Category: Execution Plans

Unavoidable

Having some key lookups in your query plans is generally unavoidable.

You’ll wanna select more columns than you wanna put in a nonclustered index, or ones with large data types that you don’t wanna bloat them with.

Enter the key lookup.

They’re one of those things — I’d say even the most common thing — that makes parameterized code sensitive to the bad kind of parameter sniffing, so they get a lot of attention.

The thing is, most of the attention that they get is just for columns you’re selecting, and most of the advice you get is to “create covering indexes”.

That’s not always possible, and that’s why I did this session a while back on a different way to rewrite queries to sometimes make them more efficient. Especially since key lookups may cause blocking issues.

Milk and Cookies

At some point, everyone will come across a key lookup in a query plan, and they’ll wonder if tuning it will fix performance.

There are three things to pay attention to when you look at a key lookup:

1. Number of executions: This is usually more helpful in an actual plan
2. If there are any Predicates involved: That means there are parts of your where clause not in your nonclustered index
3. If there’s an Output List involved: That means you’re selecting columns not in your nonclustered index

For number of executions, generally higher numbers are worse. This can be misleading if you’re looking at a cached plan because… You’re going to see the cached number, not the runtime number. They can be way different.

Notice I’m not worried about the Seek Predicates here — that just tells us how the clustered index got joined to the nonclustered index. In other words, it’s the clustered index key column(s).

Figure It Out

Here’s our situation: we’re working on a new stored procedure.

CREATE PROCEDURE dbo.predicate_felon (@Score INT, @CreationDate DATETIME)
AS
BEGIN

    SELECT *
    FROM dbo.Comments AS c
    WHERE c.Score = @Score
    AND   c.CreationDate >= @CreationDate
    ORDER BY c.CreationDate DESC;

END;

Right now, aside from the clustered index, we only have this nonclustered index. It’s great for some other query, or something.

CREATE INDEX ix_whatever 
ON dbo.Comments (Score, UserId, PostId)
GO

When we run the stored procedure like this, it’s fast.

EXEC dbo.predicate_felon @Score = 6, --Sixer
                         @CreationDate = '2013-12-31';

SQL Server wants an index — a fully covering index — but if we create it, we end up a 7.8GB index that has every column in the Comments table in it. That includes the Text column, which is an NVARCHAR(700). Sure, it fixes the key lookup, but golly and gosh, that’s a crappy index to have hanging around.

Bad Problems On The Rise

The issue turns up when we run the procedure like this:

EXEC dbo.predicate_felon @Score = 0, --El Zero
                         @CreationDate = '2013-12-31';

This happens because there are a lot more 0 scores than 6 scores.

Smarty Pants

Eagle eyed readers will notice that the second query only returns ~18k rows, but it takes ~18 seconds to do it.

The problem is how much time we spend locating those rows. Sure, we can Seek into the nonclustered index to find all the 0s, but there are 20.5 million of them.

Looking at the actual plan, we can spot a few things.

The 18k rows we end up with are only filtered to with they key lookup, but it has to execute 20.5 million times to evaluate that extra predicate.

If we just index the key columns, the key lookup to get the other columns (PostId, Text, UserId) will only execute ~18k times. That’s not a big deal at all.

CREATE NONCLUSTERED INDEX keys_only
    ON dbo.Comments ( Score, CreationDate );

This index is only ~500MB, which is a heck of a lot better than nearly 8GB covering the entire thing.

With that in place, both the score 6 and score 0 plans are fast.

Why This Is Effective, and When It Might Not Be

This works here because the date filter is restrictive.

When we can eliminate more rows via the index seek, the key lookup is less of a big deal.

If the date predicate were much less restrictive, say going back to 2011, boy oh boy, things get ugly for the 0 query again.

EXEC dbo.predicate_felon @Score = 6,
                         @CreationDate = '2011-12-31';

EXEC dbo.predicate_felon @Score = 0,
                         @CreationDate = '2011-12-31';

Of course, returning that many rows will suck no matter what, so this is where other techniques come in like Paging, or charging users by the row come into play.

What? Why are you looking at me like that?

Thanks for reading!

Going Further

If this is the kind of SQL Server stuff you love learning about, you’ll love my training. I’m offering a 75% discount to my blog readers if you click from here. I’m also available for consulting if you just don’t have time for that and need to solve performance problems quickly.

