Neutron has protection against these now as of liberty with API level
retry
operations so this shouldn't be a problem for neutron any more.
Jay Pipes wrote an awesome post about the problem itself and a solution
earlier this year:
http://www.joinfu.com/2015/01/understanding-reservations-concurrency-locking-in-nova/
Saying that though, we’re running the same setup as others have
described - which is that we have a multi-writer mode Galera cluster
with haproxy in front ensuring that only one node is actually written to
- and it’s worked very well for us thus far.
—
-Nick
On Dec 7, 2015 4:06 PM, "Matteo Panella" <matteo.pane...@cnaf.infn.it>
wrote:
On 2015-12-07 22:45, James Dempsey wrote:
+1 for designating one node as primary. This helped us reduce some
deadlocks that we were seeing when balancing sessions between DB
hosts.
Keystone most likely won't be affected by writeset certification
failures,
but other services (especially Neutron) are going to be hit by one
sooner
or later.
Unfortunately, Galera doesn't take very kindly "SELECT ... FOR
UPDATE", so
you can't really do proper load balancing (aside from designating a
different node
as master in multiple listen stanzas, each one for a different set of
services).
--
DataCentred Limited registered in England and Wales no. 05611763
_______________________________________________
OpenStack-operators mailing list
OpenStack-operators@lists.openstack.org
http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-operators
