eBay aims to move 100 per cent of its ebay.com service onto OpenStack

OpenStack has improved leaps and bounds in the past four years but it still leaves much to be desired in terms of upgradability and manageability, according to Subbu Allamaraju, eBay’s top cloud engineer.

Allamaraju, who was speaking at the OpenStack Summit in Vancouver this week, said the ecommerce giant is a big believer in open source tech when it comes to building out its own internal, dev-and-test and customer-facing services.

In 2012 when the company, which is a 100 per cent KVM and OVS shop, started looking at OpenStack, it decided to deploy on around 300 servers. Now the company has deployed nearly 12,000 hypervisors on 300,000 cores, including 15 virtual private clouds, in 10 availability zones.

“In 2012 we had virtually no automation; in 2014 we still needed to worry about configuration drift to keep the fleet of hypervisors in sync. In 2012, there was also no monitoring,” he said. “We built tools to move workloads between deployments because in the early years there was no clear upgrade path.”

eBay has about 20 per cent of its customer-facing website running on OpenStack, and as of the holiday season this past year processed all PayPal transactions on applications deployed on the platform. The company also hosts significant amounts of data – Allamaraju claims eBay runs one of the largest Hadoop clusters in the world at around 120 petabytes.

But he said the company still faces concerns about deploying at scale, and about upgrading, adding that in 2012 eBay had to build a toolset just to migrate its workloads off the Essex release because no clear upgrade path presented itself.

“In most datacentre cloud is only running in part of it, but we want to go beyond that. We’re not there yet and we’re working on that,” he said, adding that the company’s goal is to go all-in on OpenStack within the next few years. “But at meetings we’re still hearing questions like ‘does Heat scale?’… these are worrying questions from the perspective of a large operator.”

He also said the data from recent user surveys suggest manageability and in particular upgradeability, long held to be a significant barrier to OpenStack adoption, are still huge issues.

“Production deployments went up, but 89 per cent are running a core base at least 6 months old, but 55 per cent of operators are running a year-old core base, and 18 per cent are running core bases older than 12 months,” he said. “Lots of people are coming to these summits, but the data suggests many are worried about the upgrading.”

“This is an example of manageability missing in action.  How do you manage large deployments? How do you manage upgradeability?”