Kubernetes Adds Overhead: Benchmark Shows No Latency Improvement for Low‑RPM Workloads
Avoid migrating low‑RPM workloads to Kubernetes unless you need scaling.
Benchmark your workloads on Kubernetes before migrating to avoid latency penalties.
Summary
The benchmark study compares a Spring Boot 3.2.1 monolith running on bare‑metal Docker with a Kubernetes 1.29.3 cluster on a single AWS c6i.xlarge node across five target RPM levels. Each workload runs 10 iterations of a 5‑minute warm‑up and 10‑minute test, measuring mean latency, p99 latency, and throughput. Results show the monolith consistently outperforms Kubernetes, with mean latency ranging from 12 ms at 100 RPM to 89 ms at 10 000 RPM, while Kubernetes ranges from 14 ms to 92 ms. The study attributes the overhead to container runtime, CNI overlay, service discovery, and pod health checks, totaling roughly 47 ms of mean overhead. At low RPMs, this overhead represents a 26–37 % increase, but at high RPMs it becomes a negligible 3–5 % of total latency. The article notes that 60 % of Kubernetes migrations for stateless workloads under 5 000 RPM will be rolled back by 2026, per Gartner. It also highlights that disabling Calico and using host networking can reduce latency but sacrifices isolation. The benchmark underscores that Kubernetes is not a silver bullet for performance and that careful evaluation is needed before migration. The author recommends benchmarking workloads before deciding to move to Kubernetes.
Key changes
- 10 iterations on AWS c6i.xlarge (4 vCPU, 8 GB RAM) for monolith and Kubernetes.
- Monolith mean latency 12–89 ms; Kubernetes 14–92 ms across 100–10 000 RPM.
- p99 latency overhead 47 ms due to container runtime, CNI overlay, service discovery, pod health checks.
- Low‑RPM workloads see 26–37 % latency increase; high‑RPM workloads see only 3–5 % increase.
- 60 % of Kubernetes migrations for stateless workloads under 5 000 RPM will be rolled back by 2026.
- Disabling Calico and using host networking reduces latency but removes isolation.
- Benchmarking is essential before migrating to Kubernetes.