Case Study: How Runway powers AI media creation with Anyscale
Key results
The challenge
Runway ran Ray on KubeRay, but it became unmanageable once four or five researchers submitted jobs to the same cluster at once, with users accidentally misconfiguring resources and disrupting each other's jobs. The team lacked observability tooling and considered hiring someone full-time just to manage the KubeRay infrastructure.
The solution
Runway moved to the Anyscale platform, including Anyscale on GKE, so researchers could submit jobs and monitor them from the Anyscale dashboard without managing clusters, while retaining Kubernetes-level security controls.
“Using Ray was a really straightforward decision. It's hard to even compare Ray to anything – it's a bit of a default at this point for this type of thing.”
CWCindy WangStaff ML Engineer, Runway
The results, in context
On Anyscale, Runway achieved 13x faster model loading and an 85% reduction in data pipeline development and deployment time, cutting it from one week to one day. The platform supports 40-50 Runway engineers and helped the team build and launch its Gen-3 Alpha model in the summer of 2024.