ceilometer is not collecting all the available compute metrics

subscribed ~field-medium

Looking into the ceilometer-agent charm, it looks like it just takes the default polling.yaml for ceilometer [0]. This would need to be configurable as it is in the ceilometer charm. Even if you set enable-all-pollsters in ceilometer config, the ceilometer-agent units do not share the same polling info, so it won't be able to report those metrics.

I was able to test this by manually changing /etc/ceilometer/polling.yaml on a ceilometer-agent unit, adding a metric that's Origin is Pollster [1] and restarting the ceilometer agent service. In my case, it was ceilometer-agent on a nova compute and the service was ceilometer-agent-compute.service. After the restart, the added metric was available.

So to fix, we likely need to add configurability to ceilometer-agent's polling.yaml.

[0] https://github.com/openstack/ceilometer/blob/master/etc/ceilometer/polling.yaml
[1] https://docs.openstack.org/ceilometer/latest/admin/telemetry-measurements.html

Fix proposed to branch: master
Review: https://review.opendev.org/c/openstack/charm-ceilometer-agent/+/789876

Reviewed: https://review.opendev.org/c/openstack/charm-ceilometer-agent/+/789876
Committed: https://opendev.org/openstack/charm-ceilometer-agent/commit/9fd4ea0a39a9bf44dfd3563ea819a57c4eadd9ff
Submitter: "Zuul (22348)"
Branch: master

commit 9fd4ea0a39a9bf44dfd3563ea819a57c4eadd9ff
Author: Cornellius Metto <email address hidden>
Date: Wed May 5 17:58:06 2021 +0300

    Add configurability to polling.yaml.

    The ceilometer compute agent uses the default polling.yaml
    from the installed packages without the ability to configure its contents.

    This change adds two configuration options: 'polling-interval' and
    'enable-all-pollsters', borrowing from the implementation in
    charm-ceilometer. We start off with a limited set of meters as before
    and if these are not enough, the user can set 'enable-all-pollsters' to
    'true' to collect all available 'Pollster' metrics as listed in the
    documentation [1].

    Verification:
    I tested this change on a cluster built from the OpenStack base bundle
    and the ceilometer and gnocchi charms. I confirmed that extra metrics
    that originate from the Compute Agent pollster (e.g. disk.device.read.latency)
    are available in gnocchi after setting 'enable-all-pollsters' to true.

    [1] https://docs.openstack.org/ceilometer/latest/admin/telemetry-measurements.html

    Closes-Bug: #1914442
    Change-Id: I21c9a93e7dd91bced9365e44f3e6a5315976c3bb