Comment by Macha | Hacker News Reader

Comment by Macha | Hacker News Reader

Parent Full thread

Because the problem is not the numeric representation. The problem is when you get a message from 1.0.0.0.1.1.1.1 to 0.0.0.0.192.168.0.1, and 192.168.0.1 is an IPv4 only device, what are you going to tell 192.168.0.1 about the source address? There's only a 32 bit field in the protocol.

Are you going to lie and tell it it's 1.1.1.1? Congrats, you've just reinvented NAT. Enjoy your stateful boundary system and all that complexity again.

Are you going to add some extension to IPv4 and try convince everyone to upgrade all their hardware, software and configurations? Then you have the exact same problem IPv6 adoption had, except IPv6 adoption is at like 45% globally and your new scheme is at 0%: https://www.google.com/intl/en/ipv6/statistics.html

If you have some other idea, make sure to check it against the list of IPv6 transition mechanisms before commenting - it probably exists for IPv6, and ultimately all of them have been displaced by dual stacking: https://en.wikipedia.org/wiki/IPv6_transition_mechanism

> IPv6 adoption is at like 45% globally [...] https://www.google.com/intl/en/ipv6/statistics.html

> link -> The graph shows the percentage of users that access Google over IPv6

How does this work in practice, does user equipment end up with dual IPv4 IPv6? if not how do they access IPv4 only sites (of which there are still plenty)?

It's just as you have described, this is called dual stacking and is currently the most common type of IPv6 deployment.

Alternatively, cellular ISPs may make use of NAT64+DNS64. In this case, the IPv4 addresses are combined with a NAT64 prefix (most likely 64:ff9b::/96), producing addresses such as `64:ff9b::198.51.100.1`. It is effectively the same as CGNAT, but pretty much everything is single stacked IPv6 up to the IPv4<->IPv6 border relay in the ISP network.

The way it's supposed to work is the ISP delegates a prefix to the user's router, the router will advertise for that subnet and client devices will use slaac to configure their ipv6 addresses. Meanwhile for IPv4 it operates as normal. For incumbent, mostly wired, Western ISPs that means a pool of IPv4 addresses and subscriber level NAT. For everyone else, that probably means CGNAT.

My ISP gives me both an IPv4 address and a heap of IPv6, most browsers will use what is called Happy Eyeballs to try both IPv4 and IPv6 at the same time and use the one that connects the fastest.

> Are you going to add some extension to IPv4 and try convince everyone to upgrade all their hardware, software and configurations? Then you have the exact same problem IPv6 adoption had

Yeah, so I'm going to say that's just not true. If IPv6 had been significantly simpler (e.g. by mainly focusing on increasing the address space) then adoption would be much higher today – probably universal. The problem is that IPv6 changes a lot, and that implementing it is a fair amount of work. The entire business with tunnel brokers and other compatibility schemes are complex and difficult to use.

To completely change gears now would be silly and too late, but it's been almost 30 years and it baffles me how anyone does not consider the entire thing a spectacular failure. People use Python 3 as a warning about compatibility, but IPv6 is a much larger catastrofuck, and that's largely due to decisions IPv6 people made. Whether IPv6 is "better" or not doesn't even come in to play.

Are you going to lie and tell it it's 1.1.1.1? Congrats, you've just reinvented NAT. Enjoy your stateful boundary system and all that complexity again.

(Shrug) It works. What problem is being solved here, again?

Just look at your favourite game console provider's forum discussions on "NAT types" and you'll see that it really doesn't. Or try self host on CGNAT. Or realise the only direction for IPv4 prices at AWS etc. is up

The funding of the ipv6 commitee probably.

And porn for those who like overcomplicated 'enterprisey' solutions where they aren't needed.

> There's only a 32 bit field in the protocol

IPv6 is 128 bits... so we already made a change. Why can IPv7 not be 128 bits as well, but vastly more readable/recognizable than IPv6. Hex is hard on the eyes, especially for those who don't stare at it all day.

If you'd like to figure out a nicer representation for IPv6 addresses, go for it. You needn't break any compatibility for it, just invent a new representation.

I'd argue that the most common format such as 2001:db8::f00f is already pretty darn readable, and not hard to copy+paste when you need bare addresses. Base85 is an alternative if you'd like something that's purely optimized for fewer characters, but it only really helps with exceptionally long addresses like 2001:db8:424b:b4ff:fcb5:b643:f721:b703 becoming 9R}vSk6QzV!G$<lwB;|~

> 2001:db8:424b:b4ff:fcb5:b643:f721:b703 becoming 9R}vSk6QzV!G$<lwB;|~

Memorize that :) Then try to tell it over the phone.

That's a totally contrived address just to get as long of a hex representation as possible, it's pretty far from a likely situation.

That being said, telling an address over the phone (another unlikely situation, to be honest) wouldn't be that hard in hexadecimal form.

