May include but is not limited to: define client data access standards, connection timeout scenarios
There’s a nice article on TechNet on this topic and the next one, Design an efficient strategy to avoid data access throttling. I won’t publish another post for it, because there’s such a huge overlap between these subjects. The linked article above explains both of them pretty well.
As I have stated numerous times in these few posts (I think I’m constantly repeating myself) SQL Azure has the unique feature of killing your connections if you don’t play by the rules and try to monopolize all resources of the server where your database server resides. This feature is called throttling, and you should be aware of it and code your data access logic accordingly. This means that you should build in retry logic (the most preferred way for this is to use extension methods on the basic ADO.NET classes, such as SqlConnection and SqlCommand, according to the online examples floating around). Typically (a Microsoft recommendation) is to wait ten seconds and try again. Of course you can increase this time limit if it needed.
Trying to force something that fails constantly isn’t such a bright idea. So on the second or third approach, you could as well check out what the real problem is, and change your workload accordingly – if all else fails, then of course you could partition your database horizontally, but that’s a whole different topic.
Microsoft says you should do these three things in your data access layer: execute transactions in a continuous loop, catch connection termination errors and wait a little to reconnect. Now let’s see some possible reasons of database throttling. I remind you again to check out the TechNet article that is the basis of this post for a complete reference. So the following conditions result in throttling:
- Consuming more than a million locks.
- Uncommitted transactions.
- Locking a system resource for more than 20 seconds.
- A single transaction’s log file size exceeds 1 GB.
- Using more than 5 GB tempdb space.
- Consuming more than 16 MB of memory for more than 20 seconds when memory is limited.
- Exceeding maximum database size.
- Idle connections longer than 30 minutes.
- A transaction running for more than 24 hours.
- DoS attack (throttles connections from a particular (set of) IP address(es).
- Network errors.
- Failover errors.
These are the main reasons of connection throttling. Of course there’s a cure, and it comes in the form of some general best practices recommended by Microsoft. It makes such a perfect set of exam questions that I won’t leave out anything. So the best practices to avoid throttling:
- Minimize network latency (choose the closest data center).
- Reduce network usage: cache data and minimize round-trips to the server.
- Keep your connections open as short as possible.
- Set up short timeout durations in the connection string.
- Use connection pooling.
- Wrap all database operations in transactions with TRY CATCH blocks.
- Fine tune your T-SQL.
- Keep (or push) business logic in SQL Azure.
- Use stored procedures and batching.
May include but is not limited to: batch operations and performance techniques, data latency due to location, saving bandwidth cost
These topics were mentioned in some of my posts previously published, but a little review and repeat couldn’t hurt. This topic is about improving performance – both “real” performance, the raw speed of retrieving data, and both perceived performance, such as not blocking your UI thread while you’re waiting for data to be received.
When dealing with the Azure environment you’d like to be cost effective (you should be cost effective in every case, but cloud computing is a bit special, since you literally have to pay for your laziness). This means that you should design your data access strategy in a chunky manner. Roundtrips to the database over encrypted connection tend to be slow, not to mention the possibility of failure. Because of this, you’d like to query bigger chunks of data, and let the SQL Azure engine process that for you to save computing costs, too. You’d better use views and stored procedures, and implement your business logic in them, very close to the data (an arguable design, I admit). Also, you’d like to submit your changes in bulks – a good ORM such as EF or Hibernate can do this for you.
Another great way of reducing bandwidth usage is to load the data that is absolute necessary, and not anything else. The concept of lazy loading is very useful here – and mature ORM tools provide this for you nowadays. Note that lazy loading and chunky data retrieval are contradictory at best, so you have to find the right balance between them.
You can do a lot by caching your results. If you deploy an ASP.NET application to the cloud, you get a great and easy-to-use caching-feature.
The concept of sharding (or horizontal partitioning, or federating…) can help a lot to scale an application horizontally. The idea is relatively simple: you have multiple databases with the exact same schema, and store data based on some criteria (geographic location, customer ID, etc.) in them. The benefits are smaller indexes and faster query results; the drawback is complication in data access code.
The techniques above are real performance tuners, but you can easily boost the perceived performance of your applications by not blocking the main (UI) thread. Make database calls asynchronously, show nice animations while loading, and your app will be judged faster.