A custom object extends a Kubernetes installation with a user-defined API type. It lets teams represent specialized work, such as distributed training, using Kubernetes-native objects instead of managing every workload component directly. section 8.3.1
Custom objects support an operator-driven workflow:
- Register the type dynamically; administrators can add, remove, or update it independently of the cluster. section 8.3.1
- Create and inspect instances with kubectl, just as with built-in resources such as Pods. section 8.3.1
- Let a stateful controller continuously monitor submitted objects and turn their specifications into running workloads. section 8.3.1
TFJob for distributed TensorFlow training
The Kubeflow Training Operator defines a TFJob custom object with two worker replicas, an image, a training command, and an OnFailure restart policy. Applying the manifest creates a TFJob; the operator then creates the corresponding worker Pods, which can be inspected with kubectl. section 8.3.1
