Modal Rebuilds Sandboxes for 1M Concurrent Environments
Modal Rebuilds Sandboxes for 1M Concurrent Environments
Source: https://x.com/i/status/2077855275046838720
📌 Modal rewrote its sandbox control plane to remove central database bottlenecks, enabling massive parallel creation of isolated containers for AI agents and RL workloads. A demo ran one million concurrent sandboxes created in under a minute, with median start under half a second.
🧪 Why sandboxes matter
Sandboxes are short-lived, programmable environments for untrusted or agent-generated code—coding agents, app previews, and RL rollouts that need huge parallel trajectory sampling.
🧱 Old bottlenecks
Kubernetes-style designs rely on strong consistency and central stores (Postgres for Modal, etcd for K8s), creating O(containers) and O(nodes) limits under heavy churn and burst creation.
⚡ New control plane
Horizontal in-memory schedulers; workers publish state to Redis streams and act as source of truth; create path hits workers directly with no durable DB on the critical path (~two network hops + cheap CPU).
📈 Scale claims
Demo: 1M concurrent sandboxes in under a minute; tens of thousands created per second; median time-to-interactivity under ~0.5s; scheduling on the order of tens of milliseconds.
🚀 Availability
New stack is in beta (opt-in) and intended to become the default for all Modal sandbox scheduling; prior product topped ~50k concurrent per customer and millions per day.
Key facts
| Fact | Value |
|---|---|
| Company | Modal (AI/serverless compute; sandboxes core product) |
| Announced | July 16, 2026 (blog + X by Colin Weld & Connor Adams) |
| Scale demo | 1M concurrent sandboxes, created in under 1 minute |
| Create throughput | Tens of thousands of sandboxes per second |
| Start latency | Median create-to-runnable under ~0.5 second |
| Architecture | No central durable store on create path; workers + Redis streams |
Details
Modal, a high-performance AI cloud infrastructure company, rebuilt its sandbox platform so teams can spin up isolated containers at extreme scale. Demand from coding agents, preview sessions, and reinforcement-learning rollouts has outgrown traditional container platforms that bottleneck on central consistent stores.
The new design replaces Postgres-backed strong consistency on the create path with a fleet of in-memory schedulers and workers that publish state into Redis streams and serve as their own source of truth. Sandbox creation talks to workers directly, so capacity scales by adding schedulers and workers rather than fighting a global coordinator.
Modal reports creating one million concurrent sandboxes in under a minute, with median time-to-interactivity under half a second and design throughput of tens of thousands of creates per second. The prior product supported roughly 50k concurrent sandboxes per customer and millions per day; the new stack is already in beta and slated to become the default scheduler.