select ... for update breaks with multi-writer galera

[Chris Ricker]

The SQL construct "select ... for update" produces undefined results when used on galera databases with multiple writers

OpenStack uses this construct heavily in both nova and neutron, where it can lead to deadlock errors from the database

See the thread starting at http://lists.openstack.org/pipermail/openstack-dev/2014-May/035264.html for more details

[Chris Ricker (chris-ricker)]

To resolve for icehouse, we'll move to a variant of our current active-active multiple writer model

We'll still set up galera across 3 nodes but will configure the HAProxy VIP in front of galera to direct all writes to a single node, with the other two nodes configured as sequential backups to be written to only when the primary node is unavailable

This will avoid the undefined locking semantics seen when using "select ... for update" with multiple writers, but will still be HA

[Chris Ricker (chris-ricker)]

https://github.com/CiscoSystems/puppet-openstack-ha/pull/19 for the module change

[Chris Ricker (chris-ricker)]

https://review.openstack.org/94738 for the full_ha changes

[Chris Ricker (chris-ricker)]

https://github.com/CiscoSystems/puppet_openstack_builder/pull/84 for the compressed_ha changes

[Chris Ricker (chris-ricker)]

And yes, this is pretty ugly

[Chris Ricker (chris-ricker)]

For Havana:

https://github.com/CiscoSystems/puppet-openstack-ha/pull/20 for proxy configuration change
https://github.com/CiscoSystems/puppet_openstack_builder/pull/100 for composition layer changes