commotion_
See what your Linux boxes are really doing.
Metrics first, for Linux processes. Every process on every node is judged on the node, once a second. Almost always the answer is nothing is wrong, and nothing is sent.
Status in the workshop
Tell me when it’s readyThe premise
You don’t take a blood panel every morning.
Today’s stacks collect everything, ship everything, store everything, and then sell you a second product to filter what the first one collected. Commotion asks the question on the node and sends the answer, not the raw material. The bill it doesn’t create is the bill you don’t pay.
01
Per process.
Not per node, not per pod, not per container. The thing that is actually suffering is a process, and that is the unit Commotion watches.
02
Saturation, not usage.
A process pinned at 100% of a core with nothing queued behind it is working, not struggling. Commotion alerts on waiting, stalling and failing.
03
Four resources, one source.
CPU, memory, disk, network. When several go bad at once, Commotion names which one started it, and shows the counters that say so.
04
Measured, not guessed.
The link between one resource and the next is read from the kernel’s own per-task accounting. No correlation engine, no training window, no black box.
The flight recorder
The two minutes before the alert are the only two that matter.
Start collecting when the symptom appears and the cause has already left the building. Commotion keeps a full-fidelity recording in memory at all times and writes it out when a process goes bad, so every alert arrives with the run-up attached. Ask why, at 03:14, and there is an answer.
What it costs to run
Design targets at ten thousand nodes. Measured numbers will be published with the first release, alongside the method used to get them.
Where it sits
Keep the stack you have.
Commotion speaks OTLP and Prometheus remote-write, and exposes what it finds as ordinary series. Point your collector at it, keep your dashboards, keep your alerting. Nothing here asks you to move first and find out later.
Intended pricing
Flat, per node.
Not per series, not per gigabyte, not per host-hour of whatever the agent felt like sending. The architecture is what makes that honest.
While it is in the workshop, Pop Task is out now and three more apps are coming soon.
See the apps