version number evolution discussions from Lyon 2014

version-pm-evolution.txt

Broad decisions about rationalizing version object behavior based on
discussions at the Lyon QAH.  Participants: David Golden, Ricardo Signes,
Karen Etheridge, Leon Timmermans, Peter Rabbitson and Graham Knop

- version comparision should be done irrespective of the presence of
  underscores in the string used to initialize the version object

- underscore should no longer be used as a tuple separator in vstrings or
  vstring-like strings; vstrings are converted to tuples by splitting into
  *characters* (not bytes) and converting to codepoints; any elements after
  the first must be in the range 0-999

- numify/normal should produce a standarized string representation without
  underscores

- stringify should produce the best possible representation of the value
  used to initialize the version object; it should include underscores
  only if the initializing value was a non-vstring string.

- floating point numbers used as initializers are converted to a decimal
  string form at the precision limit of the architecture; people will be
  warned about this in the documentation

Examples:

Comparison:
    - version->new(1.0203)      ==      version->new("1.0203")
    - version->new(1.02_03)     ==      version->new("1.02_03")
    - version->new(v1.2.3)      ==      version->new("v1.2.3")
    - version->new(v1.2.3_0)    ==      version->new("v1.2.3_0")

Underscore no longer tuple separator:
    - version->new(v1.2.3_0)    ->      tuple (1,2,30)
    - version->new("v1.2.3_0")  ->      tuple (1,2,30)

Numify/normalize don't produce underscore:
    - version->new("1.0203")->numify    -> "1.0203"
    - version->new("1.0203")->normal    -> "v1.20.300"
    - version->new("1.02_03")->numify   -> "1.0203"
    - version->new("1.02_03")->normal   -> "v1.20.300"
    - version->new("v1.2.30")->numify   -> "1.002030"
    - version->new("v1.2.30")->normal   -> "v1.2.30"
    - version->new("v1.2.3_0")->numify  -> "1.002030"
    - version->new("v1.2.3_0")->normal  -> "v1.2.30"

Stringify should attempt to preserve string initializers:
    - version->new("1.0203")->stringify     -> "1.0203"
    - version->new("1.02_03")->stringify    -> "1.02_03"
    - version->new("v1.2.30")->stringify    -> "v1.2.30"
    - version->new("v1.2.3_0")->stringify   -> "v1.2.3_0"
    - version->new(1.0203)->stringify       -> "1.0203"
    - version->new(1.02_03)->stringify      -> "1.0203"
    - version->new(v1.2.30)->stringify      -> "v1.2.30"
    - version->new(v1.2.3_0)->stringify     -> "v1.2.30"

Another other thing I wonder about is whether "numify" should have a
minimum decimal length (e.g. padding with zeros on the right when
necessary).

we also talked about deprecating is_alpha

Just want to throw a couple of comments attached to the above:

• There was no significant discussion of the status of _'s in a distname as 'no_index' flags, and the current status quo in this regard is unchanged.
• There is broad agreement to make no significant efforts to preserve _ during runtime (also see the table above), however tools performing static analysis are still expected to preserve their current logic (with all heuristic as is).

As an aside - what the above doesn't seem to cover for me is the sanity of keeping the ->normal semantic broken (i.e. s/_/./). I did not see a case where breaking the normal() behavior will result in any extra breakage on top of everything else.

@ribasushi, I think github markdown ate some of your comments, but my reactions nonetheless:

• We're only discussing behavior of version.pm (and by implication, UNIVERSAL::VERSION) and so distribution names are outside the scope. We did say that stringification should attempt to preserve underscores in quoted versions so that other external tools that do have semantics for "_" can make use of them.
• I don't think "tools performing static analysis" is a correct term. Can you give an example of what you mean? (If you mean does MM->parse_version return "1.02_03" when our $VERSION = "1.02_03", then I concur.
• On "normal", we currently have an issue where version->parse("1.02_03")->normal is broken. JPEACOCK "fixed" it by making that fatal. We don't want "_" to be a tuple separator, and it's tricky to figure out how to represent in a round-trippable way. 1.2_3 can be v1.2_30.0 but 1.234_567 can't be v1.234_567 because we don't want underscore as a tuple separator. v1.234_.567 is a legal v-string, but looks quite odd. By having normal never keep the underscore, 1.234_567 can numify to 1.234567, normalize to v1.234.567 and stringify to 1.234_567 as it was originally. Does that answer the question?

To summarize the outcome of the discussion: we need to say in the docs something to the effect of

"->normal should not be used by anyone besides packagers, and packagers should not be using anything other than version->new($ver)->normal"

Thanks @dagolden for clarifying stuff, @ribasushi is reasonably happy.

stringify should produce the best possible representation of the value used to initialize the version object; it should include underscores only if the initializing value was a non-vstring string.

I do not remember the rationale behind this, and it doesn't make sense to me right now. Why did we want to threat vstrings special? It makes more sense to me to have the closest to the original representation.

I'm pretty sure the rationale is to remove the reliance on v-string magic, so it could hopefully be removed in the future. v-strings already ignore underscores for their string component. I suspect @dagolden may be able to confirm that.

@Leont also I believe that sentence is badly worded. It's trying to distinguish between quoted strings and v-string literals. We want to preserve underscores in all forms of quoted strings, but not in v-string literals, as the examples show. This isn't making vstrings special, but putting them on equal ground with numeric forms. version->new(v1.2.3_0) won't preserve the underscore, just like version->new(1.02_03) doesn't.

The proposed patch has been integrated into 0.9912_01 and thence to 0.9913 except for the limitation "any elements after the first must be in the range 0-999". For one thing, for dotted decimal versions (either quoted or bare v-string), any leading zeros are not significant and will just be ignored when parsing (so 1.2.0000003 => v1.2.3). The other thing is that the current code does not enforce the 3 digit limit for subsequent elements. That can be considered a TODO for now.

stringify should produce the best possible representation of the value used to initialize the version object; it should include underscores only if the initializing value was a non-vstring string.

This is confusing Module::Build, because it no longer recognizes 'v1.2.3_1' as a testing version. See #110882. I have no idea if anyone is actually depending on this, but it does break at least one test. @xdg: opinion?

My understanding is that 'v1.2.3_1' will still be seen as alpha, but v1.2.3_1 will not. The quoting being the important distinction.

Add a title of something like "# The Lyon Compromise on Perl Module Version Handling" might help. I noted a long time ago that searchability for this was low. I'm dealing with another problem related to this in briandfoy/cpan-mini-inject#11 (likely similar to #110882).