leaving perl v5.8 behind

On New Year’s Eve, I posted that I’d uploaded 114 updated distributions to the CPAN. Many of those, in addition to updating distribution metadata, made some changes to the version of perl they require, or say they may require in the future. I mentioned that in my last post. I was adding text something like this:

This library should run on perls released even a long time ago. It should work on any version of perl released in the last five years.

Although it may work on older versions of perl, no guarantee is made that the minimum required version will not be increased. The version may be increased for any reason, and there is no promise that patches will be accepted to lower the minimum required perl.

I’ve got more than one piece of boilerplate for various forms of policy, but generally they are tied to one thing, conceptually: the ongoing march of time. The policy above (“long-term”) says you should expect the code to keep working for perls up to five years old. Another says ten. Another tracks the p5p-supported perl versions, which is also effectively time-based. There are three that don’t:

  • toolchain: This says the code will abide by the consensus of The Toolchain Gang, which mean v5.8.1 until further notice. (More on that below.)
  • none: This is abandoned software.
  • no-mercy: I will change anything at any time. Do not rely on anything about this software.

I use those quite rarely. I find that mostly I’m putting long-term support onto modules, and feel good about that.

what about older perls?

I put “perl v5.8” in the title of this post, but I don’t want to cloud the point: I think some later versions are also too antique to be asked to cater to. I’ve been pretty happy to bump the minimum required version of a library to v5.12 for the simple creature comforts provided there, and I don’t feel bad about it. So, why not?

Let’s consider Test::Fatal. I wrote Test::Fatal in 2010, and I wrote it to make it easier to run the most cutting edge perl. The previous solution, Test::Exception, would break now and then on development versions of perl. In 2010, I wasn’t project manager for perl5 yet, but I was involved in p5p, and I ran the monthly snapshots as my primary development perl. I wrote Test::Fatal so more code would keep working on the newest versions of perl. We converted Moose to use Test::Fatal within a week of its release.

Test::Fatal was pretty widely adopted. Today, 16k other CPAN distributions depend on Test::Fatal, either directly or indirectly. In CPAN jargon, we sometimes say that means that Test::Fatal is “high up the river”, meaning that many other libraries are affected by changes to Test::Fatal, and so it should not change too precipitously. I was in the room when the river metaphor was first drawn on a whiteboard, and I think it’s a good one.

I also think that a big question is what it means to change too precipitously. I think, for example, that releasing a new version of a high-upriver library that suddenly makes an uncommonly used method fatal is bad form. Better to give notice via deprecation warnings or, when that’s not possible, documentation. People actively developing against the library maybe continuously integrating, and they’ll be better served by “fix this soon” than “stop everything and fix it now.”

On the other hand, I think there should be a basic assumption that eventually a platform will become obsolete and abandoned. That is: at some point, it is reasonable to assume that anybody on an n year old platform is accepting the implicit risk that they will be unable to run new versions of programs. They can, of course, keep running the old version! Also, if they really need part of the update — you know, the part that didn’t start relying on a new perl — maybe they can build their own forked version.

This is a pretty common conception throughout software. New versions of software eventually stop building on old compilers or on old operating systems. Updated websites stop working on MSIE6. Different ecosystems will have a different take on what version of the platform is too old to support, but there tends to be a norm. On the CPAN, there isn’t really such a norm. Or, maybe, the norm is v5.8, which just means the expected time window is one year longer every year.

It wasn’t always like this. Many of the libraries I maintain once supported perl v5.4 or v5.6, and when I moved them to v5.8, there was very little noise – in part, I think, because the general churn in all parts of Perl development was higher. We were doing more, so it was normal to keep moving forward. Why did it change? Well, first, consider questioning whether it changed. I think so, but I’m relying on my memories of a general vibe. This is hardly science. And as to why I think it changed? It’s hard to say. I think that perl v5.8 was the go-to for a very long time. We had two years from perl v5.5 (which was not a dev release, despite the odd number) to v5.6, then another two years to v5.8. It took five years to get the next version, v5.10, and then many people held off until v5.10.1, which was another two years. During this time, v5.8 got deeply entrenched.

But that was all ages ago. Perl v5.10 came out in 2007, during the presidency of George W. Bush, the same year as the first iPhone. Anyway, around fifteen years ago. So, when a new version of a CPAN library today starts to require a version newer than v5.8, the affected audience is people who:

  1. have had fifteen years to upgrade their perl, but chose not to
  2. continue to want to install the latest versions of libraries from the CPAN

I hold that catering to this behavior is a bad idea. It penalizes the module author or maintainer. Despite (potentially) writing all their private code targeting a recent perl, they must switch into another dialect of perl for their CPAN work. As time goes on, the gulf between these dialects grow. When a library is released in 2010 and targets v5.8, it forgoes the benefits of v5.10 that the author might like. By 2022, it’s now also forgoing the benefits of v5.12, v5.14, v5.16, v5.18, v5.20, v5.22, v5.24, v5.26, v5.28, v5.30, v5.32, v5.34, and v5.36. Some of these benefits may be small, like using say instead of print with a newline, or like using //= instead of a ternary. Still, the maintainer has to remember all this, and has to suffer through a more tedious experience, and with the knowledge that this gulf will only expand, because the far end is tied to a fixed point.

There should be a line drawn somewhere. I have tentatively declared, for me, that I think it should be around five versions, with wiggle room based on circumstances.

what if I’m affected?

I try to imagine what I’d do if I had to maintain important code that ran on v5.8. I can’t imagine it, though, because I’m living it. Most of the Perl 5 code at work runs on v5.34, but there are some significant bits of code that still run on v5.8.8, for reasons that are not particularly interesting, but also not particularly easy to work around. (Ask me again in a year, I hope it’s all on v5.30-something by then.)

