NVIDIA / gpu-operator / 39808196378

45%

main: 37%

push

github

Filter DaemonSet watch updates to readiness changes

The ClusterPolicy and GPUCluster controllers watch their DaemonSets
without a predicate, so every status update enqueues a full reconcile.
During a rollout the DaemonSet controller rewrites the status once per
pod, which at fleet scale means thousands of reconciles that compute
the same result.

Add a predicate that passes an update only when the generation, labels,
annotations, owner references, or deletion state change, or when one of
the status comparisons the reconcilers read changes: desired, current,
unavailable, or misscheduled reaching or leaving zero, updated or
available reaching or leaving the desired count, or the observed
generation catching up. Pod count changes that leave all of these equal
are dropped. The not-ready requeue still polls DaemonSets that are
mid-rollout.

Signed-off-by: Rajath Agasthya <ragasthya@nvidia.com>

27 of 29 new or added lines in 3 files covered. (93.1%)

5886 of 13054 relevant lines covered (45.09%)

0.45 hits per line