Modal ditches Kubernetes-style control to spin up 1 million sandboxes in seconds

By pushing state to workers and parallel schedulers, Modal shows how far large-scale sandboxing can go without traditional Kubernetes coordination.

Modal has redesigned its sandbox infrastructure to support around 1 million concurrent, isolated sandboxes and roughly 50,000 new sandboxes per second by abandoning traditional Kubernetes-style global coordination. Instead of a single, strongly consistent datastore and centralized scheduler, each worker node now holds its own authoritative state, while a horizontally scaled fleet of schedulers behaves more like load balancers and talks to workers directly via RPC. This design turns most O(nodes) and O(containers) operations into horizontally scalable components and avoids hitting Kubernetes’ etcd and scheduler scaling limits at very high pod and node counts. The only shared bottleneck is a Redis stream used for worker state publication, which testing suggests can handle well over 100,000 workers, enabling benchmarks of 1 million sandboxes created in under a minute with median startup-to-code times below 0.5 seconds. The work is being taken as evidence by some industry engineers that next‑gen AI and GenAI infrastructure may need architectures that decouple coordination from execution rather than extending Kubernetes, aligning Modal with efforts like Unikraft, Google Substrate, and Overdrive to rework cloud primitives for ultra-fast, large-scale isolation.