refactored ipns records to point to paths

[whyrusleeping]

Also changed the ipns dns resolution to use the "dnslink" format

[jbenet]

may want to do the parts := strings.Split(txt, "/") part first, as it's faster than b58 decoding. maybe:

parts := strings.Split(txt, "/")
if len(parts) == 1 {
    kp, err := ParseKeyToPath(txt)
    if err == nil {
        return kp, nil
    }
}
if len(parts) < 3 {
    return "", ErrBadPath
}

[jbenet]

comments above o/

[wking]

On Fri, Apr 17, 2015 at 11:48:58PM -0700, Jeromy Johnson wrote:

Also changed the ipns dns resolution to use the "dnslink" format

I don't know if there's a more current spec out there, or if there's
an RFC or some such for “dnslink”, but the cap2pfs paper was
suggesting 1:

ipfs=XLF2ipQ4jD3U …

It looks like DNS service discovery (RFC 6763) prefers equals to
separate keys from values 2, so I think we should stick with that
here. I'm also not sure how the unqualified example:

QmY3hE8xgFCjGcz6PHgnvJz5HZi1BaKRfPkn1ghZUcYMjD

is supposed to be interpreted, so I'd suggest we drop that value in
favor of the more explicit forms. That means I like:

ipfs=QmY3hE8xgFCjGcz6PHgnvJz5HZi1BaKRfPkn1ghZUcYMjD
ipns=QmY3hE8xgFCjGcz6PHgnvJz5HZi1BaKRfPkn1ghZUcYMjD
ipfs=QmY3hE8xgFCjGcz6PHgnvJz5HZi1BaKRfPkn1ghZUcYMjD/foo
ipns=QmY3hE8xgFCjGcz6PHgnvJz5HZi1BaKRfPkn1ghZUcYMjD/bar
ipfs=QmY3hE8xgFCjGcz6PHgnvJz5HZi1BaKRfPkn1ghZUcYMjD/foo/bar/baz

for the DNS TXT record format (but maybe this PR is about the symlink
format?).

It's also a bit odd to use the same hash for both the IPFS and IPNS.
A valid IPNS hash will be a public key, not a directory, so that's not
going to work for:

ipfs=QmY3hE8xgFCjGcz6PHgnvJz5HZi1BaKRfPkn1ghZUcYMjD/foo

(which asks for children of public key?). Basically, /ipns/ should
mean “please dereference this to the signed reference and then process
again” So if you had (in IPNS):

ipns=QmbqDJaoZYoGNw4dz7FqFDCf6q9EcHoMBQtoGViVFw2qv7
→ ipfs=QmbqDJaoZYoGNw4dz7FqFDCf6q9EcHoMBQtoGViVFw2qv7

Then:

ipns=QmbqDJaoZYoGNw4dz7FqFDCf6q9EcHoMBQtoGViVFw2qv7

would be equivalent to:

ipfs=ipfs=QmbqDJaoZYoGNw4dz7FqFDCf6q9EcHoMBQtoGViVFw2qv7

as it would be if there were multiple layers of mutable IPNS
references in between:

ipns=QmbqDJaoZYoGNw4dz7FqFDCf6q9EcHoMBQtoGViVFw2qv7
→ ipns=QmaV5w1y1GNdU8ZMdjJwqpCgKeNzR4fX5TAZg5Dg9TZSiT
→ ipfs=QmbqDJaoZYoGNw4dz7FqFDCf6q9EcHoMBQtoGViVFw2qv7

So I'd use different hashes for IPFS and IPNS:

ipfs=QmVyS3iAy7mvDA2HqQWm2aqZDcGDH3bCRLFkEutfBWNBqN
ipns=QmbqDJaoZYoGNw4dz7FqFDCf6q9EcHoMBQtoGViVFw2qv7
ipfs=QmVyS3iAy7mvDA2HqQWm2aqZDcGDH3bCRLFkEutfBWNBqN/foo
ipns=QmbqDJaoZYoGNw4dz7FqFDCf6q9EcHoMBQtoGViVFw2qv7/bar
ipfs=QmVyS3iAy7mvDA2HqQWm2aqZDcGDH3bCRLFkEutfBWNBqN/foo/bar/baz

