Machine-learning research isn’t something that Amazon Web Services does just for its cloud customers; it has developed machine-learning driven models on how to forecast demand for its services and deploy infrastructure accordingly.
AWS CEO Andy Jassy told attendees at the Pacific Science Center’s 14th Annual Foundations of Science Breakfast yesterday that Amazon has been using machine learning to anticipate demand for its services as deals work their way through the pipeline and predict how and where it should add capacity. “One of the least understood aspects of AWS is that it’s a giant logistics challenge, it’s a really hard business to operate,” he said.
The surge in interest around cloud computing over the last few years seems like it would make solving this challenge easy: just buy all the servers, because the workloads keep coming. “Every single day we add enough new servers to have handled all of Amazon as a $7 billion global business,” Jassy said, indicating the scale at which AWS now operates.
In practice, it’s of course more complicated than just buying servers by the truckload. Geographic demand for AWS services is spreading around the world and financial analysts are tracking Amazon’s capital spending very closely.
So AWS uses a forecasting model driven by machine-learning research to make capacity decisions, Jassy said. For example, it can pick up signals from the process its sales teams follow (enterprise sales cycles are notoriously long) to forecast demand. A lot of new customers like to start slow on AWS and then accelerate their usage as they see more benefits, Jassy said, which can lead to spikes in demand if they move faster than anticipated.
Likewise, the company has also determined where to store excess components for its data centers to be able to react quickly to a need for more capacity in a given region, he said.