A little while back I wrote an article on SQL Cache Dependencies. I had a colleague ask about these today so I think it's time to blog it. (For more information, I also did a podcast some time back with my old mates Wally McClure and Paul Glavich from the ASP.NET Podcast. It's at: http://weblogs.asp.net/pglavich/archive/2005/08/21/423160.aspx)
SQL Dependencies in ASP.NET v2
Traditional ASP Applications
In traditional ASP applications, every time the user would request a page, the application would retrieve the data from the database to render the page. Clearly this doesn’t scale very well. It works for a small number of users but not a large number. Often though, this data access can be completely avoided for certain types of data.
Consider the Production.ProductCategory table in the AdventureWorks database. Realistically, how often would it change? If we are building an application that displays the list of product categories, it doesn’t make sense to retrieve it from the database on every request. We could improve on this situation if the same user was retrieving this list regularly by holding it in the user’s Session state but that also wasn’t that helpful if there were a large number of users as many copies of the same data would be held.
A key advance in ASP.NET was the introduction of the Cache object. It provided a server-wide shared cache that could be used to hold relatively static data. We could read the list of product categories into a dataset and hold it in the Cache.
The issue with any cached data though is when to refresh the data in the cache. If we choose to refresh it once per day, we then could be showing an incorrect list for up to an entire day. If we base the refresh policy on elapsed time, we have a trade-off. The shorter the refresh time, the fresher the data but the less useful the cache is. The longer the refresh interval, the more efficient the cache becomes but the staler the data potentially is.
The ASP.NET Cache object in v1 gave us two options for cache invalidation. One was time based, the other was based on a file dependency. This was originally designed to allow for situations where the data was coming from an XML file stored in the filesystem. A filesystem watcher was set up to monitor the file. If the file had changed, the Cache would be invalidated and refreshed.
But the invalidation option that we really wanted was “when the data changes”.
Cache Invalidation Prior To ASP.NET v2
Prior to ASP.NET v2, there were a number of attempts to provide cache invalidation when the data changed. The most popular version was:
1. Create an extended stored procedure in SQL Server that simply touched a nominated file in the filesystem.
2. Add a trigger (for INSERT, UPDATE and DELETE) on the table to be monitored that called fired the trigger.
3. Set the cache to have a dependency on that file.
This option was miles ahead of where we started but it had a few downsides. First, writing extended stored procedures for SQL Server was not trivial. It couldn’t be done in managed code but a C++ template was available to use as a starting point. Second, the trigger would fire for any changes to the table, even if our query was only for certain rows in the table. Third, the mechanism was table-based and didn’t deal with queries that involved joins across multiple tables. Last, the mechanism was tied to a specific database engine and often to a specific version of a database engine.
ADO.NETv2 SqlDependency And SQL Server 2005
SQL Server 2005 added a wonderful new feature called Query Notifications. (NOTE: nothing to do with Notification Services – they often get confused). The idea here is that you could say to SQL Server “Here is a query I want you to execute. However, when you finish, I don’t want you to forget the query, I want you to remember it and, if the underlying data that it was based on changes in such a way that *if* I was to re-execute the query I’d get a different response, I want you to tell me about it”.
This is exactly what we need to support our ideal Cache invalidation scenario. We don’t want to re-execute the query, unless the data has changed. You might be thinking “that sounds like a lot of hard work for the database engine”. Well it is a bit of work but it’s in fact utilising the same mechanism that was already in place for detecting data changes underneath Indexed Views in SQL Server 2000. What it does mean is that we have some limitations on the types of query that this will work with:
· Use two part names for tables (SQL Server needs to know which schema so it knows exactly which table, rather than having to resolve that at query time)
· Reasonable complexity (not all queries can be implemented)
· Specific column lists (ie: No SELECT * FROM somewhere)
Normally we SQL Server, we send a query and get a response. We don’t expect unannounced data to arrive sometime later. So the next issue is how SQL Server can send us the notification. To do this it uses another new feature called Service Broker. Service Broker is a transacted queue that lives within the database server. Communication occurs between “services” attached to “endpoints”. We then need to set up an endpoint ready to receive the notification.
The ADO.NET team did a lot of this for us. They created a new object called the SqlDependency. It is a simple programmable object that wraps most of this functionality for us. Basically, it’s used by:
· Create a SqlCommand you wish to execute
· Wrap a new SqlDependency around the command (you pass the command to the constructor of the dependency)
· An event on the SqlDependency object fires when the data changes.