> it's pretty far from a likely situation.

I don't know, I just checked my linux box that takes ipv6 from one of my routers and it's the full 16 bytes, no zeros to skip.

I suppose example.com has a lot of consecutive zeros and is shorter to type...

> (another unlikely situation, to be honest)

Kid calls you when you're out and asks where in your local network the minecraft server is running. Or on which samba server the mod archives are. Real story.

The move to hex is to avoid people having to memorise all these offhands like "so a /27 is x.x.x.0-31". Instead every "4" in the subnet is another digit.

But also do you think 24414.64517.21200.44298.46574.49887.23623.7394 is really that much easier to type? Or 231.22.96.68.14.229.38.41.242.44.72.220.156.50.232.95 if you keep to 8 bit octets?

Maybe I'm weird, but I do in fact think that is easier. People memorize long numbers all the time, such as SSN, phone numbers, account numbers, credit card numbers, etc. People are not very good at memorizing lengthy alphanumeric sequences such as randomized passwords. I don't know why, but it "just is".

But -

> The move to hex is to avoid people having to memorise all these offhands

Point taken. It does seem in practice though people will still have a need to memorize/recognize addresses for various purposes, such as administration/dns/etc.

> People memorize long numbers all the time, such as SSN, phone numbers, account numbers, credit card numbers, etc. People are not very good at memorizing lengthy alphanumeric sequences such as randomized passwords.

It's easy. Long sequence of 10 possible tokens that we're always used to seeing together. Easier to memorize because you subconsciously use any patterns you notice, and with just 10 possibilities the patterns are there. And you've been reading sequences of digits since you were old enough to get given pocket money.

Vs long sequence of 16 possible tokens, of which 10 belong in one cognitive group and 6 in another. The extra 6 you've never memorized before in random order because they're for representing your language and that has other ingrained patterns.

Let's not forget they admitted ipv6 addresses are too long and allow you to skip parts of it so now you have to count the colons carefully...

You don't understand why a sequence of 39 digits might be harder to memorise than sequences of 10? 16 is longest of the examples that people actually memorise (no, people are not memorising their bitlocker keys), and even then they have 1 example that lasts at least 5 years to deal with, instead of however many IP addresses they have to deal with.

Also many of my account numbers and my nations equivalent of a SSN are alphanumeric so 36 vs 16 choices.

I could be wrong, but I do not think it's easier to remember 45465465464938259 then it is to remember the lyrics to a popular song. Now I assume there is some sort of brain difference playing here. But while looking it up most popular memorization strategies are either ram it into your head with repetition or converting it to songs, rhymes, poems, or other word phrases.

It's quite easy to remember things like "Mmm-mmm, went the little green frog one day".

Microsoft did. The first example is pretty much what bitlocker recovery keys are.

Speaking of which, the slash notation has not gone away in IPv6 notation.

At least the user networks are /64s, so, to some degree, CIDR notation isn't relevant for someone just trying to set up their system - netmask is now always 64 bits. And sure, ISP/Network Engineers still care about /56, /48, /32, etc... but that's their day, and takes about a week to have it ingrained in a way that you almost never could with IPv4.

> But also do you think is 24414.64517.21200.44298.46574.49887.23623.7394 is really that much easier to type? Or 231.22.96.68.14.229.38.41.242.44.72.220.156.50.232.95 if you keep to 8 bit octets?

I mean... yes, actually; I can type those blindly on a normal numpad. I'm sure somebody somewhere has made a hex number pad, but realistically on any normal keyboard entering hex means going back and forth between number and letter keys.

On QWERTY layouts, you can type ABCDEF with your left hand and 0123456789 with your right hand on the num pad.

That would be really convenient if : weren't in the middle of the keyboard. Maybe hardcore IP typists need an alternate keyboard layout where * maps to :

You're more than welcome to represent an IPv6 address in some other format. That's just a implementation detail in the UI. You don't need a new wire protocol to do that.

(There's already at least 4 different ways you can write an IPv4 address, and your OS is happy to accept any of them.)

You gotta love how they decided to use colons as part of the notation! But, oops, that is also used for port number notations. Solution? Square brackets:

  [ff00:abcd::0]:1234

Yeah it's annoying, I've come across numerous address parsing bugs because of this.

I think the reason they needed to use a different delimiter is because IPv6 addresses can also omit entire portions of zeros as a convenience feature to shorten addresses, which means there are a subset of shortened addresses that look exactly like IPv4 if not for the delimiter.

> I've come across numerous address parsing bugs because of this.

To prevent that, you can write tons of test cases. IPv6 notation parsing and generating test cases are oodles of fun.

I think two major problems that IPv6 failed to account for are the unreasonable ambiguity and "global registry" proposed by the Unique Local Address range, e.g. fc00::/7. This was even after having the opportunity to fix the previous mistakes in the fec0::/10 range.

