306-300 LPIC-3 High Availability and Storage Clusters (Exam 306) Practice Questions

In Pacemaker, an Order constraint (e.g., start ServiceA then start ServiceB) implies that if ServiceA is stopped, ServiceB must be stopped first. It handles both the startup sequence and the shutdown dependency.

The gluster volume add-brick command extends the volume. Syntax: gluster volume add-brick . This adds capacity to a distributed volume.

The -s flag defines the schedule using regex syntax T/MM/DD/d/HH. L stands for Long test. ../../7/02 means any month, any day of month, day of week 7 (Sunday), at 02:00 hours.

The MDS is required only for CephFS (the filesystem). It stores metadata (inodes, directory structure) so clients can mount the filesystem POSIX-style. It is not needed for RBD or Object Gateway (RGW).

Corosync requires the exact same authkey file on all nodes to encrypt and authenticate traffic. If the file is missing or unreadable (wrong permissions, e.g., not 400 root:root), the node cannot join the cluster ring.

DRBD is managed in Pacemaker as a promotable clone of the ocf:linbit:drbd agent (older documentation calls this multi-state or master/slave; "promiscuous clone" is not a Pacemaker term). A colocation constraint of the database with the Promoted (Master) role of that clone, e.g. pcs constraint colocation add Database with Promoted DRBD-clone or in crmsh colocation db_on_drbd inf: Database ms_drbd:Master, allows the database to run only on the node where DRBD is currently Primary. Colocation affects placement only, so it is normally paired with an order constraint (promote DRBD-clone then start Database) to make the database start after the promotion. A location constraint cannot point at another resource's role, promoting DRBD after the database has started is the wrong order, and symmetric-cluster=false only turns the cluster into an opt-in cluster.

HAProxy is monitored with a vrrp_script block (for example script "killall -0 haproxy") that the vrrp_instance references with track_script. After fall consecutive failures, an instance tracking a script with the default weight 0 transitions to FAULT state and releases the VIP; with a negative weight the value is subtracted from its priority, so a BACKUP with a higher priority wins the election and takes over the VIP. lb_algo selects the LVS scheduler, notify_master only runs a script after a transition to MASTER, and persistence_timeout sets LVS client persistence.

Corosync 3.x corosync.conf(5): "The default is knet. The transport type can also be set to udpu or udp. Only knet allows crypto or multiple interfaces per node." The crypto_cipher and crypto_hash entries add that "Encrypted transmission is only supported for the knet transport", and the maximum number of interface directives is 8 for knet but 1 for other transports. The Corosync 3.0 release notes confirm that UDP/UDPU "supports only single ring (RRP is gone in favor of Knet) and doesn't support encryption" and list "Higher throughput and lower latency" among the knet features, so moving to udp is not a performance optimization. The token default (3000 ms in current 3.x releases) does not depend on the transport, and rrp_mode exists only in Corosync 2.x.

First stop the cluster stack on the stale node B (pcs cluster stop without a node name stops the local node). Node A then holds only one of the two votes and, unless two_node is enabled, is inquorate, so lower the expected votes on the survivor: pcs quorum expected-votes 1 changes expected votes in the live cluster only (the corosync.conf value returns on reload). Red Hat documents this under "Running inquorate clusters": "This allows the cluster to continue operation when it does not have quorum" (the corosync-level equivalent is corosync-quorumtool -e 1); use it with extreme caution and only when the other node is really stopped. After verifying node A (pcs status, corosync-quorumtool -s), start node B so it rejoins the running cluster. corosync-cmapctl -s needs a key name, a type and a value, and no runtime.votequorum.expected_votes key is documented. Stopping all nodes and starting B would bring up the stale node, crm_node --force --remove plus pcs cluster setup rebuilds the cluster membership, and pcs resource cleanup does not address quorum.

GFS2 relies on the Distributed Lock Manager (DLM) to coordinate file locking across the cluster nodes. Without DLM, nodes would overwrite each other's data on the shared block device.