Open Stack

Hi all,
While doing scale tests on our infrastructure, we observed some increase in the response times of our keystone servers.
After further investigation we observed that we have a hot key in our cache configuration (this means than all keystone servers are checking this key quite frequently)
We are using a pool of memcache servers for hosting the cache and the solution does not seem ideal at this scale.

The key turns out to be the revocation tree, that is evaluated in every token validation.  If the revocation tree object stored is big enough it can kill the network connectivity
on the cache server affecting the whole infrastructure as the identity servers needs to check the key before validating a token.

On our scale tests after the cleanup, we have 250 requests/second for an object of 500KB that is a throughput of 1Gbit/sec that saturate the network link of the cache server.

We are checking other strategies like redis or mongo, but we would like to know if you have already seen this before? If so what you have done?

Kind regards,
Jose

Jose Castro Leon
CERN IT-CM-RPS                   tel:    +41.22.76.74272
                                                mob: +41.75.41.19222
                                                fax:    +41.22.76.67955
Office: 31-1-026                  CH-1211      Geneve 23
email: jose.castro.leon at cern.ch<mailto:jose.castro.leon at cern.ch>

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.openstack.org/pipermail/openstack-operators/attachments/20160621/a73fd4bd/attachment.html>