KUBEDIRECT: Unleashing the Full Power of the Cluster Manager for Serverless Computing
summary
KUBEDIRECT bypasses the API server and uses direct P2P message passing along a sequential controller chain, implementing pairwise agreements as a hierarchical write-back cache. It adds about 150 LoC per controller, uses hard and soft invalidations plus tombstone objects for safe termination, and runs on an 80‑node cluster. Upscaling latency improves 3.7×–16.9× over vanilla Kubernetes, per-function scheduling latency drops 26.7× versus Knative, and overall performance matches the clean‑slate Dirigent.
AI-written summary · about
chapters (10)
- 0:00KubeDirect and serverless
- 2:24Kubernetes vs bursty workloads
- 4:53KubeDirect design goals
- 7:25Direct P2P message passing
- 9:35Write-back cache for consistency
- 11:21Handling termination with tombstones
- 12:47KubeDirect implementation
- 13:50Evaluation results
- 14:53Limitations and future
- 15:14Q&A
related talks
34mKubeFlow’s Serverless Component: 10x Faster, a 1/10 of the Effort
34mPanel: Improving and Managing Kubernetes at Scale
12mKubespray: Driving Cost-Efficiency for AI on Kubernetes
24mPropagating Programming Paradigms: Lifting High Performance Compute Into K8s
37mKubeInception: running up to a thousand clusters in the real Cloud-Native way
45mYour K8s Infinity Gaunlet: The Marvel of KEDA and Cluster-API for Infinite Scale- 45mYour K8s Infinity Gaunlet: The Marvel of KEDA and Cluster-API for Infinite Scale
34mBuilding a Cluster Management API using Kubicorn