CNCF: Optimizing Kubernetes admission control: Where Kyverno’s CEL admission cost actually goes
The Kyverno maintainers profiled the CEL-based admission engines and found about 86% of allocated bytes went to converting the admitted object, not evaluating policies. The conversion ran once per matched policy, so cost grew with policy count. The fix builds the request value once per request and ships in Kyverno 1.19.2. In the authors' single-machine Go benchmark with a 50 KB Pod and 64 matched policies, engine time dropped from 20.01 ms to 0.36 ms. The release also fixes a 12% to 23% regression in the legacy ClusterPolicy engine introduced between 1.16.4 and 1.17.0.
What changes for platform teams
Platform teams running Kyverno with many policies should see lower admission engine CPU and memory after upgrading to 1.19.2, and the per-policy conversion cost is gone. The figures are from single-host engine benchmarks that exclude webhook, network and etcd, so teams should measure in their own clusters.
Read the original at CNCF Blog ↗
written by claude-sonnet-5-5, prompt 7cbe154, checked by publish rules