Category: DMVs


Clearing the Plan Cache

While at SQL Saturday #63 in Dallas, I got some questions about clearing the plan cache.  This evolved into an email discussion; which resulted in me writing up some example scripts.  Since I received some good feedback, I thought this would be a good opportunity to turn them into a blog post.

Warning: These scripts should almost never be run on a production system.  These are meant for debugging and testing on DEV.  If you do run these on PROD, be sure to understand the consequences.

I could see running these on PROD in a few limited cases.  For example, if you have a query that is stuck with a bad plan in cache, you can use these scripts to remove that individual plan.  Another possibility would be to clear out the adhoc plans, perhaps just after you have enabled the optimize for adhoc workloads setting.

Examining the Plan Cache

The Plan Cache is where SQL Server stores the Execution Plans for any queries or stored procedures it has run.  The purpose of this is to enable SQL Server to reuse Execution Plans for each subsequent run of the query or stored procedure.  This allows SQL Server to scale up as more requests begin to pile on.

To examine the Plan Cache you can start by querying the sys.dm_exec_cached_plans DMV.


-- look at the plan cache, list all plans
select * from sys.dm_exec_cached_plans;
go

This is not very useful by itself.  In order to make this query a bit more useful, we can link in either the Query Text or the Query Plan by using two other DMVs: sys.dm_exec_sql_text and sys.dm_exec_query_plan respectively.


-- link in the sql text
select decp.*, dest.text
from sys.dm_exec_cached_plans decp
cross apply sys.dm_exec_sql_text(decp.plan_handle) dest;
go

-- link in the query plan
select decp.*, deqp.query_plan
from sys.dm_exec_cached_plans decp
cross apply sys.dm_exec_query_plan(decp.plan_handle) deqp;
go

And if you wish to see both the Query Text and the Query Plan together, you can link multiple DMVs in the same query.


-- link in both
select decp.*, dest.text, deqp.query_plan
from sys.dm_exec_cached_plans decp
cross apply sys.dm_exec_sql_text(decp.plan_handle) dest
cross apply sys.dm_exec_query_plan(decp.plan_handle) deqp;
go

Clearing the Plan Cache

As I work through clearing the Plan Cache, I’m going to start with the broadest example, and then get more granular as we go along.  If you need to clear the entire Plan Cache, you can do so with a DBCC FREEPROCCACHE command.

I feel the need to issue another warning here, and reiterate that this is almost always a bad idea on a production system.


-- clears everything in the plan cache
dbcc freeproccache;
go

Clearing by Database

If clearing the entire Plan Cache seems a bit excessive, as well it should, then you can get a wee bit more focused by only clearing the plans for a particular database.  This can be achieved by using the DBCC FLUSHPROCINDB command.  This is an undocumented command in SQL 2008, however, you can read about it under the SQL 2005 documentation.

As with clearing the entire cache, this is probably not something you want to do with a production system unless you have a good reason.  You may have one database that is a special case, which needs to be tamed.


-- clear the plan cache for a particular database
declare @v_DBID tinyint;

select @v_DBID = dbid from sys.sysdatabases where name = 'AdventureWorks2008R2';

dbcc flushprocindb (@v_DBID);
go

Clearing by Object Type

Next, let’s look at clearing the Plan Cache for a particular type of object.  By using DBCC FREESYSTEMCACHE we can choose to clear out the Adhoc Plans, Stored Procedure Plans, Views, or Extended Stored Procedures.

Most likely, the ones you will be interested in are Adhoc Plans and Stored Procedures.


-- clear plans for a particular type of object

-- clear sql plans: adhoc sql, prepared statements
dbcc freesystemcache('SQL Plans');
go

-- clear object plans: stored procedures, functions, triggers
dbcc freesystemcache('Object Plans');
go

-- clear bound trees: views, constraints, defaults
dbcc freesystemcache('Bound Trees');
go

-- clear extended stored procedures
dbcc freesystemcache('Extended Stored Procedures');
go

Like a broken record, I gotta remind you that it is a bad idea to do this on a production system.

Optimize for Adhoc Workloads

The primary reason you may wish to clear out the Adhoc Plans from your cache, is if you have just enabled the SQL Server setting ‘optimize for adhoc workloads‘.  This setting tells SQL Server not to cache the Execution Plans for Adhoc queries until they have been run more than once.  This helps keep Plan Cache Bloat under control.

You can enable this setting by using sp_configure.


-- show advanced options - enable all options to be seen / set
use [master]
go
exec sp_configure 'show advanced options', 1
go
reconfigure
go
reconfigure with override
go

-- optimize for ad hoc workloads
sp_configure 'optimize for ad hoc workloads', 1
go
reconfigure
go
reconfigure with override
go

-- verify change
sp_configure 'optimize for ad hoc workloads'
go

Once you have enabled this setting, the Single-Use Adhoc Plans will remain in cache until either the SQL Service has been restarted, or you have explicitly cleared the Plan Cache of Adhoc Plans by using the code example above.

