Azure table storage performance - REST vs. StorageClient

Question

I am working with Azure Table Storage, and trying to figure out the best way to increase performance. The queries that I perform are very simple - either an exact select using partition key and row key, or a where clause with a list (e.g., WHERE x==1 or x==2 or x==3, etc). Once I get the data back, I don't track it in a data context (no need for change tracking, etc). Saving data is likewise, so I only add it to the context to enable the save.

At the moment, I am using the .NET library (storage client). As I don't use the change tracking and other features of the TableServiceContext, I am thinking about using the HTTP API directly. Has anyone tried both options? If so, what kind of performance difference did you see?

Thanks, Erick

Answer 1

Table storage can be a bit of a fickle beast to optimize performance. There are a variety of factors that will impact it. Here are just a few off the top of my head:

1. Using a Partition Key in every query is a must. If you are not doing this, you are doing it wrong. If you use single PK and single RK (and only those two), it is no longer a query, but a resource GET and should be relatively instantaneous.
2. Do not use OR-based queries. This will cause a full table scan and your performance will be horrible. Instead, parallelize those queries within the OR statement.
3. Partitioning strategy will have a major impact. How many partitions you have and how often you hit them (to warm them up and cause the underlying partition servers to load balance) will cause dramatic differences. The size of the partition makes a big impact here too. Sequential partition keys is often a bad idea.
4. Small requests can benefit from turning off nagling (as previously mentioned).
5. Turn off context tracking and 100 continue (see here) can help as well.

There are many more I suppose that depend on your application. However, the ones I mention are generally the ones I start with.

Answer 2

Have you turned Nagle off?

• Nagle’s Algorithm is Not Friendly towards Small Requests

Of Interest:

• Maximizing Throughput in Windows Azure – Part 1