It’s that easy. The one remaining requirement is to set up the queue to listen for the notifications. You do this by calling the Start() method on the SqlDependency object once before you get going.
This new object can be used in any ADO.NET v2 application, including in WinForms applications. However, its primary use-case is in ASP.NET Cache invalidation. You can use it in code like this:
using (SqlConnection conn = new SqlConnection(“some connection string”))
SqlCommand cmd = new
SqlCommand(“SELECT EmployeeID, EmployeeName FROM Employees”,conn);
SqlCacheDependency dependency = new SqlCacheDependency(cmd);
SqlDataAdapter employeesDA = new SqlDataAdapter(cmd);
DataSet employessDS = new DataSet(“EmployeesDS”);
Cache.Insert(“emp”, employeesDS, dependency);
ASP.NET v2 SqlCacheDependency
Even though the ADO.NET v2 team had done most of the work, the ASP.NET team decided to make it even easier to use with the Cache object. In v2, the Cache object can now have a SqlCacheDependency in addition to the previous options of file and time dependencies. This can be applied to the Cache object and also to other cached objects like user controls. You can specify a directive like:
<%@ OutputCache Duration=”2400” VaryByParam=”none”
You can also directly use the SqlDependencies in code:
The code shown is a general form of the SqlCacheDependency settings. In ASP.NET v2, when using SQL Server 2005’s special mechanisms, we use the specific string “CommandNotification” rather than database and table names.
What If I Don’t Have SQL Server 2005?
Rather than tying their implementation directly to SQL Server 2005, the ASP.NET team also decided to create a pluggable architecture so other forms of SqlCacheDependency could be provided. For example, they have provided an implementation for SQL Server 2000. As SQL Server 2000 doesn’t offer query notifications or service broker, the implementation isn’t as good as the one for 2005. It works like this:
· A table is created to hold details of change notifications. (A command has been provided to help with this: aspnet_regsql.exe –d databasename –ed –E)
· Triggers are set up on the required tables. (Again a command has been provided to assist: aspnet_regsql.exe –d databasename –et –t tablename –E)
· ASP.NET polls the change notification table to look for changes.
Obviously polling is not as efficient as being notified but it’s still a whole lot more efficient that having the application constantly looking up the data.
This mechanism can then be extended to other database engines.
8 thoughts on “SQL Cache Dependencies”
and not to say, the difficulty to incoporate the use of this cache notification into the layered data access architecture.
The notification implementation for sql 2005 is not good at all. basically it's something that wants to cover all situations, it's too complex, its performance gain, i dont know, if any, very limited, the extra cost in query plan, and the timeout pulse period can not be configured, and the too many constraint if you really want to use this mechanism. Basically, it's something that sounds nice, but not practical.
Have to disagree I'm afraid. I've used it quite a bit with cache refreshing and it works a treat. I'm not sure why you suspect it has a great cost in the query plan or why you feel it's complex. From ado.net for example, you instantiate a sqldependency object, wrap it around a command and then handle the event that occurs when the data changes. It's hard to imagine it being much easier than that.
What timeout pulse period are you looking to configure? It is event driven not polled. The only restrictions are on the types of query that can be used but these are no different to those for indexed views.
Does this operate successfully with stored procedures?
Good Work Greg! Very detailed post on SQL Cache dependencies, well my opinion is bit different ASP.NET cache is good but only on a single server environment. As we try to scale out to a server farm; we start facing problems. I think instead of relying on SQL Cache dependencies in ASP.NET Cache we should go for distributed caching because it is scalable and highly reliable.
Is this something that can be done for some users and not for others. For instance wrap it in something like
if StaffID = 13 then
do cachedependecy and refresh
regualr db refresh
On a per-user basis you could decide where you get the data from (ie: cache or not) but that's certainly not the standard use case.
This was a indeed a very good article.
There are a few things that I would like to know – What are the differences of Data Caching and SQL Cache dependancy.
We had implemented the SQL Cache dependancy in our last project but had to discontinue it as due to notifications some of the transactions were very slow.
Currently I am designing a framework for my client and I have implemented Data caching.
The problem with this feature is that, currently I am doing a POC on a single Web Server, but the application has to be deployed on the proper multi server environment (maybe web farm). If I am using Data caching then it doesnt help as multiple users would access the same cache, and if cache is invalidated, then the data would be inconsistent for some users. One approach is attaching the SessionID (UserID) with the Cache object for each user, but then if you have more number of concurrent users then I think it is not advisable to have this feature as the Server would become very heavy.
I read an article on Velocity, but cannot implement that for a POC.
Thanks in advance.