One index or two?

Question

I have the following index created on a table in my database:

CREATE INDEX [idx_index1]
on [table1]
(col1, col2, col3)

The server is suggesting the following 'missing' index:

CREATE INDEX [idx_index2]
on [table1]
(col1, col2)
INCLUDE (col3, col4, col5, col6....)

It seems logical to me to amend the existing index definition to include the suggested columns, rather than creating a new index that needs to be maintained. A query that selects on col1 and col2 could use index1 just as effectively as index2. Am I correct or am I maybe missing something?

Answer 1

And so enters the art of performance tuning and indexing strategies...

It seems logical to me to amend the existing index definition to include the suggested columns

I'm going to take your quote and write a third index definition:

create index [idx_index3]
on [table1] (col1, col2, col3)
include (col4, col5, col6....);

That should be the CREATE INDEX statement that corresponds to your quoted statement.

That very well may be a prudent solution, but it depends. Here are a couple of examples when I say that it depends.

If you have a common workload that mostly consists of queries like this:

select col1, col2, col3
from table1
where col1 = 1
and col2 = 2
and col3 = 3;

Then your idx_index1 index would be solid. Perfectly narrow, it's an index that satisfies that query with no extraneous data in it (not taking into account the clustered index definition, if one at all).

But if you have workload that consists of queries mainly like the following:

select co11, col2, col3, col4, col5
from table1
where col1 = 1
and col2 = 2;

Then idx_index2 would be wise, as it is what's called a covering index preventing the need for a key lookup back to the clustered index (or a RID lookup back to the heap). That nonclustered index definition would solely encompass all of the data that query needs.

With your recommendation, it would be well suited for a query like the following:

select co11, col2, col3, col4, col5
from table1
where col1 = 1
and col2 = 2
and col3 = 3;

Your idx_index3 recommendation would be a covering index that satisfies the search criteria for the above query.

The point I'm trying to get at, is in an isolated question like this we can't answer this definitively. It all depends on what the common and frequent workload is. Of course you could always define all three of these indexes to handle each sample query type, but then comes into question the maintenance that'll be required to keep these indexes updated (think: INSERTs, UPDATEs, DELETEs). That's the overhead of indexes.

You need to dissect and evaluate the workload, and determine where the advantages will be best in place. If the first sample query is the most common by far being executed dozens of times a second, and there is a very infrequent query like the third sample query, then it wouldn't make sense to bloat the leaf level pages of the index with the INCLUDE nonkey columns. It all depends on your workload.

If you understand prudent indexing strategies, and you understand your common workload, then by applying both of those you will be able to come up with what is the best route to take.

Answer 2

You are indeed correct and have discovered why it's important for a DBA to always review the "suggestions" put forward by the missing index DMVs etc.

Consider that the suggestions offered by the missing index DMVs are put forward in isolation, meaning that SQL Server decided that an index of the recommended structure would benefit the query, irrespective of what other index structures may already exist.

Answer 3

A little more, on one of the implications of Thomas' answer:

He said:

Of course you could always define all three of these indexes to handle each sample query type, but then comes into question the maintenance that'll be required to keep these indexes updated (think: INSERTs, UPDATEs, DELETEs). That's the overhead of indexes.

So, another big question becomes: how often is the table updated?

Consider first an example of a table that is constantly updated, like for example, a retail ORDERS table reflecting website consumer activity... there, you want to be conscientious about having multiple indexes, because they increase the work done by constant updates, and therefore constantly affect the database's performance.

On the other hand, consider a table that is only updated as part of website setup -- the table being updated ONCE for most values, and values infrequently added -- there, update slowdowns are pretty much not a consideration. Multiple indexes could slowdown database index rebuilds & reorgs, but as long as they are fast enough, FEEL FREE: if multiple indices speeds up the reads, go for it.

A middle case could be a table that is normally only updated in a batch process overnight. There, update slowdowns from multiple indexes would not affect daytime performance -- they would only affect (1) the time taken, to run that nightly batch maintenance, (2) the performance of any concurrent processes, and (3) the time taken for database maintenance tasks like index reorganization. So, as long as processes in those 3 arenas are running fast enough for you... create the indexes that speed up queries.

HTH...