It was entirely inappropriate for small scale personal networking and assumed an "administrative scope" that's too cumbersome and without purpose for most implementations.

I wonder if a dedicated "local" network prefix with a fixed /64 mask and an operating system recognized "alias" would have been a good mechanism to introduce. Something like fec0::/64 with an option to call it "local::". Then you might be able to more easily move from a world of 192.168.10.1, 192.168.10.2 to a world of local::1, local::2, that being translated to fec0::1 and fec0::2.

What do you mean by global registry? ULAs do have a local bit, by setting it to 1 (i.e., fd00::/8) you can have your own ULA range. I am using ULA in my own home network.

Right, and the local bit being 0 is currently undefined, but the original and lasting idea (apparently) is to create a central registry for it, so that in some future case where you have to merge corporate networks with their own ULAs there is zero chance that there would ever be a collision between the two.

Even in so far as the local bit being 1, there have been at least two attempts to create a "voluntary global registry" for your randomly generated prefix, mostly for the same "collision avoidance" stated above.

In either case, somewhat unintuitively, those are still _global_ addresses, and don't have the same level of "non local routing" restrictions that the RFC1918 address space explicitly gained.

It's been a mess, and at some point, the IETF just seems to have given up on it.

> those are still _global_ addresses

They are meant to be globally _unique_ (or effectively so, by utilizing RNG you ensure that it is very, very unlikely for two organizations to have the same ULA), but they are not globally routable.

> Even in so far as the local bit being 1, there have been at least two attempts to create a "voluntary global registry" for your randomly generated prefix, mostly for the same "collision avoidance" stated above.

Like most things people don't like about IPv6, you don't like this because you misunderstand how it's supposed to work.

There is no need for a ULA registry because you're supposed to pick a random prefix within the ULA space. The ULA space has 7 bits already defined, so you're picking a random prefix from a space 57 bits big. The chance that you will pick the same prefix as some other specific person is effectively zero.

(the birthday paradox does not apply because we don't care if your prefix matches anyone else on earth, just the one specific person/organization that you want to merge networks with).

> In either case, somewhat unintuitively, those are still _global_ addresses, and don't have the same level of "non local routing" restrictions that the RFC1918 address space explicitly gained.

Err, what? ULAs are routable and "global" addresses in exactly the same way that RFC1918 addresses are routable and global. They are valid addresses within the address space and the only thing that makes them "not globally routable" is that all routers have certain prefixes baked into the firmware as non-routable. If you don't trust your router not to route them for IPv6, why do you trust it for IPv4?

> Like most things people don't like about IPv6, you don't like this because you misunderstand how it's supposed to work.

Thanks for explaining my own motivations to me, while at the same time denying that any discussed changes are relevant to anyone.

> There is no need for a ULA registry

I was merely summarizing what the public opinion on the matter happens to currently be. Which suggests that the current recommendation to pick a random address and be happy with it is inadequate. The original posts dissatisfaction with being able to deploy IPv6 _conveniently_ also hints that improvements would be welcome by most people.

> ULAs are routable and "global" addresses in exactly the same way that RFC1918 addresses are routable and global.

RFC1918 explicitly defines it's addresses as local and not globally unique. RFC4193 specifies that these addresses are meant to be globally unique. RFC1918 requires a more specific level of filtering than RFC4193, and the definition of "site local" has always had some ambiguity to it.

So, within the very specific and nuanced points I was trying to make, no, they're different.

Agree with the parent that there is no need for a ULA registry for locally assigned addresses under RFC 4193 (which is every one I've ever seen). But there is that option for globally assigning addresses that I'd be intrigued to see if anyone does anything with...

Also important to note that while RFC1918 is roughly equal to RFC4193/ULA, that "site local" is another beast entirely (FECO ) that was, as you noted, ambiguous and was deprecated in rfc3879

Well -that's not entirely true. RFC4193/ULA has a Local Bit that, in my experience, is almost set to 1, so the prefix is FDxx.... At a prior company we rolled out about 25 million nodes all in RFC4193 space (non-routability was a design objective) - and, just to be within spec, we did indeed grab a random /48 for each customer instance we rolled out (and made very good use of the 65K networks we had available within that /48).

But - what if that Local bit is set to 0 - then:

   Set to 1 if the prefix is locally assigned.
   Set to 0 may be defined in the future.

Presumably there was some thought that there might be a Global Registry that the OP was referring to?

Concur with the rest- and, honestly, RFC 4193 is one of the simply written RFCs out there - understandable within 30 minute read - I'd encourage everyone to dig into it.

Because 99.99% of people never see an IP address let alone interact with them.

> Hex is hard on the eyes

Thankfully there is near-100% compliance with PTR reverse-dns for IPv6 addresses and networks. /s

Reply on news.ycombinator.com