Adobe likes to talk about its public cloud partnerships with Microsoft and others, but it doesnt often talk about its private cloud strategy. Its no secret that there are plenty of good reasons for using a private data center and Adobe mana...
Date of publication: 07/11/2017
by Frederic Lardinois
likes to talk about its partnerships with and others, but it doesnt often talk about its strategy. Its no secret that there are plenty of good reasons for using a private data center and manages a few of these around the globe. For most businesses, opting for a comes down to cost, but for s Advertising Cloud, which previously flirted with the likes of AWS, moving back to a mostly came down to performance.
The Advertising Cloud as a brand is a relatively new concept for the company, though the actual product has been around for quite a long time. Its the companys service for helping brands manage and optimize advertising spend across channels. The service manages over $3.5 billion in annual ad spend from the likes of Allstate, Ford, MGM and Southwest Airlines, but to get those ads in front of the right eyeballs, it has to make billions of real-time pricing decisions per day to ensure that its users win the right advertising bids.
To build this , the team opted for using , the massive open source project thats building the equivalent of an AWS for businesses that want to manage their own clouds inside their own data centers.
Ten years ago, started betting on , but Nicolas Brousse, s director for operations engineering for the Advertising Cloud, tells me that the team always faced challenges with regard to latency. The only way to beat the market was to go , he told me. By 2013, the team started evaluating its options in earnest.
This was before became the de facto standard for s, so everything from to VMWare and was on the table. We ended up with because we liked the modular approach thats close to the Linux philosophy, said Brousse. We could take the pieces we care about and build around that. He also added, though, that the decision wasnt just about itself, it was about what were trying to accomplish.
In 2016, acquired the and with that, it gained even more expertise. TubeMoguls Joseph Sandoval, for example, was previously the director of cloud platform engineering at Lithium, where he made on the technology (A lot of people who made that early decision didnt come out unscathed, he rightly noted). Once at , Sandoval spearheaded the adoption of for the Advertising Cloud under the code name Cloud Mogul.
Brousse tells me that the teams move to improved raw compute performance by 3x over AWS. Now, the bottleneck isnt compute but the network edge, he said. He and Sandoval also noted that a lot of the tooling around the project is now far more mature and that the latest releases have been extremely stable.
Today, about 90 percent of the Advertising Clouds real-time bidding processes run on and the plan is to move to 100 percent very soon. Thats over 100,000 of compute nodes and growing, which makes this one of the larger deployments.
The team is still placing some workloads on the , though, and is looking at how it can make cloud bursting a reality for its services. Not all parts of the Advertising Cloud are all that latency constraint, for example, but Sandoval also told me that he wants the to be available to him so he can quickly stand up a cluster for overflow compute from the . In a , after all, you dont get the seemingly unlimited compute resources that the large s can provide.
Currently, the team that manages this part of s consists of only four engineers. That speaks to how easy it is to manage this service, but it also means that all of the engineers involved here need to be able to do each others jobs if necessary. Sandoval notes that this means the team has to automate as much as possible and for this, too, the tooling is now in place.
s use case is, of course, a bit of an edge case, given that the Advertising Cloud has some pretty clear constraints. Still, isnt the only company that is making similar decisions, whether its because running 24/7 workloads tends to get rather expensive on s like AWS, Azure or the Google Cloud Platform, or whether these s cant offer the kind of dependable performance characteristics that some workloads simply demand.