Netflix Introduces Single-Tenant Titus for VM-Level Isolation on Kubernetes

Netflix implemented Single-Tenant Titus to provide VM-level isolation for stateful and high-throughput workloads, eliminating noisy neighbor interference caused by shared kernel resources like last-level cache and memory bandwidth.
The solution allocates dedicated EC2 instances to specific workloads while preserving the existing Titus API, deployment pipelines, and Compute/Platform team ownership of host patching and scaling.
Fleet efficiency improved because moving sensitive workloads to dedicated nodes removed anti-affinity fragmentation, allowing denser packing of the remaining multi-tenant fleet and reducing self-inflicted disruptions from repacking.
Future improvements include replacing HTB network shaping with eBPF-based Expected Departure Time and evaluating Intel RDT, AMD PQoS, and NVIDIA MIG for hardware partitioning on the multi-tenant side.

Implementation & Scheduling

Workloads opt in via an allowlist-based federation layer that tags tasks with single-tenant labels and allowed instance families.
Autoscaling provisions just-in-time EC2 instances matching requested CPU, memory, and GPU resources to minimize unused capacity on dedicated nodes.
Scheduler uses pod anti-affinity, preemption-exempt priority classes, and node taints to prevent colocation and unrelated workload placement.

Isolation benchmarks use differential testing with 'havoc' containers (UDP floods, cache/memory stress) to quantify noisy-neighbor impact via throughput and latency efficiency metrics.
Launch latency is higher for single-tenant due to full boot and image pull costs, mitigated by lazy-pulling snapshotters (stargz, SOCI) and CRIU checkpoint/restore.
Small instances incur higher fixed platform overhead (tens of percent) compared to large instances (couple of percent), making the mode suitable for large, long-running workloads rather than small, bursty ones.