Tink vs metrics-server
Kubernetes metrics-server scrapes kubelets so kubectl top and HPA can see CPU and memory for a few minutes. That is not a mechanic. Tink installs in one command, watches the Linux server, explains what is wrong, and helps you fix it.
metrics-server powers kubectl top. Tink answers “why is this VPS sick?”
Accidental sysadmins land on metrics-server because every k3s install mentions kubectl top. Then the volume fills, nginx dies, and top still looks fine — disk, certs, and SSH are not millicores. A filling /var is a Tink warning; it is not a metrics-server sample unless the kubelet already reported CPU.
Tink is the other job: detect the issue on the machine, say why it happened in plain English, propose the command, and run it only after you approve. Keep metrics-server if you already operate Kubernetes and need HPA. Use Tink if you run Linux servers and want a mechanic.
| Feature | Tink | metrics-server |
|---|---|---|
| Setup time | 30 seconds (one curl | sh command) | Hours — cluster RBAC, an APIService, kubelet summary scrape, and still no alerts |
| What you get | Working monitoring, diagnosis, and approved fixes | In-memory CPU and memory for kubectl top and HorizontalPodAutoscaler |
| Pricing | Free (Scout) / $9 / $29 per machine per month | Free software (Apache 2.0). You still pay for Kubernetes, kubelets, and the engineer who owns the APIService |
| Hidden costs | None — fully managed | API aggregation, kubelet reachability, no history, and a second tool that actually watches the Linux host |
| Monitoring approach | Agent on the server — CPU, disk, services, logs, certs, ports | None on the OS — metrics-server only caches kubelet Summary API usage for a few minutes |
| Configuration | None after install — heuristics and AI | args, kubelet-insecure-tls, metric-resolution, and which nodes still answer the Summary API |
| Plain-English diagnosis | Yes — AI explains root cause, impact, and fix | No — kubectl top shows millicores. You still SSH or kubectl describe to learn why |
| Fix execution | Proposes and executes approved commands with an audit trail | Metrics only — metrics-server cannot free disk, restart nginx, or rotate a cert |
| Alerting | Built-in across 8 channels | None — metrics-server is not Prometheus and does not page anyone |
| Predictive alerts | Yes — disk fills in ~6 days, memory and CPU trends | Never — metrics-server discards samples after minutes. There is no trend |
| SSH brute-force detection | Built-in — parses auth.log every scan | Never — metrics-server does not read auth.log, journald, or host firewall state |
| Machine offline detection | Agent presence monitoring with multi-channel alerts | A missing node in kubectl top is a kubelet that stopped answering — only if you already noticed |
| Public status page | Shareable URL with 90-day history | None — metrics-server has no customer-facing status page and no 90-day store |
| Weekly fleet digest | Automated Monday digest + daily brief when issues are open | None — millicores in RAM are not a plain-English fleet narrative |
| On-call tracking | Built-in /oncall command + incident acknowledgment | Not included — wire Prometheus or another incident tool if you want pages |
| Conversation interface | Telegram, WhatsApp, web dashboard, CLI | kubectl top. No mechanic you text when /var filled on a VPS that is not in the cluster |
| Learning curve | None — works after install | Steep — Kubernetes, API aggregation, kubelet Summary API, and why kubectl top is empty |
| Best for | Freelancers, small teams, accidental sysadmins (1-50 Linux servers) | Teams that already run Kubernetes and need HPA / kubectl top, not a sick-VPS mechanic |
When metrics-server is the right choice
Keep metrics-server when you already want that exact job:
- kubectl top and HPA — you need millicores for autoscaling, not a mechanic on a quiet VPS.
- An existing Kubernetes cluster — kubelets, API aggregation, and someone who owns why top is empty.
- Short-lived resource usage — CPU and memory right now, with no history, logs, or disk.
- Hosts that only exist as Nodes — you already run Kubernetes, not a sick-box SSH loop.
The real cost of “just run metrics-server”
metrics-server is in every Kubernetes getting-started guide. A working ops loop for five Linux boxes is not:
- kubectl top still means you SSH in and change the box by hand
- Host fill, nginx, certs, and SSH brute-force never appear unless they show up as millicores
- A filling /var is a Tink disk warning on every scan — metrics-server will stay quiet until CPU already moved
- A single VPS with no cluster still needs a mechanic, not an APIService
For a 5-server team, Tink Mechanic at $45/month is cheaper than metrics-server plus the engineer who keeps kubelet scrapes and API aggregation alive, and every scan includes a diagnosis kubectl top will not type.
No APIService homework. No silent host. One command install.
Start MechanicAlso compare: Tink vs kube-state-metrics · Tink vs cAdvisor · Tink vs Prometheus