doing simulations of xcache at US scale. Due to wide interest in doing the same for different sites, I will be giving hands-on tutorial on it during the next Friday ADC analytics meeting. 

In essence, it seems that a two layer xcache setup would be sufficient. Now investigating differences between analy and prod queues data accesses. 

Working out a plan to deploy Tensorflow as a service. Options are:

* TFAAS - from Valentin K.

* TF serving - from Google

* TFjs

Also looking at spark with TF