Intuition, Performance, and Scale
Intuition, Performance, and Scale
Join the DZone community and get the full member experience.Join For Free
Maintain Application Performance with real-time monitoring and instrumentation for any application. Learn More!
Intuition is a double-edged sword. A blessing and a curse as it were. Intuition is knowing something with a reasonable sense of certainty without any justification for why you know it. It's this inability to clearly explain yourself when acting on intuition that is frustrating to you and those around you. Sometimes people will just trust your instincts but other times they will challenge you. There are many outcomes of such challenges. You may fail to convince anybody of your instincts. But, there are times when the challenge will lead you to a better understanding of your intuition. And that is exactly how I arrived at a simple realization of how performance and scale balanced to build usable, high volume web sites and services.
First, I should probably review performance vs. scale because there is a good deal of confusion about what the two qualities really are. Performance is a measure of responsiveness. How much time elapses while the service requested by the customer is carried out by the system. It doesn't matter whether this is a web page, a web service, or event. Time is the unit of measure for performance.
Scale measures how much work the system can perform within a given set of performance parameters. If the maximum acceptable response time for a web service is 500 milliseconds, scale will specify the maximum transactions per second the system can reach before response time exceeds the limit. Unlike performance, which can be accurately measured in most systems, scale is always an estimate. The system paths required to perform a customer request are pushed to their limits and the peak throughput is extrapolated. This is always a bit of science and a bit of faith as bottlenecks cannot accurately be predicted, especially at very high transaction rates.
As engineers we have a tendency towards optimizing performance. I suspect that one reason is that it's more exact. I can focus on individual hot spots, improve those, and measure an improvement in performance. Scale on the other hand is a bit more difficult. It often involves making seemingly illogical changes to the system that will often decrease performance. And that last bit is where engineers really object. How can you suggest approaches that are slower when the goal is increasing the throughput of the system?
A classic case in point is the ongoing debate about the performance advantages of stored procedures vs. application driven queries. The argument for stored procedure performance has always been based on several points. One is that stored procedures will have their statements parsed and query plans stored. This is true, but the reality is that most databases today do the same for all SQL so frequently used application SQL loses little to stored procedures here. The other main argument is that it is more efficient to process data near the data rather than pulling it into the application. This is mostly true. I say mostly because some aggregate and ordering operations may be more efficient to do in compiled code if the data sizes involved are relatively small. In those cases the data transfer overheads are often less than the database computational overheads.
So if stored procedures have an advantage over application SQL why would I be opposed to them. Am I a performance heretic? Because stored procedures don't allow the system to scale as well as moving the workload into the application servers. In fact, I go further than just avoiding stored procedures to reducing the complexity of queries as much as possible. I'm not a fan of ordering, grouping, or aggregation, especially if the application needs the full result set anyway. Look at any system architecture and the most difficult and expensive component to scale is the database.
Database servers are typically attached to SAN storage. Most sites use fiber channel for SAN so even if all the other hardware is equivalent in price to the application servers, database servers require SAN interfaces. If you are using a commercial database product, then depending upon the vendor and your contract, database servers also require additional license fees. It isn't unusual for a database server to cost anywhere from 150% to 300% more than an application server of equivalent performance. So from a cost perspective alone, I would prefer to push work to my application servers away from my database servers.
The bigger challenge comes as the database server reaches capacity. There are only two options available when this occurs. Scale out or scale up. Scale out requires a significant engineering effort while scale up puts you onto the slippery slope of bigger and bigger hardware and a less than desirable step function in capacity costs. So again, to the extent that I can delay capacity challenges on the database, I can maintain a simpler and less expensive architecture. Scaling out application servers is much simpler and for most application architectures comes almost for free until the number of servers becomes unwieldy.
This brings us to the fundamental "aha" moment. In building web applications to scale, you will regularly trade performance for scale. That's not to say that you will never pay attention to performance. Your performance has to be good enough to keep your customers happy. It turns out providing performance beyond that level though doesn't make them any happier and if you add that performance at the cost of scale, you're painting yourself into a dark corner.
If your goal is to create a site for the masses, then your architecture will need to revolve around protecting those components that are expensive or difficult to scale. Peak performance is not your primary concern. Acceptable performance as you bring on customers or your customers increase their use of your service is the challenge. Rather than measure each design against optimal performance for one customer, measure it against acceptable performance for all current and future customers. And be willing to give up a bit of response time for an increase in scale. Ultimately most customers won't pay you more for extra performance but you will earn more by having more customers or having them do more with your service.
Published at DZone with permission of Dan Pritchett , DZone MVB. See the original article here.
Opinions expressed by DZone contributors are their own.