[IGNITE-11871] [ML] IP resolver in TensorFlow cluster manager doesn't work properly

• Type: Bug
• Status: Open
• Priority: Minor
• Resolution: Unresolved
• Affects Version/s: 2.7, 2.8
• Component/s: ml
• Assignee: Alexey Zinoviev
• Reporter: Alexey Zinoviev
• Created: 24/May/19 11:30
• Updated: 20/Aug/20 13:37

Description

TensorFlow cluster manager requires NodeId to be resolved into IP address or hostname to pass the address/name to TensorFlow worker. Currently, it uses strategy "return first" and returns the first available address/name. As a result of that, in the case when the server has more than one interface cluster resolver might work incorrectly and return different addresses/names for the same server.

To fix this problem we need to update TensorFlowServerAddressSpec so that it returns the same address/name for the same server all the time. If a server has multiple network interfaces we need to find a "GCD", a network with all Ignite nodes.