Kubernetes Client Benchmarks (Python)
December 9, 2025 ยท View on GitHub
This repo benchmarks five Python Kubernetes clients using 5000 Deployment resources. Each client implements the same async methods: create_one, get_one, watch_all, delete_one. The suite focuses on end-to-end latency and overhead for create/get/watch/delete paths rather than application logic.
The resource under test is a Deployment with a deliberately large Pod template designed to stress model construction and serialization across clients.
Results
Benchmark results in this repo were collected against kind (Kubernetes in Docker), which provides a fast, consistent local environment for comparing client overhead under the same cluster conditions.
Combined benchmark results (objects per second).
| Client | Objects | POST | GET | Watch | DELETE |
|---|---|---|---|---|---|
| kubesdk | 5000 | 297.7 | 883.0 | 4222.3 | 2912.0 |
| kubernetes_asyncio | 5000 | 425.2 | 705.9 | 862.3 | 1586.1 |
| kr8s (async) | 5000 | 45.3 | 53.0 | 74.4 | 47.8 |
| lightkube (async) | 5000 | 44.2 | 55.9 | 3406.4 | 57.8 |
| official | 5000 | 38.1 | 52.4 | 507.4 | 1382.6 |