Cheat sheet & self-check

Pick an answer to see if you got it, and why.

1. Q1. A PVC is Bound, but the pod is stuck in ContainerCreating with 'FailedAttachVolume'. Which step failed?

   1. AProvisioning
   2. BAttach: the controller couldn't attach the volume to the node (see the VolumeAttachment and the attacher logs)
   3. CDNS
   4. DImage pull
   Show answer

   B. Bound means the volume exists. Attach connects it to the node; mount then puts it into the pod. The event name tells you which step.

   From lesson 01 · CSI architecture

2. Q2. What runs as a DaemonSet in a CSI driver?

   1. AThe provisioner
   2. BThe node plugin, which stages and mounts volumes on each node
   3. Cetcd
   4. DThe snapshot controller
   Show answer

   B. Mounting happens on the node where the pod runs, so every node needs the node plugin (plus the node-driver-registrar sidecar).

   From lesson 01 · CSI architecture

3. Q3. What does allowVolumeExpansion: true on a StorageClass enable?

   1. AAutomatic growth
   2. BGrowing a PVC by editing its requested size (if the driver supports it)
   3. CShrinking volumes
   4. DFaster IOPS
   Show answer

   B. You edit spec.resources.requests.storage upwards; the resizer sidecar and the node plugin grow the volume and filesystem. Shrinking isn't supported.

   From lesson 01 · CSI architecture

4. Q4. With the default 3 replicas, what happens when a node holding one replica dies?

   1. AThe volume is lost
   2. BThe volume keeps working from the remaining replicas and Longhorn rebuilds a new replica on another node
   3. CThe pod is deleted
   4. DData becomes read-only forever
   Show answer

   B. Longhorn synchronously writes to every replica. Losing one leaves the volume 'degraded' until a rebuild restores full redundancy.

   From lesson 02 · Longhorn in practice

5. Q5. Is a replica a backup?

   1. AYes
   2. BNo: replicas protect against node or disk failure, but a deletion or corruption is replicated too; backups (to S3/NFS) protect against that
   3. COnly for RWX
   4. DOnly with 5 replicas
   Show answer

   B. Replication is availability. Backups are recovery. You need both.

   From lesson 02 · Longhorn in practice

6. Q6. How does Longhorn provide ReadWriteMany (RWX) volumes?

   1. AIt can't
   2. BThrough a share-manager that exports the block volume over NFS to all pods
   3. CBy copying files
   4. DThrough iSCSI multipath
   Show answer

   B. RWX volumes are served by an NFS share-manager pod, so nodes need an NFS client installed.

   From lesson 02 · Longhorn in practice

7. Q7. What do Ceph MONs do?

   1. AStore the data
   2. BKeep the authoritative cluster map and require a quorum (like etcd for Ceph)
   3. CServe S3
   4. DMount volumes
   Show answer

   B. Monitors hold the cluster maps. Run 3 (or 5) so a majority survives failures. OSDs store the data.

   From lesson 03 · Rook-Ceph

8. Q8. Which Ceph interface gives Kubernetes ReadWriteMany shared filesystems?

   1. ARBD
   2. BCephFS
   3. CRGW
   4. DMGR
   Show answer

   B. RBD is block (typically RWO), CephFS is a POSIX filesystem (RWX), and RGW is S3/Swift object storage.

   From lesson 03 · Rook-Ceph

9. Q9. With a pool replicated 3× and failure domain 'host', what does CRUSH guarantee?

   1. AThree copies on the same host
   2. BThe three copies are placed on three different hosts
   3. CCopies only on SSDs
   4. DNothing
   Show answer

   B. CRUSH rules spread replicas across the chosen failure domain (host, rack, zone), so one host failure never loses all copies.

   From lesson 03 · Rook-Ceph

10. Q10. Is a CSI snapshot, stored on the same storage system as the volume, a complete backup?

    1. AYes
    2. BNo: if the storage system is lost, its snapshots are lost too; copy data off the system for a real backup
    3. COnly for RWX
    4. DOnly with Ceph
    Show answer

    B. Snapshots are fast and great for quick rollback, but they share the fate of the storage. Backups must live elsewhere.

    From lesson 04 · Snapshots & backups

11. Q11. What does 'application-consistent' mean for a database backup?

    1. AThe backup is encrypted
    2. BThe database was flushed or paused (or used its own backup tool) so the copy is a clean, restorable state, not a crash image
    3. CThe backup is compressed
    4. DIt's taken at night
    Show answer

    B. A crash-consistent snapshot is like pulling the power: databases usually recover, but not always cleanly. Hooks or native tools give consistency.