So, what do I do when somebody (like, say, me) releases a new version of something that used to work on v5.8 but now requires v5.12? Normally, nothing. I look at the changelog and if it’s terribly important (say, a CVE), we port that fix to our own production environment. We deploy from a private CPAN mirror that has been pinned to the versions we use. This prevents us from shipping random versions to production without review. It’s something like Pinto, but a bit different. (It’s called XPAN, and we never really managed to make it generally useful software.)

Of course, most people don’t use XPAN or Pinto, and the idea of it’s a bit weird unless you’re a bit of a CPAN-head. On the other hand, there’s Carton, which follows the model used by many deployment systems, now. You pin the versions you’re using one time, and then bump them deliberately later. Carton isn’t the only solution in this space for Perl. It’s not even the only one by Miyagawa, who also wrote Carmel.

If you need to pin your versions now, before more things start needing new perls, I think those are good solutions. Also, I think that’s it’s an appropriate level of onus. My position is that over time, the cost of not upgrading should be paid more by the non-upgrader. The measure of time over which the burden shifts may not be clear cut, but I don’t think we should worry much about perls over ten years old.

can’t you just fork?

Let’s say I wanted to make Test::Fatal use v5.20. One option would be to fork Test::Fatal into Test::Fatal2. The API wouldn’t change, just the required version of perl. Then I’d go find all my own downstream code that used Test::Fatal and make it use Test::Fatal2. But I’d have to also fork that code, if it was on the CPAN, for similar reasons: if its version was thus going up, it has to be forked, too, right?

Meanwhile, if Test::Fatal has a downstream dependent which has another upstream dependency that depends on Test::Fatal, and that upstream doesn’t switch to Test::Fatal2, it’s a bit of a mess all over, especially since either Test::Fatal is now unmaintainer or there are two different maintainers.

Now multiply that across n things I might want to use a newer perl.

Now imagine that someday I want to make Test::Fatal2 use v5.30, and am pressured to instead make Test::Fatal3.

If something has to fork, it should be the CPAN itself. That is: provide an alternative CPAN mirror whose index is all 5.8-safe versions. This was offered years ago (by David Cantrell, if I recall correctly), but didn’t last. Really, though, I maintain that the right solution is pinned versioning.

perl, backcompat, and the toolchain

Perl has a great track record on backward compatibility. Breaking changes have been very rare, and generally pretty small. This has been for many reasons, but I can say that a large part of this policy was worked out and made explicit by Jesse Vincent and me in the v5.12-v5.18 era, and the goal was to make it as easy as possible for people to upgrade their perl often. We also spent time working with downstream vendors to make sure they could keep packaging new perls in a timely way. Again, I think the right priority here is to make it easy to upgrade perl, and easy to upgrade important libraries on that recent perl. Making it easy to upgrade important libraries forever on very old perls is not a plan I endorse.

Among other reasons that I’ve mentioned, I also think that spending time on improving perl itself is significantly less valuable and rewarding when those improvements are more or less not available.

So, what about the toolchain? I mentioned, above, that there’s a “toolchain” policy in my set of boilerplates, which says it will stick with v5.8.1. This is the result of a set of conversations at the Perl QA Hackathon, an event which was later given the more apt name “Perl Toolchain Summit”. It was an annual event (last held before COVID) where folks who maintained PAUSE, CPAN clients, Test::Builder, and similarly “Very Core to The CPAN Working” code got together to solve problems and get things done. I have always had a very high opinion of the value of those summits.

At the summit held in Lancaster, one of the questions was about what version of perl “the toolchain” could require. “Toolchain” meant, here, the CPAN installer (CPAN.pm), the distribution install tools (ExtUtils-MakeMaker, Module-Build), and Test::Builder, which powered all the basic testing code used by everything on CPAN at the time. The big questions were:

  1. Could we all agree to drop support for v5.6?
  2. Could we all agree to require later v5.8 versions, which had significant Unicode fixes?

The answers at the time were “yes” and “no” respectively, but the “no” was very tentative. We agreed that if toolchain maintainers hit significant problems in v5.8.1 that would be fixed by later v5.8 versions, we could bump to those versions, up to and including v5.8.4, which shipped with Solaris 10.

This decision was part of “The Lancaster Consensus”, which agreed a number of other things, too. I remember the process very well, as well as the little hotel we stayed at. Also, I had “Life in a Northern Town” in my head all week.

But I digress.

I think there are few things I’d point out regarding the Lancaster Consensus and its agreement about perl v5.8.1:

  • It was meant as much to free people from v5.6 as to keep them on v5.8.1.
  • It was about a very limited set of libraries that form the absolutely final layer of version compatibility. Without their functioning, nothing else can be installed.
  • It was ten years ago. (If it had said “perl from 10 years ago” instead of “v5.8.1”, it would’ve meant v5.8.1 when signed and v5.18 now.)

what’s next?

I am not planning to go crazy and update all my libraries to v5.34. (I acknowledge that someone will say that I have already lost my mind in requiring v5.12 in many things.) On the other hand, I am not planning to put in any effort on continued compatibility with v5.8. As I release more software, I will require newer versions of perl. That’s what’s next.

I’ve done a decent amount of work, over the years, on a little set of programs that gather and analyze dist metadata from the CPAN. I’ll probably write a bit more about them, and how I have used them to see things about versions and prerequisites. I also might write a bit about Miyagawa’s report on what version of perl is using cpanm and some similar work I did recently. But I might not, we’ll see where I get to in the next week or two.

I’ll also keep trying to make progress on bug backlogs, build improvements, and improvements to perl. Wish me luck!

Written on January 8, 2023
📚 cpan
🐫 perl
🧑🏽‍💻 programming