Mellanox / network-operator / 23810671614

push

github

feat: add NicNodePolicy CRD for per-node-group DOCA driver management (#2279)

Add comprehensive support for heterogeneous clusters where different
node
groups require different DOCA driver versions via NicNodePolicy (NNP).

Webhook validation:
- Section conflict detection prevents same component (ofedDriver,
  rdmaSharedDevicePlugin, sriovDevicePlugin) in both NCP and NNP
- Node selector overlap detection rejects NNPs targeting the same nodes
- Runtime overlap re-check catches drift from node re-labeling

DaemonSet isolation:
- Unique DS/ConfigMap names via NameSuffix (e.g. mofed-...-gpu-nodes-ds)
- ds-owner label on both DaemonSet and pod template for ownership
tracking
- Node discovery scoped by NNP nodeSelector to avoid phantom DaemonSets

mofed.wait label management:
- NNP controller sets mofed.wait based on OFED pod readiness on its
nodes
- NCP fallback path excludes NNP-managed nodes from label cleanup
- NNP deletion keeps mofed.wait=true while pods terminate, then
converges

DOCA driver upgrade:
- Per-NNP upgrade state managers with isolated requestor IDs
- NCP upgrade cleanup scoped to exclude NNP-managed nodes

Controller deduplication:
- Shared NicPolicyCR status accessors for polymorphic status updates
- Extracted helpers: requeueWithDelay, updatePolicyCRStatus,
  setupOFEDCatalog, watchStateSources
- ResolveNodesByPolicy as single source of truth for node resolution

Also adds heterogeneous-cluster-support.md with architecture docs
and mermaid flowcharts.

804 of 1193 new or added lines in 27 files covered. (67.39%)

4834 of 7786 relevant lines covered (62.09%)

43.7 hits per line