Repository navigation
Supported versions
Kubernetes 1.35–1.37
Slurm 25.11 or later, with REST data parser v0.0.44. The Helm chart defaults to Slurm 26.05 on Ubuntu 26.04.
Enhancements
Slurm controller HA. Deploy a primary slurmctld and backups behind a virtual IP for failover. Set controller.ha.enabled: true and supply an existing ReadWriteMany volume for shared state.
Large NodeSets. Batched Kubernetes writes, indexed lookups, and cached Slurm node state reduce API traffic. The operator also fixes an out-of-memory failure at high replica counts.
NodeSet draining. Health checks and automation can request a drain with nodeset.slinky.slurm.net/pod-cordon-requested. During scale-in, the operator can prioritize a Pod with an earlier deadline.
Node feature discovery. Slurm jobs can select workers by Kubernetes-discovered attributes, such as network switch IDs, using --constraint=k8s/<feature>.
Accelerator deployments. The chart adds GKE A3 Ultra alongside A3 Mega. Guides cover Cloud TPU v5e with JAX, NVIDIA GPU DRA for IMEX channels on GB200 and GB300 NVL72 systems, and SR-IOV DRA for network virtual functions.
Breaking changes and upgrades
Pod binding and topology. The pods/binding webhook is off by default. Slurm receives topology for new NodeSet Pods during reconciliation. High-throughput sites that need topology at bind time should set webhook.podsBinding: true.
Metrics scraping. --metrics-secure now enforces HTTPS and requires credentials authorized to read /metrics. Update existing scrapers if the flag is enabled. Webhook metrics are on by default.
REST API without accounting. When accounting is disabled, slurmrestd loads only openapi/slurmctld. Clients must stop calling /slurmdb/ endpoints on those clusters or enable accounting.
Upgrading from 1.2 Upgrade slurm-operator-crds before slurm-operator, or upgrade both together with crds.enabled=true. Upgrade guide
Deprecations and removals
taintKubeNodes. The operator no longer applies taints through this deprecated option. Apply Kubernetes taints directly and configure matching NodeSet tolerations and node selectors.
NodeSet settings. Move scalingMode, ordinalPadding, and pinToNode under scaleStrategy. Replace pruneSlurmNodeRecords: NodeNotFound with Auto, which also handles invalid Slurm registrations.
SSH and SSSD settings. Migrate NodeSet and LoginSet extraSshdConfig to a mounted ConfigMap and sssd.conf to a mounted Secret. Helm migration guide