Hostnames and resolution on Ceph hosts

Why this matters in production

The orchestrator identifies hosts by name and matches placement against that name. A mismatch between what the host reports and what the cluster knows produces a host that appears healthy and never receives daemons — a symptom with no obvious cause.

Setting the hostname

hostnamectl set-hostname ceph-osd-01.example.com
hostnamectl status
hostname
hostname -f
hostname -s

Decide whether the cluster uses short names or FQDNs and apply it consistently. Mixing them across a fleet is what produces the mismatches.

cephadm bootstrap --mon-ip 10.20.0.10 --allow-fqdn-hostname
ceph orch host add ceph-osd-01.example.com 10.20.0.21

--allow-fqdn-hostname is needed if hosts report FQDNs, and the names used in host add must match what the hosts report.

Resolution

# forward
getent hosts ceph-osd-01
dig +short ceph-osd-01.example.com

# reverse
dig +short -x 10.20.0.21

# from every host to every other
for h in ceph-mon-01 ceph-osd-01 ceph-osd-02; do
  getent hosts "$h" || echo "FAIL forward $h"
done

Both directions, from every host. /etc/hosts entries are an acceptable fallback where DNS is unavailable, provided they are consistent across the fleet — inconsistent hosts files are worse than no entries at all.

Diagnosing a mismatch

ceph orch host ls
# HOST                     ADDR         LABELS  STATUS
# ceph-osd-01              10.20.0.21   osd

ssh ceph-osd-01 hostname
# ceph-osd-01.example.com     ← mismatch

The cluster knows ceph-osd-01; the host reports the FQDN. Placement matches on the cluster’s name and the daemon deployment reports the host’s, so they never align.

ceph orch host rm ceph-osd-01
ceph orch host add ceph-osd-01.example.com 10.20.0.21 --labels osd

Removing and re-adding with the correct name is the fix. Do it before OSDs are deployed if possible.

Verifying across the fleet

for h in $(ceph orch host ls --format json | jq -r '.[].hostname'); do
  reported=$(ssh "$h" hostname 2>/dev/null || echo UNREACHABLE)
  [ "$h" = "$reported" ] || printf '%-30s cluster=%s reported=%s\n' "$h" "$h" "$reported"
done

Anything printed is a mismatch.

Quiz

Production discipline

Establish a single naming convention at bootstrap and verify every host reports what the cluster recorded; the mismatch is silent and its symptom — a healthy host with no daemons — points nowhere obvious. Run the comparison loop after any host provisioning change.

Cross-course references

• Kubernetes: node name mismatches produce the same silent scheduling failure
• Linux: consistent naming across a fleet is a prerequisite for any orchestration tool