Node Taints and Affinities
Scheduling and placement controls for Foundation workloads.
Component Category
Cluster operations / scheduling and placement
Component Description
Node taints, tolerations, and affinities are Kubernetes scheduling controls used to steer workloads onto the right nodes.
Why It Is Used
In BullSequana AI Foundation, these controls help reserve specialized hardware for the workloads that need it most. They are especially important for GPU inference, high-storage services, and other components that should not compete with general-purpose workloads for the wrong node types.
Learn More
Platform placement
The platform uses these node labels and taints:
| Workload | Node label | Taint tolerated | Strength |
|---|---|---|---|
| PostgreSQL (CloudNativePG) | workload=high-memory | dedicated=high-memory | Preferred; required with [common] enforce_affinity |
| Ceph OSDs | workload=storage | dedicated=storage | Preferred; required with [common] enforce_affinity |
| vLLM and Triton GPU runtime profiles | nvidia.com/gpu.present=true | nvidia.com/gpu | Required |
| Time-sliced and fractional GPU profiles | nvidia.com/device-plugin.config=timeslicing or fractioning | nvidia.com/gpu | Required |
| Speaches in CUDA mode | workload=gpu | nvidia.com/gpu | Preferred |
| Agent sandbox pods and the Kata runtime | workload=sandbox | workload=sandbox (NoSchedule) | Required |
[common] enforce_affinity (COMMON_ENFORCE_AFFINITY, false by default) turns the database and storage preferences into required node affinity and their tolerations from PreferNoSchedule into NoSchedule. It does not change GPU placement. See GPU node configuration and Enable GPU for AI components.
Placement is not the only GPU control. In tenant namespaces, every GPU pod must be scheduled by KAI Scheduler into one of the tenant's GPU queues, which bound how many GPUs the tenant uses. See Tenant GPU scheduling.
Interacts With
KServe,vLLM, andKAI Scheduler, which manage GPU-aware model workloads.Kata Containers, whose runtime runs only on the sandbox node pool.Rook CephandPostgreSQL, when storage-heavy workloads should be steered to specific nodes.PrometheusandGrafana, which help operators verify workload placement and resource behavior after scheduling rules are applied.