Export actual system uptime

[pxeger]

Host operating system: output of uname -a

Linux <redacted> 5.9.8-arch1-1 #1 SMP PREEMPT Tue, 10 Nov 2020 22:44:11 +0000 x86_64 GNU/Linux

node_exporter version: output of node_exporter --version

node_exporter, version 1.0.1 (branch: tarball, revision: 1.0.1)
  build user:       someone@builder
  build date:       20200616-18:35:17
  go version:       go1.14.4

node_exporter command line flags

None

Are you running node_exporter in Docker?

No

What did you expect to see?

The actual system uptime as reported by the first value in /proc/uptime.

What did you see instead?

Various "hacks" to get to the uptime which are all bad:

• sum(node_cpu_seconds_total{cpu=0}) - total time the kernel has reported the CPU doing anything, too inaccurate
• node_boot_time_seconds from /proc/stat's btime, which requires you to get the current time on the server and subtract it

[ArthurSens]

Prometheus has a built-in function for getting the current time. The PromQL below works quite nicely:

time() - node_boot_time_seconds

Is there any reason not to use it?

Another alternative is:

node_time_seconds - node_boot_time_seconds

[pxeger]

node_time_seconds is what I was missing. Thanks, although I'll leave it open because uptime to me is the most basic stat to report.

[ArthurSens]

Cool, glad you got that going :)

About adding an exclusive metric for uptime, I don't think that's gonna happen. Just like there isn't a metric for "used memory" for example.

Each metric created adds some extra storage usage. If some information can be retrieved by doing some simple math between others, then there is no need to create that extra metric.

e.g.,
memory used = total memory - free memory
uptime = current time - boot time

[septatrix]

To my knowledge btime in procfs will return false results (when used to calculate uptime) after changes in time occured due to e.g. daylight savings, leap seconds etc. I do not know of any way to mitigate these effects.

If there is a way then everything is fine otherwise however I think it would be reasonable to reopen this issue.

[greg-zamza]

Here is a specific example when node_time_seconds - node_boot_time_seconds does not work:

The server, which is an LXC container, stores not its own data in /proc/stat, but the host server's data. Thus, it is impossible to monitor the server's uptime in an LXC container using node_exporter, however, if you run the uptime command, it will show the correct value. So, using node_boot_time_seconds for uptime is wrong. The correct way is to use the value from /proc/uptime.

[SuperQ]

Running in a container that way is explicitly not supported. You need to be running the node_exporter in the host PID namespace.