[Metricbeat][System module] psi: pressure stall information

[b-deam]

Describe the enhancement:

Since 4.2, Linux exposes contention metrics for CPU, Memory, and IO in procfs under the /proc/pressure path for a system as a whole, or for cgroup2 the subdirectory for each cgroup controller in the cgroupfs mount point contains cpu.pressure, memory.pressure, and io.pressure files.

An extra system metricset (perhaps called psi?) that collects these metrics on a defined interval would be extremely useful, both for specific cgroups and for systems as a whole.

Further Reading:

• https://www.kernel.org/doc/html/latest/accounting/psi.html
• https://lwn.net/Articles/759658/
• https://facebookmicrosites.github.io/psi/docs/overview.html

[elasticmachine]

Pinging @elastic/agent (Team:Agent)

[elasticmachine]

Pinging @elastic/integrations (Team:Integrations)

[ruflin]

@exekias @fearful-symmetry Would be great to get your take on this.

[exekias]

Thanks for bringing this up! Actually @fearful-symmetry have been working cgroups v2 support here: #27242, I see Pressure is actually retrieved.

Would you see this information useful if attached to the process documents?

[b-deam]

Not so much process documents - but rather the system as a whole (reading directly from /proc/pressure, kind of like system load averages are for the system as a whole.

I gave the PR a quick skim, but couldn't tell if that implementation will also allow collection at the host level if no cgroups are in use? Is that correct?

[exekias]

That's correct!, @fearful-symmetry would love your thoughts on this

[fearful-symmetry]

Considering we're already adding this in cgroups, I don't see any reason why we couldn't add global pressure stats as well. I'm not sure where this would go, though. Maybe just a new linux/pressure metricset?

[b-deam]

I agree, linux/pressure makes more sense than system IMO - I also noticed (if you weren't already aware) that the existing prometheus/profcs package already includes a method to retrieve these.