Many bloggers, including respected author Nicholas Carr, recently applauded Amazon for offering a Service Level Agreement (SLA) for its S3 storage service in response to competition from Nirvanix. Though SLAs have long proved an effective technique for managing corporate information technology, in Web infrastructure the SLA is nothing more than a scam that only a used car salesman could love. (Thanks, Paul D. Cocker, for the lemon.)
See what the Amazon SLA offers:
we commit to 99.9% uptime, measured on a monthly basis. If an S3 call fails … this counts against the uptime. If the resulting uptime is less than 99%, you can apply for a service credit of 25% of your total S3 charges for the month. If the uptime is 99% but less than 99.9%, you can apply for a service credit of 10% of your S3 charges.
Sounds great, doesn’t it? Let’s calculate “99% uptime,” which is the same as 1% downtime. A month has 43,830 minutes, so your site will be down 1% of that time, 438 minutes, or over 14 minutes per day on average. Suppose just 14 minutes of downtime occurs during your daily peak period. How will that downtime impact your revenue? Will being able to “apply for a service credit of 25%” compensate you for lost income, troubleshooting time, and unhappy customers?
Now consider the worst case of having all 438 minutes of downtime occur in one big failure during, say, the peak selling hours of the day after Thanksgiving. Then what recourse do you have? The dirty little secret is that the SLA protects the infrastructure provider, not the customer.
Amazon’s 99.9% uptime commitment means average daily downtime of 1.4 minutes. To put this statistic in perspective, I hosted a website for family and friends at a data center that, according to my measurements, delivered slightly worse than 99.9% uptime over a 3 month period. The number of complaints I received and the time I spent dealing with downtime astounded me. Do you think the infrastructure provider graciously provided me the 100% refund to which I was entitled according to its generous SLA? Think again.
I would have been fired for delivering 3 months of 99.9% availability back in the 1980s when I managed a corporate data center.
Web SLAs today are no better than the “9/90 warranty” my father offered years ago on his used car lot. If you shopped carefully, you could buy an awesome car at my dad’s store, but the warranty was only good for the first 9 seconds or 90 feet, whichever came first.