Open Stack

[Joe]: For reliability purpose, I suggest that the keystone client should provide a fail-safe design: primary KeyStone server, the second KeyStone server (or even the third KeySont server) . If the primary KeyStone server is out of service, then the KeyStone client will try the second KeyStone server. Different KeyStone client may be configured with different primary KeyStone server and the second KeyStone server.

[Adam]: Makes sense, but that can be handled outside of Keystone using HA and Heartbear and awhole slew of technologies.  Each Keystone server can validate each other's tokens.
For cross-site KeyStone HA, the backend of HA can leverage MySQL Galera cluster for multisite database synchronous replication to provide high availability, but for the KeyStone front-end the API server, it's web service and accessed through the endpoint address ( name, or domain name, or ip address ) , like http://.... or ip address.

AFAIK, the HA for web service will usually be done through DNS based geo-load balancer in multi-site scenario. The shortcoming for this HA is that the fault recovery ( forward request to the health web service) will take longer time, it's up to the configuration in the DNS system. The other way is to put a load balancer like LVS ahead of KeyStone web services in multi-site. Then either the LVS is put in one site(so that KeyStone client only configured with one IP address based endpoint item, but LVS cross-site HA is lack), or in multisite site, and register the multi-LVS's IP to the DNS or Name server(so that KeyStone client only configured with one Domain name or name based endpoint item, same issue just mentioned).

Therefore, I still think that keystone client with a fail-safe design( primary KeyStone server, the second KeyStone server ) will be a "very high gain but low invest" multisite high availability solution. Just like MySQL itself, we know there is some outbound high availability solution (for example, PaceMaker+ColoSync+DRDB), but also there is  Galera like inbound cluster ware.

Best Regards
Chaoyi Huang ( Joe Huang )


From: Adam Young [mailto:ayoung at redhat.com]
Sent: Tuesday, March 17, 2015 10:00 PM
To: openstack-dev at lists.openstack.org
Subject: Re: [openstack-dev] [opnfv-tech-discuss] [Keystone][Multisite] Huge token size

On 03/17/2015 02:51 AM, joehuang wrote:
It's not reality to deploy KeyStone service ( including backend store ) in each site if the number, for example, is more than 10.  The reason is that the stored data including data related to revocation need to be replicated to all sites in synchronization manner. Otherwise, the API server might attempt to use the token before it's able to be validated in the target site.

Replicating revocati9on data across 10 sites will be tricky, but far better than replicating all of the token data.  Revocations should be relatively rare.

When Fernet token is used in multisite scenario, each API request will ask for token validation from KeyStone. The cloud will be out of service if KeyStone stop working, therefore KeyStone service need to run in several sites.

There will be multiple Keystone servers, so each should talk to their local instance.

For reliability purpose, I suggest that the keystone client should provide a fail-safe design: primary KeyStone server, the second KeyStone server (or even the third KeySont server) . If the primary KeyStone server is out of service, then the KeyStone client will try the second KeyStone server. Different KeyStone client may be configured with different primary KeyStone server and the second KeyStone server.

Makes sense, but that can be handled outside of Keystone using HA and Heartbear and awhole slew of technologies.  Each Keystone server can validate each other's tokens.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.openstack.org/pipermail/openstack-dev/attachments/20150319/2df14e3f/attachment.html>