Resource Governor

If you are using Resource Governor, then you have the option of clearing the Plan Cache for a particular Resource Pool.  Resource Governor allows you to create various Resource Pools and then designate a certain percentage of CPU or Memory to that pool.

In order to get a list of the Resource Pools that are defined on your system, you can query the sys.dm_resource_governor_resource_pools DMV.


-- list the resource governor pools
select distinct name as 'Resource Governor Pools'
from sys.dm_resource_governor_resource_pools;
go

Once you have identified the Resource Pool that needs to have its associated Cached Plans cleared, you can clear the plans by using the DBCC FREEPROCCACHE command.


-- clear the default resource governor pool
dbcc freeproccache ('default');
go

-- clear the internal resource governor pool
dbcc freeproccache ('internal');
go

-- clear the default resource governor pool
dbcc freeproccache ('YOURRESOURCEGOVERNORPOOL');
go

Again, please do not do this on a production system unless you are certain this is what you need.

Clearing a Single Plan

If you have a particular query which is causing you problems, for example, it got compiled with a bad plan, then, one option is to remove the offending plan from the cache, which will cause a recompile.  There is no guarantee, however, that the new plan will be any better than the old plan.

You may wish to capture the existing plan so that you can compare it with the new plan.  This way you can explore what the differences are, and focus your tuning efforts there.

To find the query plan for a particular query we will need a search string; the SQL from your query.  We will place your query inside of mine, so we can search the Plan Cache for it.  It is not necessary to use the entire text of the query, that would be a bit cumbersome. All we need is enough to make it unique, or unique enough that we can find it in the results.


-- or it may be even better to search by the sql text
-- so that you only see the ones that interest you
-- also link in the query plan so that you can analyze it
select decp.usecounts, decp.size_in_bytes, dest.text, decp.plan_handle, deqp.query_plan
from sys.dm_exec_cached_plans decp
cross apply sys.dm_exec_sql_text(decp.plan_handle) dest
cross apply sys.dm_exec_query_plan(decp.plan_handle) deqp
where dest.text like '%PLACE YOUR SQL CODE SNIPPET HERE%';
go

Once you have identified the offending plan, it is a simple matter to remove it from the cache.  Use the plan_handle with the DBCC FREEPROCCACHE command to clear it from the Plan Cache.


-- once you find the plan you wish to prune
-- place the plan_handle in the following statement
dbcc freeproccache (0x0600010010450926B8E00C07000000000000000000000000);
go

So this is the only one of the options where I can see a valid case for doing it on Production.  All the same, be careful and understand what you are doing.  It is important to clear only the Query Plan that is needed, and no more.

Conclusion

It is important to keep tabs on what is going on in your Plan Cache.  Being able to clear out all or part of your Plan Cache can be a powerful debugging technique.  Please bear in mind that this is not something that should be done on a Production server without careful consideration of the side effects.

However, playing with this in DEV can help you understand more of what is happening under the covers of SQL Server.

Back to the Drawing Board

Like many DBAs, I have a variety of scripts I use to check on the health of my servers.  Some are borrowed, others stolen, and a few I have written myself.  A common theme amongst them is the need to filter either user or system processes depending on the situation.

Filtering User Processes

Typically I have done this by adding a where clause which filters the results on session_id.


-- incorrect method to filter for user sessions
select *
from sys.dm_exec_sessions
where session_id > 50

-- incorrect method to filter for user requests
select *
from sys.dm_exec_requests
where session_id > 50

You see, I have always been taught that system processes will always have a session_id of one through fifty, while user processes will always be greater than fifty.

Recently, I had the opportunity to work with a Microsoft engineer while troubleshooting some systems.  As we were running through my scripts he mentioned that this was no longer the case; system SPIDs can exist above fifty.  When I asked him how I should go about filtering user and system processes, he said he did not know of a way.

System Processes Have SPIDs Above 50

I was crestfallen to say the least.  Here I was, using an outdated technique and thinking I was cool. However, I was sure there must be a way to achieve the results I was looking for.

As luck would have it, the following week I was taking a SQLskills class with Paul Randal (b | t). During one of the breaks, I asked him about this and if he knew of a new method for filtering user and system processes.  True to form, he did some research, reached out to one of his contacts at Microsoft and had an answer for me within the hour.

New way?  What’s this about a new way?

Apparently, this behavior has existed since SQL 2005 but is not widely known.  The correct way to filter user processes is to query sys.dm_exec_sessions and filter on the is_user_process column.


-- list user processes
select *
from sys.dm_exec_sessions
where is_user_process = 1

-- list system processes
select *
from sys.dm_exec_sessions
where is_user_process = 0

So, in order to leverage this in other DMVs, you will need to use a join clause.  My updated code looks like this…


-- correct method to filter user requests
select der.*
from sys.dm_exec_requests der
	join sys.dm_exec_sessions des
	on der.session_id = des.session_id
where des.is_user_process = 1

An Exercise for the Writer

It always amazes me how long you can use a product and still stumble upon ‘new’ features at any moment.  So now, I will be reviewing my existing scripts and making updates where appropriate.