Hey Yi, thanks for the write-up. Has your team been able to achieve high utilization of GPU using TensorFlow Serving? How do you distribute traffic evenly between the different pods? Do you use batching? Happy to chat and compare notes as my team is working on similar problems.