[openstack-dev] [ceilometer] Pollsters and counters

Jiang, Yunhong yunhong.jiang at intel.com
Thu Jan 17 03:19:09 UTC 2013


This mail is mostly triggered by https://review.openstack.org/#/c/19810/.

Currently we split pollster into small pieces so that we can disable some of them. However, IMHO this is not efficient. For example, we have to create two glance client and access glance twice to fetch the image information (considering a big . We looking in libvirt for 3 times for each domain.

I'd suggest to combine the pollsters for same resource (or source) into one pollster, and use configuration options (no matter pollster-specific or global) to disable some counters. This will be more efficient but possibly we may lost some backward compatibility.

One thing need discussion is, in my original comment to the https://review.openstack.org/#/c/19810/, I suggest to achieve this in multiple publisher. However, after more consideration, I think this is not so good an idea. The reasons are, 

a) as discussed before, the pipeline definition should not be changed in deployment unless expert user. While in deployment, user may want to disable some counter specifically. 

b) If we really have several publishers (3 or 4, or even more), that means user have to excluded the counters in each pipeline definition and that's not so convenient. After all, pipeline is mostly a per-publisher configuration.

Any idea?

BTW, I noticed the glance pollster is one counter per pollster while swift pollster is all counters in one pollster. I think we need keep it same for all resource pollster.

Thanks
--jyh



More information about the OpenStack-dev mailing list