What's the best way to maintain a package repository that have the major version in the import path?

[alex....@gmail.com]

vgo suggested that package developers put the major version into the import path e.g. foo/v1. 

However dealing with a VCS like git, this thought occur to me, what would it look like when the project needs to move from v1 to v2?

In git we can rename the file in the entire history which messes things up for everyone so that's a bad thing to do. 

We can also copy paste v1 into the v2 folder, but then we lose the entire history of the files and would need to refer to the history of the other folder so this is a huge pain.

We can also create a dummy repository that just submodules v1 and v2 to the actual source repository but this isn't pretty and feels kinda hackish.

I also thought maybe do all development on v0 and then have v1/v2/etc... just be wrappers to v0.

The last one makes the most sense to me, so I guess question is, is there a better way to handle multiple API version paths? Especially when there are many sub packages in the same repository. 

[Sam Whited]

On Wed, Feb 21, 2018, at 22:35, alex....@gmail.com wrote:
> vgo suggested that package developers put the major version into the import
> path e.g. foo/v1.
> However dealing with a VCS like git, this thought occur to me, what would
> it look like when the project needs to move from v1 to v2?

This wasn't clear to me either, but it was pointed out that you can just change the "module" line in your go.mod file, eg.

module "example.net/mypackage/v2"

and you don't have to actually create a /v2 tree or move your code around at all.

—Sam

[alex....@gmail.com]

I don't think changing the module name would be a good idea for singletons tho. 

And having a v0 has the added benefit of using newer codes while maintaining backwards compatible API.

So no more backporting and all the mess of having multiple branches.

[Axel Wagner]

On Thu, Feb 22, 2018 at 7:18 AM <alex....@gmail.com> wrote:

I don't think changing the module name would be a good idea for singletons tho. 

Have you read https://research.swtch.com/vgo-import? It talks about singletons and also how to solve that with different import paths.

Note, that this is also independent of *how* the different import paths are represented and distributed; for singletons it only matters what import paths the compiler sees, whether they are committed physical directories or whether the vgo tool automatically resolves them.

 

And having a v0 has the added benefit of using newer codes while maintaining backwards compatible API.

So no more backporting and all the mess of having multiple branches.

If you don't want multiple branches, you can commit your version-directories and keep the same module name for all of them. If you prefer having a single directory and different branches for different version, you can have the version in the module line instead.

To the compiler, both are equivalent, the tool will resolve them to the same thing.

 

On Thursday, February 22, 2018 at 1:49:53 PM UTC+8, Sam Whited wrote:

On Wed, Feb 21, 2018, at 22:35, alex....@gmail.com wrote:
> vgo suggested that package developers put the major version into the import
> path e.g. foo/v1.
> However dealing with a VCS like git, this thought occur to me, what would
> it look like when the project needs to move from v1 to v2?

This wasn't clear to me either, but it was pointed out that you can just change the "module" line in your go.mod file, eg.

    module "example.net/mypackage/v2"

and you don't have to actually create a /v2 tree or move your code around at all.

—Sam

--
You received this message because you are subscribed to the Google Groups "golang-nuts" group.
To unsubscribe from this group and stop receiving emails from it, send an email to golang-nuts...@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[alex....@gmail.com]

On Thursday, February 22, 2018 at 5:29:23 PM UTC+8, Axel Wagner wrote:

Have you read https://research.swtch.com/vgo-import? It talks about singletons and also how to solve that with different import paths.

Note, that this is also independent of *how* the different import paths are represented and distributed; for singletons it only matters what import paths the compiler sees, whether they are committed physical directories or whether the vgo tool automatically resolves them.

Yup, the package I'm working on really cannot have 2 instances even in different import paths as it deals with a shared resource and code that must be run on the main thread.

So I have to choose between daisy chaining v1 to v2 to v3 etc.. or have a central v0 where all the v1, v2, etc.. packages import.

Daisy chaining would mean I would only have to code API translations for the latest API but then it's debug hell and if one version in the chain breaks, 

it means fixing all the newer versions. Also there's a performance hit going through many translations.

Having a v0 means non of all the daisy chaining problems but it means more work when there's a breaking change as then I'll have to update all versions.

Also it means that all version packages have to be from the same release or it breaks. 

If I just have one module at the root, is it right to assume all sub packages would be of the same commit and that there will be only one version of the module?

