metrics-server VS prometheus

- name: Installer Metrics Server shell: | kubectl apply -f https://github.com/kubernetes-sigs/metrics-server/releases/latest/download/components.yaml environment: KUBECONFIG: /etc/kubernetes/admin.conf - name: Patcher pour ignorer TLS (dev seulement) shell: | kubectl patch deployment metrics-server -n kube-system --type='json' \ -p='[{"op": "add", "path": "/spec/template/spec/containers/0/args/-", "value": "--kubelet-insecure-tls"}]'

# Download metrics-server manifest curl -sSL https://github.com/kubernetes-sigs/metrics-server/releases/latest/download/components.yaml -o metrics-server.yaml # Edit to add --kubelet-insecure-tls flag for self-signed certificates kubectl apply -f metrics-server.yaml # Verify installation kubectl get pods -n kube-system | grep metrics-server kubectl top nodes kubectl top pods

kubectl apply -f https://github.com/kubernetes-sigs/metrics-server/releases/latest/download/components.yaml

# patches/metrics-server.yaml machine: kubelet: extraArgs: rotate-server-certificates: true files: - content: | [metrics] address = "0.0.0.0:11234" path: /var/cri/conf.d/metrics.toml op: create cluster: extraManifests: - https://raw.githubusercontent.com/alex1989hu/kubelet-serving-cert-approver/main/deploy/standalone-install.yaml - https://github.com/kubernetes-sigs/metrics-server/releases/latest/download/components.yaml

Kube-State-Metrics is only one part of the Kubernetes metrics ecosystem. Metrics-Server is another commonly discussed component that you should also include in your observability strategy. Whereas Kube-State-Metrics reports metrics about the states of objects in your cluster, Metrics-Server exposes real-time CPU and memory resource utilization metrics for your Nodes and Pods.

apiVersion: metal.sidero.dev/v1alpha1 kind: ServerClass metadata: name: masters spec: qualifiers: hardware: - system: manufacturer: QEMU memory: totalSize: "12 GB" configPatches: - op: add path: /machine/network/interfaces value: - deviceSelector: busPath: "0*" dhcp: true vip: ip: "10.1.1.50" - op: add path: /machine/network/nameservers value: - 1.1.1.1 - 1.0.0.1 - op: replace path: /machine/install value: disk: none diskSelector: size: '< 100GB' - op: replace path: /cluster/network/cni value: name: none # name: "custom" # urls: # - "https://raw.githubusercontent.com/kubebn/talos-proxmox-kaas/main/manifests/talos/cilium.yaml" - op: replace path: /cluster/proxy value: disabled: true - op: replace path: /machine/kubelet/extraArgs value: rotate-server-certificates: true - op: replace path: /cluster/inlineManifests value: - name: cilium contents: |- apiVersion: v1 kind: Namespace metadata: name: cilium labels: pod-security.kubernetes.io/enforce: "privileged" - op: replace path: /cluster/extraManifests value: - https://github.com/kubernetes-sigs/metrics-server/releases/latest/download/components.yaml - https://raw.githubusercontent.com/alex1989hu/kubelet-serving-cert-approver/main/deploy/standalone-install.yaml

Just add `--kubelet-insecure-tls`

that solves the problem!

https://github.com/kubernetes-sigs/metrics-server/issues/196

Before deploying anything into this namespace, it’s important first to verify that the metrics server is deployed and configured in your cluster. This metrics server is part of how Kubernetes implements autoscaling rules. Check whether it’s installed by running:

Metrics Server is a Kubernetes addon that collects CPU and memory resource utilization information at the Node and Pod level. It's a lightweight, single-cluster, Kubernetes-only alternative to more complex monitoring solutions like Prometheus.

To monitor and analyze rate limiting metrics, we're using a combination of Redis and Prometheus. We're storing rate limiting metrics in Redis and then using Prometheus to scrape the metrics and display them in a dashboard. Here's an example of how we're storing rate limiting metrics in Redis:

In this post, we compare two forecasting models, Chronos (Chronos‑Bolt) and Toto, on telemetry from Prometheus and OpenSearch. We judge them with two easy metrics: MASE for point accuracy and CRPS for the quality of uncertainty.

Stack & Versions: Go 1.21, PostgreSQL 16, Kafka 3.6, Kubernetes 1.29, prometheus/prometheus v2.48.0 for metrics

For monitoring infrastructure, Prometheus and Grafana are widely used for pipeline metrics collection and alerting. For orchestration that includes built-in run observability, Apache Airflow tracks run history, task durations, and failure states in a web UI. Python with SQLAlchemy is the standard stack for custom pipeline implementation with relational state management.

The OpenTelemetry ecosystem has three main components: instrumentation, the Collector, and exporters. Instrumentation is how you integrate OpenTelemetry into your application, using language-specific SDKs to create spans, record metrics, and propagate context. The Collector is an optional middleware component that receives, processes, and exports telemetry data to one or more backends. Exporters are the plugins that actually send your data to a specific destination, like Prometheus or Jaeger.

The fix is switching from naive serving to optimized serving, which means deploying the same model differently. High-performance teams running Llama-3-70B in production have converged on a specific stack: vLLM or SGLang as the inference engine, Prometheus for observability, and Runpod as the infrastructure layer that lets them deploy and iterate without managing a Kubernetes cluster. This guide works through that stack in ROI order: quantization (VRAM footprint), serving engine selection (throughput), speculative decoding (latency), and deployment mode (cost-scaling).

There are a lot of free and paid tools to monitor these variables. I almost always do this type of test in a VM (easier to clean up the mess when it all breaks) and I like to use Prometheus but honestly Perfmon in Windows or Top in Linux gives you all you really need.

Prometheus + Alertmanager