[azernik]

Having done some work on implementation/migration:

EVERYTHING needs to work.

There's a hell of a lot of software (and, even worse, hardware) that encodes assumptions about IPv4. Everything from userspace asking specifically parsing only IPv4 addresses, to only listening for IPv4 incoming connections, to variables using the IPv4-only type, &c. And let's not even start on the L3 hardware acceleration built into so much networking equipment.

You can't flip the switch when 75% of the stuff on your device is ready. The remaining 25% will break, badly. You need to switch over every single piece of code. And then before you get any benefit out of it, you need to do the same on every remote system you want to talk to.

This is the magic and the terror of internetworking. Before IP, you had to go through this nightmare for changes at many different layers of the stack. With the clean-ish separation of IP, only one layer was make-or-break like this: layer 3. Any change to layer 3 is horrifically difficult, purely as a deployment/management challenge.

[kazen44]

this comments hit up a good point,

the lower in the OSI stack you go, the harder it is to initiate change.

There is still some (mostly industrial) hardware that has lackluster TCP/IP support at best (for instance, assuming the netmask is always a /24...).

migrating these networks to ipv6 is not possible, and doing 6to4 adds a crapton of complexity.

[azernik]

It's not about "lower".

We can, and do, make massive , non-backwards-compatible changes in layers 1 and 2 every few years and no one is the wiser. Layer 3 is special. It makes that flexibility in all the other layers possible, but keeps none for itself.

[kazen44]

mind you though, that layer 3 and 4 are somewhat interlinked.