Test report: EKS qualification plan

Test report

This page is for contributors and records a particular design or test snapshot. For current user guidance, see capabilities and limitations and the operation guides. Historical results do not establish current release or AWS validation.

Run only in an explicitly selected disposable AWS environment. Record the operator, launcher and runtime image digests, fork commit, node architectures, Kubernetes and CSI versions, StorageClass, region, zones and IAM roles.

1. Verify private networking and runtime-only bucket credentials. Start a homogeneous fork fleet and confirm schema 1 and strict removal capability for every member. Verify stable per-Pod peer DNS resolves while Pods are not Ready; predecessor recovery must not depend on old Pod IPs.
2. Generate uniquely numbered acknowledged durable writes while scaling an Ordered Bucket fleet and a PersistentFleet down and back up. Verify every acknowledgment after membership changes and survivor restart.
3. Exercise strict shutdown with follower-only obligations, the last follower, delayed S3 operations and transient control-plane outages. Failed or unknown completion must preserve the exact current operation and disk.
4. Crash the operator before issuance, after runtime completion, after proof capture and after each workload/PVC effect. Reconcile without duplicate or stale effects. Replace named resources to verify UID preconditions.
5. Exercise the selected disk policy with real EBS CSI. Verify PVC protection, exact PV/handle binding, attachment cleanup and the actual EBS outcome. A missing PVC alone does not prove physical volume deletion.
6. Exercise coordinated restart, image change and fleet deletion. Check explicit downtime permission, fresh disks, permanent bucket reservation and every acknowledged write after recovery.
7. Lose a launcher before proof capture and try unsafe cross-host disk reuse. Verify both remain blocked; do not use force detach or finalizer removal.
8. Partition S3 across node-lease expiry after fresh acknowledged writes. Verify fail-stop and unchanged disks, then explicit same-host administrative recovery with changed Pod IPs and stable peer DNS. Preserve logs before Pod replacement and verify every acknowledgment. Deliberately delay one retained witness: the corrected runtime must remain unready while that witness is unreachable, then recover every acknowledgment when it returns. Keep it unavailable through retry exhaustion and verify startup fails without sealing the predecessor or recording loss; restart both retained disks and verify recovery again.

Archive commands, timestamps, artifacts and failure results for the exact run. A local MinIO or Kind pass does not close any unexecuted AWS gate.