Build TFServing images for different TF versions

[jlewi]

We should probably build different versions of TFServing images corresponding to different TF versions.

[jlewi]

Is the HTTP Proxy TF Version specific?

[lluunn]

No that should work for all TF versions.

[pdmack]

Some strategy discussion in #608

[ankushagarwal]

Should we use the same strategy as jupyter notebook images?

 gcr.io/kf-images/tensorflow-serving-1.4.1-cpu  
 gcr.io/kf-images/tensorflow-serving-1.4.1-gpu  
 gcr.io/kf-images/tensorflow-serving-1.5.1-cpu  
 gcr.io/kf-images/tensorflow-serving-1.5.1-gpu  
 gcr.io/kf-images/tensorflow-serving-1.6.0-cpu  
 gcr.io/kf-images/tensorflow-serving-1.6.0-gpu  
 gcr.io/kf-images/tensorflow-serving-1.7.0-cpu  
 gcr.io/kf-images/tensorflow-serving-1.7.0-gpu

[lluunn]

We should have different tf version, cpu/gpu, and py2/py3 combination for TF serving image as above, e.g. gcr.io/kf-images/tensorflow-serving-1.6.0-py3-cpu

• jupyter notebook installs py2 and py3 together, but I think we have to separate that in serving case.
• Jupyter notebook has parameterized Docker image.
• The logic is embedded in the workflow

I think it's a little hard to read and maintain. And it's hard for users if they want to customize.

How about listing the parameters in a file (yaml or json), and the workflow runs a script to read from the file and build the image?

cc @ankushagarwal

[ankushagarwal]

I think it's a little hard to read and maintain. And it's hard for users if they want to customize.

We should think of a simpler unified approach for building our images - jupyter notebook images, tf-operator images and tf-serving images - #666 would be a good place to track this.

[jlewi]

@lluunn Is this item done? Can we close this issue?