Netflix Titus Adds Single-Tenant VM Isolation for Containers

Netflix has introduced a new mode for its container orchestration tool Titus, allowing one task to run on its own EC2 instance with single-tenant virtual machine isolation. This mode is designed for stateful, high-throughput workloads that require performance isolation, rather than protection against malicious tenants. While shared-kernel tuning can improve efficiency, it does not fully eliminate the gap in performance caused by shared hardware resources.

Netflix found that MicroVMs like Firecracker and Kata, as well as gVisor, introduced too much overhead to be practical alternatives. Instead, Netflix relies on the Titus federation layer and an allowlist to control which instance families a task can use. The autoscaler selects the smallest instance that meets the task's resource requirements, and additional tools like pod anti-affinity and a non-preemptible priority class enforce the isolation.

Netflix has implemented measures such as disabling defragmentation and keeping a small warm pool of pre-warmed instances to reduce launch latency. While this approach offers significant performance benefits for certain workloads, it comes at the cost of reduced multi-tenant density.