For symlinks stored in an IPFS filesystem or as IPNS values, it makes
more sense to me to store them as link objects, like 3, but with an
optional child-path array:

type IPFSSymbolicLink struct {
hash Multihash // IPFS hash, or an IPNS reference
children []string; // names of children for drilling into the
// directory that the Multihash eventually resolves to.
}

and no name (that would be a key pointing to the symlink, not part of
the symlink value) or aggregate ancestor size (which is poorly defined
once you allow mutable IPNS references).

[jbenet]

@wking the format of the DNS TXT record is now (after ipfs paper DRAFT 3 was written):

dnslink=<path>

For example:

dnslink=/ipfs/QmWbKNc2sutt3Mboemu6Wsr6qU4hhCwekjg4WMK2dsfJvE/foo/bar/baz
dnslink=/ipns/QmdsnptcQjjAGzE5WiDJSupA5UmGJtVS4hzsrVdMkBjAah/foo/bar/baz

and it's general for non-ipfs use cases:

dnslink=/dns/foo.com/foo/bar/baz
dnslink=/bittorrent/<infohash>
...

Of course, this needs both to catch cycles and a recursive lookup limit.

Why dnslink ? i don't believe it's in use. perhaps it should be just link, but i worry that may clash with something.

relevant: jbenet/random-ideas#28

[whyrusleeping]

I addressed the comments above, but just realized that path.Resolver doesnt resolve ipns paths yet

[jbenet]

but just realized that path.Resolver doesnt resolve ipns paths yet

maybe it should be a test case

[whyrusleeping]

I mean, the logic for even thinking about ipns isnt there yet. hasnt been implemented yet

[jbenet]

@whyrusleeping do you mean: handling a DNS -> IPNS -> IPFS record? (i.e. right now it just handles DNS -> IPFS?

[whyrusleeping]

@jbenet no i mean, if i do:

resolver := path.NewResolver(...)
mypath := path.Path("/ipns/QmBlahBlah/something/else")
resolver.ResolvePath(mypath)

It wont work. the resolver doesnt understand ipns

[jbenet]

@whyrusleeping ah right, not the basic resolver.

[whyrusleeping]

not sure what the best way forward here is. because the ipns resolution stuff uses the straight ipfs resolver... so adding ipns into the path.Resolver directly may make things funky

[jbenet]

so adding ipns into the path.Resolver directly may make things funky

let's not do that.

namesys package should probably have its own resolver? or how low-level do you need it?

[whyrusleeping]

travis is awesome: https://travis-ci.org/ipfs/go-ipfs/jobs/59675730

but otherwise i think this is good to go

[jbenet]

this is hard to read, maybe use vars?

newpathSeg := append(respath.Segments(), extensions...)
newpath := path.FromSegments(newpathSeg...)
return resolveRecurse(n, newpath, depth+1)

[jbenet]

@whyrusleeping

• shut down iptb in test
• path validity + tests
• hard to read
• tests dont pass at all

[whyrusleeping]

@chriscool will do!

[whyrusleeping]

i love these random little failures: https://travis-ci.org/ipfs/go-ipfs/jobs/59829161#L2957

reliable testing is hard to get right

[chriscool]

The failure is:

expecting success:
test_kill_repeat_10_sec $IPFS_PID

not ok 19 - 'ipfs daemon' can be killed

#
# test_kill_repeat_10_sec $IPFS_PID
# 

Maybe 10 seconds are not enough on Travis, or maybe something can be done so that ipfs daemon exits faster.

[whyrusleeping]

We should probably take some time to investigate ipfs shutdown a little better, ten seconds should be enough

[whyrusleeping]

@jbenet I beleive that i have addressed your concerns

[jbenet]

iptb is so useful. ❤️

[jbenet]

LGTM, RFM. i think this needs to be rebased.