Reliable Pacemaker Galera Resource Agent Script
This document is intended to capture the problems and requirements for
Pacemaker OCF “Resource Agent” (hereafter RA) to improve Galera Cluster
management under Pacemaker Resource Manager
- Reboot Whole cluster (Power outages scenario)
- RA script doesn’t determine the latest Galera GTID version. It always
relies on “primary controller” as a donor. Under some circumstances
Pacemaker cannot assemble Galera cluster.
- Reboot any node from cluster
- Add a new node to active cluster
- Advanced features
- Currently puppet manifests use cs_shadow as a method for cluster
management. It’s not possible to use crm_attribute to store attributes in
configuration as cs_shadow will revert values back
- Write a new RA script for Galera with the following requirements
- RA script allows to bootstrap cluster even when wsrep_cluster_address has
all nodes specified.
- RA script introduces timeout where pacemaker waits for 60-120 seconds until
all nodes specified in CIB became online after reboot or outage.
- After 60-120 seconds RA script must start the process of Primary Component
election which is the node with the latest GTID. This timeout is specified
as node attribute and can be changed by administrator. If all nodes
specified in CIB are UP the election process will be started immediately.
- RA script dertemines Galera GTID state and set it as node attribute. RA
gets GTID from mysqld –wsrep-recover or SQL query
SHOW STATUS LIKE ‚wsrep_local_state_uuid
- The node with the latest GTID will become Galera Primary Controller. It
will be started with empty gcomm:// string. All other nodes will join to
Galera Primary controller to synchronize their state.
- If the node bootstrapped after timeout it will discard its configuration.
This usually happenes when it’s stuck performing fsck.
- When new a node is added to cluster it will join cluster normally.
- Remove cs_shadow
- Remove cs_shadow from manifests to allow to store node attributes
This change doesn’t affect master node upgrade. Openstack upgrade should be
disabled as this change impacts on HA logic.
Other end user impact
Other deployer impact
- Write Galera OCF script
- Perform all set of destructive tests
All set of destructive tests: Reboot single node, reboot whole cluster, add a
new node from Fuel UI
The documentation should indicate how to increase/decrease Bootstrap timeout.