Wu Yi

@Yancey1989 thought `TrainingJob` is more general, not only paddle training.

Support @putcn 's idea! Proxing and filter is simple enough and easy!

Thought kubernetes will use `--device` argument to attach GPU device to the container, and this issue is very possible that has something to do with docker.

Online of course. Offline inference can use the same method as training.

是需要申请的。请问是需要申请国家工程实验室 集群资源还是百度内部集群资源？

@Yancey1989 Do you got time to fix this?

的确，包含"sub-command"的程序总体看大部分的command line的设计都遵守： ``` $ [command] [verb] [target...] [options] ``` 的规范，比如：`git clone`, `git checkout`, `pip install`, `kubectl create`, `docker pull`, `yum list`。 不实用"sub-command"的unix命令基本也都是一些动词，比如：`ls`, `mkdir`, `kill`。 > 另外一种选择是把pfs的操作客户端和paddlecloud命令分开为两个，如： 赞同这种方式。其实hadoop的命令的设计是比较奇葩的，个人觉得可以不必参考。

> 另外一种选择是把pfs的操作客户端和paddlecloud命令分开为两个，如： 补充：后来想了下，分成两个命令还是不合适，这样用户得下载两个二进制了。只下载一个二进制直接执行还是会简单很多。

Supporting for the high-performance network is indeed what we needed, but using host network will break the design of Kubernetes. Kubernetes design the network to be "flat", means containers should...

@drinktee Yep. Still, I think this solution should be a "workaround", it's not a best technical choice.