hwmon: Introduce chipName label carrying the human-readable name of the hwmon device

[horazont]

The chip label generation has been changed in #334 to prefer the unique device path (e.g. the location on the PCI bus) due to #333.

We propose the re-introduction of the human-readable device name (e.g. coretemp) as a separate label.

This allows to mitigate the downsides of the solution to #333 for cases where it does not matter that multiple devices may have the same human-readable name (e.g. aggregation or where at most one device of a type is present).

For cases where no name file exists in the hwmon directory, the same value as for the chip label is used.

@rtreffer @brian-brazil

[brian-brazil]

Thanks for the PR. Each label should add distinctness to the timeseries. As chip is already distinct, chipName doesn't add anything.

[horazont]

@brian-brazil Thanks for the feedback!

The concrete issue we are having is that we want to monitor GPU temperatures, which are available via hwmon. However, since the fix for #333, it is hard to address exactly those, because the PCI path is not guaranteed to be stable (and varies over machines), the sensor label is simply "temp1".

I see that this is not a problem for many devices, because you could, for example, easily select all coretemp sensors by matching for chip=~".*coretemp.*". The GPUs however are scraped simply with their PCI path as name, that’s not useful at all.

chipName adds value here, even if its not distinctness. If this is a hard limit, does it make more sense to prepend the name to the chip label? That would allow matching on that.

[brian-brazil]

The way to approach this would be using the machine roles method https://www.robustperception.io/how-to-have-labels-for-machine-roles/

Messing around with label values would just cause confusion, and lead to users using regex. Regexes are a smell.

[horazont]

I’m not sure how machine roles apply here, but you may be able to clarify. My idea to applying machine roles would be to have something like (via textfile-collector, omitting metric metadata):

gpu{chip="0000:00:02_0_0000:01:00_0"} 1

and then query using:

node_hwmon_temp_celsius and on (chip) gpu

Does that make sense?

This seems overly cumbersome for me, but the hack with prepending the name to the chip label doesn’t feel good either.

[brian-brazil]

You'd have both chip and chipName on the new metric.

[horazont]

So more like:

sensor_groups{chip="0000:00:02_0_0000:01:00_0", chipName="amdgpu"} 1

and then query using the group_left approach?

[brian-brazil]

Yes

[horazont]

Okay thanks. We still feel that it makes sense to use labels for these cases, but the other way works too, I guess.

Closing as it seems to violate the policy of the project and won’t be merged.

[SuperQ]

We could add these annotation metrics to the exporter without having to resort to doing these in the textfile. Just not as part of the original metrics themselves.

[horazont]

@SuperQ So a pull request implementing that would be fine?

[SuperQ]

Yes, I think it would be good.