Add process page fault metrics

[andrzej-stencel]

What are you trying to achieve?

I want to add process page fault metrics to the semantic conventions. Here are the specific metrics I have in mind, relative to how Linux reports them:

• The number of major ("hard") page faults that the process has made
  • This is reported in /proc/[pid]/stat Linux file, described as majflt field in the man page
• The number of minor ("soft") page faults that the process has made
  • This is reported in /proc/[pid]/stat Linux file, described as minflt field in the man page
• The number of major ("hard") page faults that the process's waited-for children have made
  • This is reported in /proc/[pid]/stat Linux file, described as cmajflt field in the man page
• The number of minor ("soft") page faults that the process's waited-for children have made
  • This is reported in /proc/[pid]/stat Linux file, described as cminflt field in the man page

I propose the following names for the metrics:

• process.memory.page_faults with the attribute type of either hard or soft
• process.memory.child_page_faults with the attribute type of either hard or soft

Additional context

I have read through the discussions on putting information in metric names vs. attributes (#2589 , #2726). I considered various options for the metric names and attributes, including putting the type in the metric name and putting the child into an attribute (couldn't find a good name for this attribute). Please let me know if you have any thoughts on this.

Related issue to add these metrics in hostmetrics receiver: open-telemetry/opentelemetry-collector-contrib#14084.

[andrzej-stencel]

Oops only now have I noticed that there's a similar convention already in place in the system namespace. The metric name is system.paging.faults with the type attribute of major or minor.

Should this proposal be changed to propose process.paging.faults? What about the child page faults - make them process.paging.child_faults?

[andrzej-stencel]

I'll take no comments as good news and will prepare a pull request to add process.paging.faults metric with a type attribute of major or minor. I will not add the child faults for now.

[rbailey7210]

Pleased pause for now, we are considering refactoring existing metrics semantic conventions. Please come to ongoing discussions. See #2753.