Modal rebuilt its execution platform to create one million isolated sandboxes in under a minute, addressing the bursty demand produced by coding agents and other AI workloads. The company reports median time from startup to running code below 0.5 seconds and a sustained rate near 50,000 sandbox creations per second.
The design avoids extending Kubernetes at extreme scale. Kubernetes scheduling and its central etcd data store accumulate work as node and pod counts rise, while rapid sandbox churn creates additional writes. Modal instead treats scheduling more like load balancing: multiple scheduling servers work in parallel, contact workers directly, and let each worker act as the authority on its own available resources.
Workers publish state into one Redis stream, which remains a shared bottleneck. Modal says load testing indicates that design is viable beyond 100,000 workers, but the result is the company’s own benchmark rather than an independent production audit. The architecture deliberately gives up some global coordination to keep the sandbox creation path simple and horizontally scalable.
That tradeoff fits short-lived execution environments used by AI agents, where a conventional container may spend more time starting than running. Workflows that need shared memory, overlapping security boundaries or stronger global consistency may still require a separate coordination layer.