We've been trying to get gnocchi working us, and have been hitting some
performance bottlenecks. It seems that the 'host' field for the instance
resource-type flaps between host id and host names. This makes the gnocchi
dispatcher calls an update per resource which is a HTTP PATCH call. The
large amount of HTTP calls causes gnocchi to be unable to keep up with
ceilometer, essentially making it unworkable for us in production.
I've tried removing this update code and can get ~5x the performance. So I
guess the questions are,
1) why is the host information flapping? Are there difference services
generating this data?
2) If gnocchi data can be viewable by the user, should we only have host id
in the resource? (similar to what nova shows to user)
3) Anyone running gnocchi in prod, can you let us know how is it going?
Any pointers to track down this issue will be appreciated! There is a
bug open for this issue but I hope this email can reach more people.
FYI, we are running Mitaka ceilometer and gnocchi, but we have patched
ceilometer to use the Newton gnocchi dispatcher to help with performance.
Core Services, NeCTAR Research Cloud,
The University of Melbourne
OpenStack Development Mailing List (not for usage questions)