What's the function of tfReplicaType: Master ?

[lxl910915]

Why and when kubeflow removed Master?
In a doc, we see: The master only acts as the chief and doesn't do any training. When master started and intilized, how master waits ps and worker to completed?

[gaocegege]

Master now is replaced by chief.

[jlewi]

Earlier versions of TensorFlow used "Master" to indicate the process that coordinates distributed training. TensorFlow then changed to calling this "Chief".

We added "Chief" to be consistent with TensorFlow terminology; but we've kept "Master" for backwards compatibility with older versions of TF.

[lxl910915]

@jlewi
In version v1alpha2, is master removed? The tf_job yaml file has no master. We are very confusing for this.

[johnugeorge]

As @jlewi said, Master type is still kept for backwards compatibility. https://github.com/kubeflow/tf-operator/blob/master/pkg/apis/tensorflow/v1alpha2/types.go#L135

[lxl910915]

@johnugeorge
Do you mean tf_job only need to contain ps and worker ? no "Master", and no "Chief" ?