mlspawner uses Subnets from the internet

[sluetze]

To avoid conflicts with local docker networks, the code uses the subnets 172.33. upwards.
This IP-Ranges are public ip ranges. The networks only exist locally on each host, but there may be a issue, when a spawned container wants to contact a service/ip/Packetresource which resides in the same created subnet. even though this case may not have a high probability, it theoretically can occur.

ml-hub/resources/mlhubspawner/mlhubspawner/mlhubspawner.py

Lines 27 to 31 in 7644f35

	# we create networks in the range of 172.33-255.0.0/24
	# Docker by default uses the range 172.17-32.0.0, so we should be save using that range
	INITIAL_CIDR_FIRST_OCTET = 172
	INITIAL_CIDR_SECOND_OCTET = 33
	INITIAL_CIDR = "{}.33.0.0/24".format(INITIAL_CIDR_FIRST_OCTET)

I would recommend to use another private subnet range, as specified by https://tools.ietf.org/html/rfc1918 Chapter 3. Since the 192.168.x ranges are often used in private networks, and only have a small subnet-possibility I would use a 10.x.x.x Net-Range

[ClementGautier]

The problem is that 10.x.x.x is often used for Kubernetes as default CIDR for services (kubeadm --service-cidr defaults to 10.96.0.0/12, Rancher RKE also uses similar ranges for cluster cidr). Ideally, not using fixed ip and relying on kubernetes as a scheduler would solve that.