SGTM.

What if users want to add sidecars to do things like

• handle logging
• download/stage data e.g. in the case of Pachyderm

I think assuming there will be only one container is two limiting

The constraint of naming the container 'tensorflow' is fine when there are sidecars. But if there is only a single container per pod (which is not an uncommon scenario), it doesn't make sense of have this restriction.

So we could drop the requirement that a container be named TensorFlow if there is a single container but I'm not sure I like the idea of only imposing the requirement of having a container named TensorFlow when there is more than 1 container.

If there is a single container that isn't named "TensorFlow" can't we just provide a very clear error message that explains how to fix things?

I meet a problem when I am fixing #532 , I found that it is hard to set the default port for tensorflow container if we do not know which one is it.

Consider that we have two containers in the pod and one is tensorflow container named "my-name", and a sidecar named "sidecar". We need to set the default port for the tensorflow container since it does not have one. Then we can not do it without the name "tensorflow" since we do not know which one is the real tensorflow container.

What do we want to do in 0.3.0? Do we leave the behavior as is or do we change it?

We can not change it since we need it to find the port and generate cluster spec.

Per the above closing this issue.