    From lesson 04 · Snapshots & backups

12. Q12. What's the safest way to test a Velero restore?

    1. ARestore over production
    2. BRestore into a new namespace (namespace mapping) or another cluster, then verify the app and its data
    3. CDelete production first
    4. DDon't test restores
    Show answer

    B. Namespace mappings let you prove restores work without touching running workloads.

    From lesson 04 · Snapshots & backups

13. Q13. Why use --direct=1 in fio tests?

    1. AIt's faster
    2. BIt bypasses the page cache, so you measure the storage, not RAM
    3. CIt enables encryption
    4. DIt's required by Kubernetes
    Show answer

    B. Without direct I/O, reads may come from memory and writes may only reach the cache, which gives flattering, meaningless numbers.

    From lesson 05 · Storage performance

14. Q14. Which metric matters most for etcd and database commit performance?

    1. ASequential throughput
    2. BLatency of small synchronous writes (fsync/fdatasync), especially the high percentiles
    3. CTotal capacity
    4. DRead IOPS
    Show answer

    B. Every commit waits for a flush to stable storage. Slow or spiky fsync latency directly slows the whole system.

    From lesson 05 · Storage performance

15. Q15. Why does a 3-replica distributed volume often have higher write latency than a local disk?

    1. AIt uses more CPU
    2. BEach write must travel over the network and be acknowledged by all replicas before completing
    3. CIt compresses data
    4. DIt's a bug
    Show answer

    B. Synchronous replication trades latency for durability. Fast networks and disks reduce the cost; they can't remove it.

    From lesson 05 · Storage performance

16. Q16. What does a StatefulSet guarantee that a Deployment doesn't?

    1. AMore replicas
    2. BStable pod names (db-0, db-1), stable per-pod storage, and ordered creation and updates
    3. CFaster startup
    4. DAutomatic backups
    Show answer

    B. Each replica keeps its identity and its own PVC across restarts and rescheduling, which clustered databases rely on.

    From lesson 06 · Stateful patterns

17. Q17. Why use a database operator instead of a plain StatefulSet?

    1. AOperators are required
    2. BThe operator encodes the database's operational knowledge: replication setup, failover, backups, upgrades
    3. CStatefulSets can't use PVCs
    4. DOperators are faster
    Show answer

    B. A StatefulSet keeps pods and disks stable, but knows nothing about primaries, replicas or WAL archiving. Operators do.

    From lesson 06 · Stateful patterns

18. Q18. When is a managed cloud database usually the better choice?

    1. ANever
    2. BWhen the team lacks database operations expertise, or needs features like cross-region replicas and point-in-time restore without building them
    3. COnly for tiny databases
    4. DWhen using Kubernetes
    Show answer

    B. Running databases well is a skill. Managed services trade control and cost for less operational burden.

    From lesson 06 · Stateful patterns

19. Q19. Why alert on predicted fill time rather than only '90% used'?

    1. AIt's fancier
    2. BA fast-growing volume can go from 70% to full in minutes, while a slow one can sit at 90% safely for months
    3. CPercentages are inaccurate
    4. DPrometheus can't compute percentages
    Show answer

    B. predict_linear gives the time left, which is what the on-call engineer actually needs to act on.

    From lesson 07 · Production patterns

20. Q20. What's the risk of reclaimPolicy: Delete on a StorageClass used for important data?

    1. ASlower volumes
    2. BDeleting the PVC (even by mistake, e.g. deleting a namespace) deletes the underlying disk and its data
    3. CVolumes can't be expanded
    4. DSnapshots stop working
    Show answer

    B. Use Retain (or backups plus deletion protection) for critical data, so an accidental namespace deletion isn't a data-loss event.

    From lesson 07 · Production patterns

21. Q21. Before draining a node that hosts replicas of a replicated storage system, what should you confirm?

    1. ANothing
    2. BThat all volumes are healthy (fully replicated), so taking one node's replicas offline can't leave any volume without a copy
    3. CThat the node has free CPU
    4. DThat DNS works
    Show answer

    B. Draining a node while another replica is already degraded can take a volume offline. Check health first; use the storage system's drain policy.

    From lesson 07 · Production patterns