Lost In The Woods

When you find unused indexes, whether using Some Script From The Internet™, sp_BlitzIndex, or Database Telepathy, the first thing most people think of is “wasted space”.

Sure, okay, yeah. That’s valid. They’re in backups, restores, they get hit by CHECKDB. You probably rebuild them if there’s a whisper of fragmentation.

But it’s not the end of the story.

Not by a long shot.

Today we’re going to look at how redundant indexes can clog the buffer pool up.

Holla Back

If you want to see the definitions for the views I’m using, head to this post and scroll down.

Heck, stick around and watch the video too.

LIKE AND SUBSCRIBE.

Now, sp_BlitzIndex has two warnings to catch these “bad” indexes:

• Unused Indexes With High Writes
• NC Indexes With High Write:Read Ratio

Unused are just what they sound like: they’re not helping queries read data at all. Of course, if you’ve rebooted recently, or rebuilt indexes on buggy versions of SQL Server, you might get this warning on indexes that will get used. I can’t fix that, but I can tell you it’s your job to keep an eye on usage over time.

Indexes with a high write to read ratio are also pretty self-explanatory. They’re sometimes used, but they’re written to a whole lot more. Again, you should keep an eye on this over time, and try to understand both how important they might be to your workload, or how much they might be hurting your workload.

I’m not going to set up a fake workload to generate those warnings, but I am going to create some overlapping indexes that might be good candidates for you to de-clutter.

Index Entrance

The Votes table is pretty narrow, but it’s also pretty big — 53 million rows or so as of Stack 2013.

Here are my indexes:

CREATE INDEX who ON dbo.Votes(PostId, UserId) INCLUDE(BountyAmount); 
CREATE INDEX what ON dbo.Votes(UserId, PostId) INCLUDE(BountyAmount); 
CREATE INDEX [where] ON dbo.Votes(CreationDate, UserId) INCLUDE(BountyAmount); 
CREATE INDEX [when] ON dbo.Votes(BountyAmount, UserId) INCLUDE(CreationDate); 
CREATE INDEX why ON dbo.Votes(PostId, CreationDate) INCLUDE(BountyAmount); 
CREATE INDEX how ON dbo.Votes(VoteTypeId, BountyAmount) INCLUDE(UserId);

First, I’m gonna make sure there’s nothing in memory:

CHECKPOINT;
GO 2
DBCC DROPCLEANBUFFERS;
GO 

Don’t run that in production. It’s stupid if you run that in production.

Now when I go to look at what’s in memory, nothing will be there:

SELECT *
FROM dbo.WhatsUpMemory AS wum
WHERE wum.object_name = 'Votes'

I’m probably not going to show you the results of an empty query set. It’s not too illustrative.

I am going to show you the index sizes on disk:

SELECT *
FROM dbo.WhatsUpIndexes AS wui
WHERE wui.table_name = 'Votes';

And I am going to show you this update:

UPDATE v
SET v.BountyAmount = 2147483647
FROM dbo.Votes AS v
WHERE v.BountyAmount IS NULL
AND   v.CreationDate >= '20131231'
AND v.VoteTypeId > 2;

After The Update

This is when things get more interesting for the memory query.

We’re updating the column BountyAmount, which is present in all of the indexes I created. This is almost certainly an anti-pattern, but it’s good to illustrate the problem.

Pieces of every index end up in memory. That’s because all data needs to end up in memory before SQL Server will work with it.

It doesn’t need the entirety of any of these indexes in memory — we’re lucky enough to have indexes to help us find the 10k or so rows we’re updating. I’m also lucky enough to have 64GB of memory dedicated to this instance, which can easily hold the full database.

But still, if you’re not lucky enough to be able to fit your whole database in memory, wasting space in the buffer pool for unused (AND OH GODD PROBABLY FRAGMENTED) indexes just to write to them is a pretty bad idea.

After all, it’s not just the buffer pool that needs memory.

You also need memory for memory grants (shocking huh?), and other caches and activities (like the plan cache, and compressed backups).

Cleaning up those low-utilization indexes can help you make better use of the memory that you have.

Thanks for reading!

Going Further

If this is the kind of SQL Server stuff you love learning about, you’ll love my training. I’m offering a 75% discount to my blog readers if you click from here. I’m also available for consulting if you just don’t have time for that and need to solve performance problems quickly.