Beyond Kubernetes at Modal: How to Scale 1 Million Concurrent Sandboxes in Seconds
Modal describes replacing centralized sandbox scheduling with horizontally scaled schedulers that choose workers from cached state and ask those workers directly to create sandboxes. Worker state is published asynchronously to Redis; durable metadata stays out of the creation path. In the company’s test, one million sandboxes were created in under a minute, with median time to run user code below half a second.
Those are vendor-reported benchmarks, and the new design trades global coordination for scalability. Modal calls a single Redis stream its nearest remaining bottleneck and says testing kept it viable beyond 100,000 workers. InfoQ quotes one LinkedIn reaction that scale brings new problems at each order of magnitude, and another that Modal sidestepped Kubernetes rather than extending it; neither is an independent benchmark.