I85PREP
⌘ K
LPIC-3 High Availability and Storage Clusters study materials
LPIC-3 High Availability and Storage Clusters study resources

LPIC-3 High Availability and Storage Clusters: High Availability Distributed Storage

LPIC-3 High Availability and Storage Clusters · Work through your exam, one topic at a time.

All guide sections

High Availability Distributed Storage

High Availability Distributed Storage

This domain tests storage that survives node loss without shared LUNs: Ceph's RADOS components and GlusterFS volumes. The exam expects you to reason about replica counts, quorum, and recovery rather than memorize only commands.

Core rules

  • Ceph monitors form quorum; use an odd number, normally three. OSDs store data, MDS serves CephFS, and MGR exposes metrics. CRUSH computes placement, so no central index is consulted.
  • Pools define size (copies) and min_size (write quorum). size=3, min_size=2 tolerates one OSD loss and keeps writes available; min_size=1 permits writes with no redundancy. PG count affects balance; autoscaling can manage it.
  • Use ceph -s, ceph osd tree, and ceph health detail. Before planned OSD downtime, ceph osd set noout; unset it afterward. ceph osd out triggers rebalance.
  • GlusterFS bricks are export directories. Volumes can be distributed, replicated, dispersed, or use an arbiter. Replica 3 tolerates one brick loss; replica 2 without an arbiter risks split-brain. gluster volume heal <vol> info shows pending repairs; the self-heal daemon repairs files.
  • Both systems need time synchronization and failure-domain-aware placement.

Worked example

A three-node Ceph cluster has a pool with size=3 and min_size=2. One node fails: each placement group still has two copies, writes continue, and ceph -s reports degraded while recovery proceeds. If min_size=3, writes block until the third copy returns; if min_size=1, writes continue but a second failure can lose data. For planned maintenance, set noout so OSDs are not marked out and rebalanced immediately; unset it when the node returns.

Traps

Two monitors do not provide quorum tolerance. min_size=1 defeats replication. Ignoring PG counts causes imbalance. Gluster replica 2 can split-brain without an arbiter. Healing is not instant, and force at volume creation does not fix a bad layout.