Purpose
Ceph is tolerant of bad hardware right up until it is not. A dying disk is absorbed as slow ops for weeks; clock skew is absorbed until it takes MON quorum with it; a single power feed is absorbed until the feed drops a whole failure domain at once. This checklist is run before a host joins the cluster, while the defects are still cheap to fix.
Order of work
The list runs in three passes. The device pass is storage-class and
disk-health: every OSD disk is identified by class, because the class is
what CRUSH rules select on, and screened with SMART before it carries data.
The host pass is entropy, power and time - the three host-level
properties that fail cluster-wide rather than per-OSD, with time synchronisation
the one that most often takes a cluster down. The placement pass is
fail-domain, which ties the physical rack position to the CRUSH bucket the
host will be added under.
Before you start
Bring the rack elevation, the PDU and feed assignment for each host, the disk model and firmware inventory, and the NTP or chrony source the site uses.
Verification
The host is ready when every critical item is checked. Disks that fail SMART screening are replaced before deployment, not after they are carrying replicas. Mark an item N/A rather than leaving it unchecked when it does not apply to this cluster.
Escalation
Marginal SMART results are a purchasing conversation, not an operational one. Route them to whoever owns the hardware contract.