The ‘target-role’ meta attribute controls the desired state of a resource. If set to ‘Stopped’, the cluster will not attempt to start the resource automatically, which matches the described behavior where the cluster forms but resources don’t start. This is a common oversight when testing or manually stopping resources.

The ‘after-sb-2pri disconnect;’ policy tells DRBD to simply disconnect and remain in split-brain when both nodes think they are primary (a two-primary split-brain). For automatic recovery, this should typically be set to a policy like ‘call-pri-lost-after-sb’ or similar, which would allow one node to demote itself based on the chosen resolution algorithm.

STONITH (Shoot The Other Node In The Head) is essential for data integrity. It ensures a node believed to be dead or malfunctioning is forcibly powered off or isolated before allowing another node to start the resource and access the shared data (like a DRBD disk). This prevents the scenario where two nodes could simultaneously access the same storage, which would corrupt data. A promotion score is relevant for multi-state resources.

GFS2 requires one journal per node that will mount the filesystem. If nodes are sharing a single journal, they will contend for it, causing excessive cache invalidations (explaining the high miss rates) and slow performance. Adding a dedicated journal for each active node with ‘gfs2_jadd’ is a standard performance tuning step for this symptom.

A resource group in Pacemaker is a set of resources that are located together, started sequentially, and stopped in the reverse order. The output shows fs_data, vip_app, and svc_app in ‘app-group’. This means they must all run on the same node (node1), fs_data starts first, then vip_app, then svc_app. Stopping would happen in the reverse order: svc_app, then vip_app, then fs_data.

The ‘token’ timeout defines how long a node waits for a token (the right to broadcast) before assuming the token is lost and starting a new membership consensus. On a fast, reliable network, this value can often be safely reduced from its default. This allows the cluster to detect failures and reconfigure more quickly, leading to faster failover. The other options either don’t exist (ringnumber 0), would slow things down (increasing timeouts), or relate to ring redundancy, not speed.

For multi-state resources, the ‘promotion-score’ meta attribute is key. A positive promotion-score (like 100) is added to a node’s total score when calculating where to place the *Master* role. The ‘promoted-on’ is not a standard attribute; location constraints or node attributes are used to pin the master role to a specific node. Option 1 correctly uses promotion-score. Option 0 uses a non-existent ‘target-role’ value. Option 2 uses master-max (correct) but resource-stickiness applies to the base resource, not specifically the master role. Option 4 uses a non-existent ‘preferred-node’ attribute.

In a cluster environment, LVM must use cluster-aware locking. This is controlled by the ‘locking_type’ parameter in lvm.conf. Setting ‘locking_type = 3’ enables clustered locking, which is required for cLVM to coordinate access to the volume group across multiple nodes. The other options are incorrect: ‘use_lvmetad’ is a caching daemon setting, ‘locking_library’ is not the standard parameter, and ‘volume_list’ is used to filter which VGs are auto-activated on boot, not for cluster coordination.

The `crm node standby ` command (or `pcs node standby`) puts a node into standby mode. This tells the cluster to move all resources off the node and to not assign any new resources to it. The node remains online and part of the cluster quorum. This is the standard method for planned maintenance. ‘fence’ would invoke STONITH. ‘resource migrate-all’ isn’t a standard command, and ‘resource ban’ is a lower-level operation not typically used for this simple, common task.

The ‘target-role’ meta attribute controls the desired state of a resource. If set to ‘Stopped’, the cluster will not attempt to start the resource automatically, which matches the described behavior where the cluster forms but resources don’t start. This is a common oversight when testing or manually stopping resources.0