FivexL built a scalable, secure, and elastic infrastructure to keep the Newstream SDK running fast. On FivexL’s side, there were two cloud engineering consultants involved - Artem Zhelezov and Maria Zubchenko. Here is what CTO of New Stream, Andrey Kolpakov, says about the beginning stages of working with FivexL: “The first task was to evaluate the existing infrastructure. Then FivexL suggested a few options, we eventually decided to go with AWS, and we started creating the new infrastructure.”

The work was planned in several iterations.
The first solution was a hybrid cloud: AWS + on-demand servers. The hybrid cloud itself went through several iterations - the initial version was built around HashiCorp Consul and Nomad. But as ECS Anywhere came into the picture, the infrastructure fully migrated to AWS ECS, which reduced the number of the self-hosted services and thus reduced the operators’ workloads.

In the meantime, CI/CD was set up with GitLab runners operating on Amazon to automate application delivery. All the services that were running on Amazon were defined as code using Terraform. The infrastructure became immutable, with the only remaining issues being fault tolerance and scalability, as servers in the private datacenter were at risk of physical impact - power outages, video card issues, etc.

To address the lack of resilience, the team began migrating to GPU Instances at AWS.

Another important consideration for any startup building a product based on resource-intensive Computer Vision and AI is cost optimization. The FivexL team and New Stream worked hard to reduce cloud costs. A custom auto-scaling based on a processing queue was implemented to eliminate underutilization of expensive cloud GPU Instances. This engaged servers only for processing tasks, which helped avoid overpaying for idle resources. Further cost analysis showed that computing resources for video transcoding were a significant expense. Transcoding operations and video storage were offloaded to the Vimeo service to lower overall costs significantly. Finally, replacement of the self-hosted Redis cluster with AWS SQS allowed for better resiliency and additional cost savings. As a result, the cloud cost went down by 40%.

The team also improved overall infrastructure security by separating AWS accounts, introducing IDS, migrating to temporary keys via AWS SSO, and providing role-based access for developers.