v1.12.0

What's Changed

Compute

• Ubuntu NVIDIA GPU node images with automatic GPU-aware AMI selection
• GPU passthrough and MIG partition state are now visible across the EC2, EKS, and ECS console
• EC2 Launch Templates: capture reusable launch configuration and launch instances straight from a template, in both the API and the console

Kubernetes (EKS)

• Run GPU workloads on EKS — NVIDIA GPUs are exposed to pods through the device plugin
• Control-plane disaster recovery: snapshot etcd and restore an entire cluster from a snapshot
• Multi-node clusters now create end-to-end reliably
• Cluster teardown converges cleanly and garbage-collects orphaned network topology
• Nodes carry the eks.amazonaws.com/nodegroup label, so nodegroup-targeted scheduling works
• Hardened nodegroup readiness, NLB naming, launch retries, and GPU control-plane tolerations

Containers (ECS)

• GPU passthrough to ECS task containers
• Tags API for ECS resources

Identity (IAM)

• Edit inline policies for IAM users, roles, and groups directly in the console
• GetAccountSummary is now supported

Security

• Presigned S3 URLs: generate time-limited GET/PUT links authenticated via SigV4 query-string auth

Networking

• Routed-NAT hardened with a per-VPC capacity cap
• Reassociating an Elastic IP flushes stale network state, so a reused address routes correctly
• Fixed gateway next-hop reachability via a static OVN MAC binding
• spx get nodes surfaces OVN NB/SB Raft roles

Storage (EBS / block)

• Faster volume snapshots and uploads through bounded parallel chunk upload with coalesced checkpoints
• Eliminated write amplification on the live checkpoint path

Reliability

• Re-running spx admin init --force now preserves existing keys, credentials, and the CA, so re-initialising to apply a fix no longer orphans encrypted volumes, S3 data, or joined-node trust — only config and the server certificate are refreshed
• Stopped VMs resume cleanly after a host reboot
• EBS unmount and stopped-instance updates are now idempotent
• GPU admission is gated on available CPU and memory, and leaked GPU slots are reclaimed when an instance crashes
• NATS KV buckets replicate at cluster size for high availability, instead of a single replica
• Load balancers wedged in provisioning are reaped, and IMDS bootstrap is hardened
• Guest-write backpressure bounds host memory use for block storage

Observability

• Control-plane, S3, and block-storage logs now stream to Elasticsearch via an OTLP bridge
• Distributed traces are emitted from the QMP collector and the web console
• S3 request metrics include transfer bytes and real S3 error codes, with new block-storage IO metrics
• AWS error messages and schedulable node capacity are surfaced in the console