[Axel Wagner]

On Thu, Feb 22, 2018, 12:43 <alex....@gmail.com> wrote:

On Thursday, February 22, 2018 at 5:29:23 PM UTC+8, Axel Wagner wrote:

Have you read https://research.swtch.com/vgo-import? It talks about singletons and also how to solve that with different import paths.

Note, that this is also independent of *how* the different import paths are represented and distributed; for singletons it only matters what import paths the compiler sees, whether they are committed physical directories or whether the vgo tool automatically resolves them.

Yup, the package I'm working on really cannot have 2 instances even in different import paths as it deals with a shared resource and code that must be run on the main thread.

So I have to choose between daisy chaining v1 to v2 to v3 etc.. or have a central v0 where all the v1, v2, etc.. packages import.

Both seem fine.

Daisy chaining would mean I would only have to code API translations for the latest API but then it's debug hell and if one version in the chain breaks, 

it means fixing all the newer versions. Also there's a performance hit going through many translations.

I don't believe so. There's may be an increase in compile time though.

Note also, that you don't have to Daisy-Chain down - AIUI you can daisy chain up too. i.e. v1/singleton aliases into v2/singleton, aliases into v3/singleton... That way any cost would only be paid by people on stale versions. Yes, if vN breaks, that would also break vM for M<N, but IMO such is life, wenn you have different packages depending on one common one. But you wouldn't have to touch intermediate versions either.

In general, this seems like a problem that should just be mostly avoided - if you absolutely have to have something like that, isolate the Singleton logic into a single, isolated package and try not to touch it, ever. And also, of course, don't constantly release new major versions.

The real question though, is how other package management semantics would solve this better. Is this actually accidental complexity or just systematic complexity caused by "I need to both break compatibility *and* have different versions coexist *and* still share Singleton state between them?

Having a v0 means non of all the daisy chaining problems but it means more work when there's a breaking change as then I'll have to update all versions.

Also it means that all version packages have to be from the same release or it breaks. 

If I just have one module at the root, is it right to assume all sub packages would be of the same commit and that there will be only one version of the module?

--

[alex....@gmail.com]

On Thursday, February 22, 2018 at 9:53:20 PM UTC+8, Axel Wagner wrote:

Daisy chaining would mean I would only have to code API translations for the latest API but then it's debug hell and if one version in the chain breaks, 

it means fixing all the newer versions. Also there's a performance hit going through many translations.

I don't believe so. There's may be an increase in compile time though.

I was thinking of cases where the behavior is so different it needed to emulate old behavior so over time all the emulation adds up and would be more complex than just emulating based on v0/latest.

 

The real question though, is how other package management semantics would solve this better. Is this actually accidental complexity or just systematic complexity caused by "I need to both break compatibility *and* have different versions coexist *and* still share Singleton state between them?

The vgo blogs talk about coexistence of multiple versions so was just figuring out if it makes sense and how would that look like.

I just like to be prepared for worst case scenarios, so stuff I was worried about may or may not actually happen. But better be safe than sorry.

[Axel Wagner]

Sure, I'm just saying that if you start with how you'd do this in a more familiar packaging scheme, you might be able to transfer it to vgo - I'm not convinced vgo makes this at all harder, it's simply a hard problem to begin with.

[David Collier-Brown]

On Thursday, February 22, 2018 at 6:43:07 AM UTC-5, alex....@gmail.com wrote:

Yup, the package I'm working on really cannot have 2 instances even in different import paths as it deals with a shared resource and code that must be run on the main thread.

So I have to choose between daisy chaining v1 to v2 to v3 etc.. or have a central v0 where all the v1, v2, etc.. packages import.

Daisy chaining would mean I would only have to code API translations for the latest API but then it's debug hell and if one version in the chain breaks, 

it means fixing all the newer versions. Also there's a performance hit going through many translations.

If I understand you correctly, v2's foo() has to call v1's foo() in the daisy-chain case.   This was one of the two common cases in Multics (and Solaris, and Linux glibc).  The other case was v1's foo() calling a wrapper around v2's foo().

We called these "updaters" and "downdaters", and implemented them in the same library, under the covers.

Feel free to drop me a line if you are interested, and see also https://leaflessca.wordpress.com/2017/02/12/dll-hell-and-avoiding-an-np-complete-problem/ fore some of the background.

--dave

dav...@spamcop.net