Turning On the Tap
How to Increase the Reliability of Your IT Infrastructure Using Predictive Analytics REGISTER >
Turning On the Tap
Powerset, the search-engine company using EC2 and S3, dreams of challenging established competitors such as Google Inc. by offering searches that use natural language, rather than stilted phrases and imprecise key words. It is a task that involves intense computational capacity. The young company, which has raised $12.5 million in funding, wants "to avoid spending resources on anything else than our core business," says CEO Barney Pell. Time was also an issue for Powerset. "There is a risk that building your own infrastructure can take longer than you thought," says Pell. Instead, he outsourced most of his computing and storage jobs to Amazon.
Amazon was significantly cheaper than other large data-center companies Powerset investigated, says Pell. "A number of the big players charge about one dollar per CPU hourwe are paying a fraction of that." Powerset might have gotten its costs down as low as five cents to seven cents per hour by building its own systems, says Pell, without factoring in development or human-support costs. "If we purchased our own equipment and amortized the cost over three years, the cost per CPU hour would be competitive. But it's easy to write it down on an accounting basis, and you still have to spend the cash. This is like taking the amortized number without spending the cash. Amazon does the investment for you. It's a fantastic value proposition for us," Pell says.
Beyond the price, the payment model is a critical part of the equation, says Pell. "Elasticity is important. We don't know how fast demand for our service will grow, and we don't know how many queries we will need to support each second," he says. "Users come in burststhere are sudden spikes of traffic and interest. We can model demands and the needs of hardware, and we're certainly going to be wrong. If we overestimate peak capacity, we spend too much money; if we underestimate, we will be turning away users or giving them a suboptimal experience. But a system that identifies bursts of traffic and lights up second and third strings of servers solves that problem."
Powerset plans to launch for public use in 2007, and is using EC2 for its intensive offline activities: the background work of reading, processing and indexing Web pages that underlies search. "The computers doing that work could almost be buried underground and not even connected," says Pell. Still to be decided: whether Powerset will rely on Amazon for the "run-time" application that powers actual searches. "We are investigating if we can make serving user queries work on EC2if our technology works in their infrastructure," says Pell. "That could be another massive game changer for us."
In many ways, the user experience with these services is different from a traditional hosting relationship. Webmail.us executives have never visited an Amazon facility, for example. "It's all self-service," says Boebel. "We started playing around with it, then engaged with one of their business development people. They worked with us pretty closely." S3 involves no client software and no hardware on the user's site. "There is just some simple code to execute on your server, and you're done. You could set up a server to back up to S3 in an hour and never touch it again," he says. Unlike traditional rack rental, says Boebel, Amazon has intelligence built into its network of servers. "The logic is there to handle failure. You can't just lease a server to do that."
Another difference: Amazon doesn't offer service-level agreements that guarantee the amount of time services will be available. The company argues that its own reliance on the platform is enough be give outside users confidence. "Amazon is a customer of our platform," says Selipsky. "That's important philosophically. We believe if it's good enough for our customers, it's good enough for us to use. We have mission-critical data stored in S3, and we have groups using EC2. It's important that we really understand all the good things and bad things we put out there, and we get incredibly good feedback from our own people."
That's good enough for Webmail.us, up to a point. "I'm very comfortable, because they rely on the same infrastructure for their own business," says Boebel. "If S3 goes down, they're hurting." Webmail.us has not seen any downtime since it went live with the storage service in September 2006; in early January, Amazon S3 did experience its first significant outage. Boebel likes the fact that Amazon stores multiple copies of its data, including at least one copy in a separate data center. But Webmail.us CEO Pat Matthews says that SLAs of the sort offered by Rackspace matter a lot for his core business, and he expects other companies may find the absence of SLAs more of an issue.
In the long run, though, says Tier 1 Research's Golding: "SLAs are useless. An SLA is not a guarantee of good service, it's a bad rebate. Companies architect a solution as well as they can, up to the limits of the existing technology. In the on-demand world, it makes sense to do away with them. If customers don't have a commitment, and they don't like what they're getting, they will go somewhere else."