Before I get into this, it is not a knock on Uptime Kuma. It's a genuinely amazing, easy-to-use piece of software. If you run a homelab or a small fleet and you're not using it, you probably should be. It's free, self-hosted, beautiful, and it does the thing it was built to do better than almost anything else at any price.
There's always a "but," though, so before we get to it I want to spend a little time on what Uptime Kuma does well.
TL;DR:
Uptime Kuma is excellent at telling you when a service becomes unreachable, but it cannot explain why a service is slow or unhealthy while still responding. That requires internal metrics from tools like Prometheus, Grafana, and Alloy. Reachability monitoring and systems monitoring solve different problems, and mature environments typically use both.
Where Uptime Kuma excels
Uptime Kuma answers one question extremely well: is this reachable? It'll ping a host, hit an HTTP endpoint and check the status code, watch a TCP port, validate a TLS cert's expiry, query a DNS record, check a keyword on a page, watch a Docker container, even poke a game server. It checks on a tight interval, shows you a clean history, and when something stops responding it fires a notification through basically any channel you can name. Ninety-plus notification integrations. Status pages you can hand to your users. Two-factor auth. A genuinely nice UI.
For "tell me the moment my website, my reverse proxy, my Plex, or my Home Assistant stops answering," it's close to perfect. The interval is short, setup is measured in minutes, and there's practically no maintenance. It has earned a famously loyal userbase for a reason.
I'm not here to tell you it isn't the answer, or to convince you to ditch it for something else. I still think everyone running infrastructure of any size should have something like it watching their endpoints. I have it running in a Proxmox container on my own homelab. But there's a gap I noticed while using it, and this post is about the specific moment when you ask Uptime Kuma a question it wasn't designed to answer, and what you do when that moment arrives.
Growing pains
There usually comes a time, as your homelab or business grows, when your database starts feeling slow. Queries that used to be instant are taking a little longer. It's not a real problem yet, but you can tell something is off.
The Uptime Kuma dashboard is all green. The database port is answering, the HTTP healthcheck returns 200, every light on the board is on. Uptime Kuma is correctly reporting that the service is up.
And it's not wrong. That's the thing. The service is reachable. But "reachable" and "healthy" mean different things, and you've just walked into the space between them. If the disk that database lives on is pinned at 100% IO utilization because a backup job and a big query are fighting over it, your queries are queuing behind that contention, and from the outside the port still answers in time to pass the check. The board is green, the database is slow, and there's no contradiction.
Uptime Kuma doesn't see any of that, and the reason it can't isn't a missing feature, it's the architecture. It checks your systems from the outside looking in. It has no way to see what's happening inside your servers. What are the disk, memory, CPU, and kernel actually doing?
What you need at that moment is something standing inside the box, reading the system from within. That's a different category of tool.
Reachability versus internals
There are two kinds of monitoring, and once you see the split you understand why mature setups end up running both.
Reachability monitoring (Uptime Kuma) asks the basic questions:
- Can I get to it?
- Is the port open, the page loading, the cert valid, the container running?
It reports what it can see from the outside, which is exactly what you want for "is my service up and can my users reach it." It's easy, simple, and honest about what it knows.
Systems monitoring (the Prometheus world) asks questions that are a little more involved:
- What's going on inside the machine?
- How busy is each CPU core?
- How much memory is actually available once you account for cache?
- What's the disk IO utilization, the queue depth, the read and write latency?
- How much network throughput, how many dropped packets?
- Is memory slowly leaking over days?
It's an internal view, and it answers why a service is behaving the way it is.
Neither replaces the other. Reachability tells you that something is wrong. Systems metrics tell you why. The database scenario above needs both: Uptime Kuma to eventually notice if the slowness becomes an actual outage, and system metrics to explain the slowness long before it gets there.
The internal view
The standard way to get the inside view on Linux is a tiny agent called node_exporter. It's a small binary that runs on the box, reads metrics straight from the kernel, and exposes them for a time-series database (Prometheus) to collect. Pair it with Grafana for dashboards, and for logs, pair Loki with a shipper. The traditional choice there was Promtail, though Grafana has since moved Promtail into long-term support and now steers you toward Grafana Alloy, which handles both metrics and logs in a single agent. (I wrote a comparison of those two separately.)
With either node_exporter or Alloy running, the database scenario stops being a mystery. The exact moment things felt slow, you can pull up:
- Disk IO utilization on that box, and watch it pin to 100% right when the slowness started.
- The specific disk and the read/write split, so you can see it was the backup volume contending with queries.
- CPU broken out by mode, so you can rule out CPU as the cause.
- Memory availability over the past week, so you can see whether pressure had been building.
And if you have Loki collecting logs alongside the metrics, you can line up the disk IO spike against the log line where the backup job kicked off, and the whole story assembles itself in one view. Uptime Kuma told you the service was up. The system metrics tell you the backup job is strangling your database disk, which is what you actually need to know to fix it before it hits production.
(If the PromQL behind those dashboards is unfamiliar, I wrote up the five queries you actually need to monitor a Linux server separately.)
The honest part: this is more work
Standing up a stack to see inside your servers is not as simple as setting up Uptime Kuma, and that simplicity is a real part of why Uptime Kuma is so loved. Moving to system metrics comes with a cost.
node_exporter or Alloy goes on every server, with Prometheus running somewhere to collect from them. Grafana dashboards have to be built, or imported from the community and then tweaked until they're readable instead of overwhelming. Alert rules have to be written to fire on real problems without crying wolf. Metric and log retention have to be configured. And then the whole thing needs ongoing maintenance.
This is the irony nobody warns you about: you now have a monitoring stack that itself needs monitoring, which is partly why you want predictive disk alerts on the box running Prometheus.
None of it is hard, exactly. But it's an ongoing process with no end, and it's a different commitment than the near-zero maintenance of an Uptime Kuma container you set up once and edit when new services come online. Uptime Kuma is the right tool for reachability and status pages, and it costs almost nothing to run. System metrics cost more, but they become relevant the moment you start wondering why services aren't behaving, even though they're still showing green.
So what do you do with this
The real takeaway here isn't a product, it's the distinction, because that understanding outlives any particular tool. Outside-in tells you something broke. Inside-out tells you why. Uptime Kuma is one of the best outside-in tools ever made, and it'll happily keep doing that job for you forever. It just wasn't built to explain the why.
When you do need the why, you've got two options: run the inside-view stack yourself (node_exporter or Alloy, Prometheus, Grafana, Loki), which is completely viable and a great way to learn, or hand it to someone who runs it for you.
For full disclosure, that second path is the reason I built Irin Observability, a managed version of that inside-view stack for small teams and homelabs that have outgrown pure reachability checks but don't want a second full-time job maintaining a metrics pipeline. It's meant to sit alongside something like Uptime Kuma, not replace it, because reachability and internals are different jobs and the mature answer is to run both.
Either way, keep the green board. Just add the view from inside the box when you start asking why.
