE UNIQUE NONCLUSTERED INDEX [IX-Panel-SubID-I-LineageID] ON [dbo].[Panel]
(
    [SubId] ASC
)
INCLUDE ([LineageId])

When he told me this, I thought he was nuts. But I just checked index usage since this index was created and found the following:

PK_Panel (232,394 seeks / 2,133 scans)
IX-Panel-SubID-I-LineageID (25,528 seeks / 3644 scans)

I was a bit shocked to see this. Under what circumstances would this new index every get used? Why would SQL Server ever select it?

Or maybe a better question would be, why would SQL Server select the new index to do a seek, instead of the clustered index? Approximately 25K times, it thought seeking on the new index was a better choice.

In case this helps, the LineageId essentially indicates where a panel was created and there are ~35 distinct values it could contain.

  • When SubId is the "clustering" key, it means that the index "contains" entire table, so index scan =~= table scan. It seems that many of your queries need only the SubId and sometimes LineageId fields, not entire table, but they need some sort of scan (it is probably about the joins? just guessing here without the actual queries to check) and scanning nonclustered index containing two columns is much faster (as it means reading less data) – jkavalik Oct 6 '15 at 19:18
  • The query optimizer will decide to use that index if it thinks it is more efficient than accessing the table via the clustered index (which is the table). – Max Vernon Oct 6 '15 at 19:18
  • 1
    Out of curiosity, what is the LineageId column used for? – Max Vernon Oct 6 '15 at 19:19
  • @MaxVernon - The LineageId essentially indicates where a panel was created and there are ~35 distinct values it could contain. – Randy Minder Oct 6 '15 at 19:21
  • For the edit - probably only the primary key (and optionaly the LineageId) are touched by some queries (maybe join is used to only check the existence of a record, not selecting any data from that table). In that case again the smaller index may mean fewer pages to read to do the seek. – jkavalik Oct 6 '15 at 19:26
6

Creating an index on the column involved in the clustering key may seem a little strange. One asks why create another index, when one already exists?

The clustered index is the table. That is to say your clustered index for this table is similar to an index like:

CREATE INDEX IX_Panel
ON dbo.Panel(SubId ASC)
INCLUDE (
    LineageId
    , Buck
    , Lot
    , GlassType
    , ETA
);

Pretty clearly, this is not the same as an non-clustered index on the clustering key, which in your case only "includes" the LineageId column:

CREATE UNIQUE NONCLUSTERED INDEX [IX-Panel-SubID-I-LineageID] ON [dbo].[Panel]
(
    [SubId] ASC
)
INCLUDE ([LineageId]);

As a test, I created a mock-up of your table, then insert over 700,000 rows into it:

INSERT INTO dbo.Panel(LineageId, Buck, Lot, GlassType, ETA)
SELECT (ROW_NUMBER() OVER (ORDER BY o1.object_id, o2.object_id) % 35)
    , SUBSTRING(o2.name, 1, 15)
    , 'lot'
    , 'GlassType'
    , o3.name + o2.name
FROM sys.objects o1
    , sys.objects o2
    , sys.objects o3;

I then ran the following query to see stats about both indexes:

SELECT o.name
    , ps.index_id
    , ps.index_type_desc
    , ps.page_count
FROM sys.dm_db_index_physical_stats(DB_ID(),OBJECT_ID('dbo.Panel'),-1,0, NULL) ps
INNER JOIN sys.objects o ON ps.object_id = o.object_id;

The results of the query above are:

enter image description here

Clearly, the non-clustered index is smaller than the clustered index. In this case about 1/5th the size. Being a smaller index means the query optimizer will choose to use it when the index fulfills the requirements of the query, in one way or another.

For instance,

SELECT SubId
FROM dbo.Panel;

will only need to read 1,304 pages from disk by scanning the non-clustered index, instead of having to read 7,105 pages to scan the clustered-index.

enter image description here

On the other hand, for a query that references columns that are not included in the non-clustered index, SQL Server will likely use the clustered index since it fulfills the requirements of the query.

Take for instance, the following query, where you might intuitively think the query optimizer might pick the non-clustered index since it can use that to fulfill the WHERE clause, however the simple act of having to look up the Buck column means it is faster to simply seek the clustered index for the values in the WHERE, then return the 3 columns in the SELECT clause:

SELECT SubId
    , LineageId
    , Buck
FROM dbo.Panel
WHERE LineageId = 6
    AND SubId >= 27
    AND SubId <= 42;

enter image description here

This is a bit of a simplification, I would highly recommend looking at Brent Ozar's Index pages

5

Choosing a narrower index is generally better, especially if it covers, meaning those queries that use seeks probably only need SubID and possibly LineageID. Though even if the index doesn't cover, it may still choose the narrower index and perform lookups for the additional column(s) if the row count is small enough that this is still less expensive overall.

Unless the queries actually require all or most of the columns in the table, the clustered index would be a poor choice, because it is spread across a lot more pages, which means more I/O to get at the subset of columns you need. If these can all be obtained from the narrower index, it will be the cheaper way to access the data.

Cost is largely determined by the number of pages that need to be read in order to satisfy the query. The clustered index is the whole table, so it is never going to be skinnier than any non-clustered index. This means that, generally, for the same number of rows, the clustered index will require more I/O than the non-clustered index. The scales tip when the non-clustered index doesn't satisfy the query, which means SQL Server needs to choose between going to the clustered index and going to the non-clustered index and performing lookups. The latter is judged based roughly on number of estimated rows and column widths, so there really is a lot of "it depends" that will come into play.

Sadly, the index usage stats DMV doesn't distinguish between a singleton seek and a seek disguising a range scan (be it 2 rows or 2 million rows), so just seeing a number of seeks vs. scans doesn't really tell us the whole story. And we don't have enough information about your actual queries to show you exactly why this choice was made, but I hope I've indicated the likely reason at least at a high level.

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.