Incompatible name restrictions breaking customers.

[mattmoor]

In v1, image names were allowed to contain: [a-z0-9_-]{4,30}.
In v2, this was restricted to: [a-z0-9]+(?:[._-][a-z0-9]+)* (for all components)

While I understand the sentiment that v1 was overly permissive, we have customers of our v1 API, who are broken by 1.8+ where the v2 naming restrictions have been applied to all image name validation paths.

This breaking change seems to fundamentally conflict with this stated objective of v2:

These name requirements only apply to the registry API and should accept a superset of what is supported by other docker ecosystem components.

The pattern being employed by most (possibly all?) affected customers is to use hostnames for their image names, e.g. foo.bar-baz.com. Therein lies the problem because hostnames are slightly more permissive (in some ways) than the current regular expression:

<hname> ::= <name>*["."<name>]
<name> ::= <letter-or-digit>[*[<letter-or-digit-or-hyphen>]<letter-or-digit>]

(source: StackOverflow quoting spec)

In particular, we have customers with hostnames of the form foo--bar.baz.com.

While it still leaves the possibility that v2 name validation breaks valid v1 images, my proposal is to relax the v2 name restrictions to allow valid hostnames as path components.

I believe that means changing this regular expression to:

// Allows a hostname part (also allowing underscores)
var NameSegmentRegexp = regexp.MustCompile(`[a-z0-9](?:[_-]*[a-z0-9])*`)
// Allows several hostname parts, each separated by a dot.
var RepositoryNameComponentRegexp = regexp.MustCompile(
    NameSegmentRegexp.String() + `(?:\.` + NameSegmentRegexp.String() + `)*`);

[mattmoor]

@stevvooe @dmcgowan @dmp42 @ncdc @vbatts @jlhawn

Folks care to weigh in on this? Has the ship sailed on potentially allowing this into 1.9? I'm happy to get a PR together ASAP, if that's a possibility.

[mattmoor]

@dmcgowan I updated the regular expression, thanks for spotting the bug.

This RFC: https://www.ietf.org/rfc/rfc2396.txt includes:

      host          = hostname | IPv4address
      hostname      = *( domainlabel "." ) toplabel [ "." ]
      domainlabel   = alphanum | alphanum *( alphanum | "-" ) alphanum

Technically foo.bar.baz. is a fully-qualified-domain-name, but I'm open to disallowing the trailing ., the main part we need to support is for one or more domainlabel segments separated by .s.

I could go either way on support for adjacent _, but it is specifically called out in the linked issue. It would be easy enough to move this from where it is now to a group with the \..

I will send a PR for this ASAP.