Open Stack

> We would need to make the change everywhere instance-related events are
> generated in nova so that the metadata for those events is always the same. I

I think currently the notification_to_metadata() in both ceilometer/compute/notifications.py and ceilometer/plugin.py fetch only some field in the notification, depends on metadata_keys definition. 

IMHO it should be ok to use extra usage information only for "compute.instance.delete.start", which is a special case. And we can create a notification class in ceilometer specifically for "compute.instance.delete.start" event , fetching the metering information from extra usage info, put the corresponding metering information, either through the pollster, or publish it directly.

> think it would be better to change info_from_instance() in nova/notifications.py
> so that the callers don't have to be updated, but I don't know if that module

As that code in compute node also, so it should be ok to retrieve such information, but that means such information will be included in all instance event, including the usage audit, not sure if overkill.

--jyh

> would have the information it needs to collect the appropriate stats from the
> hypervisor. For events generated when the instance doesn't exist we would
> want to substitute suitable constants (probably zero).