Configuration readonly archives

Salut Francois
> We've been bitten a few times at Ta Mère by the problem of changing
> metacello configuration.

I agree. This is why I also think that released should point to version
and not symbolic version.
What I do is that we can use versionner to release more often
configurations pointing to the versions.
Christophe will continue to improve it.
We are discussing in the team
- I was thinking that each project/package should have its own
configuration saved in his repo.
     -- then the developer may choose to publish it to a central place.

- other are telling to me that other language have a central place.
This is ok too.
I would like to have a catalog of frozen versions per pharo version. Now
this is two years that we request an engineering
to build a project validator and a catalog but we failed.

Stef

Hi François

It sounds like you need snapshots. That has nothing to do with Metacello.

Stephan

Stef wrote
>I agree. This is why I also think that released should point to version
>and not symbolic version.

I strongly disagree. A configuration should only change when
there is a meaningful change. Changes in dependencies that do not
change relevant behavior should not propagate.

We achieve that by using #'releasex.x' style versioning. Changes that
are not supposed to change observable behavior and bug fixes get
new numbered versions within the release.
#development versions are updated in place.

In Moose we  did not encounter a situation where a month-old numbered
version configuration was still loadable in the latest Pharo.

There is a use case for snapshots of exact versions. They don't belong
in configurations.

Focussing improvements in Versioner on releasing more often numbered
versions is counter productive. It just generates more work to keep
configurations working.

I have been fixing lots of configurations on ci, and nearly all problems
have to do with not using #'releasex.x' style symbolic versioning.
There are two problems:
- Configurations using numbered versions. They break as soon as a  
  dependent project changes functionality or does a bug fix, often as
  a result of underlying code in Pharo changing.
- Configurations using #stable, where the dependent project does
  a new release.

Now Seaside, Grease and Magritte have switched to this style of
versioning, the number of problems has reduced drastically.

Stephan

We will discuss that on a black board at ESUG.
A release is a release and when you release code you do not want to rely
on something that somewhere could change.
Christophe told me that they are doing that in Java world and this is
working.

Now for development you load your baseline and you load symbolic version
of foreign code.
Stef
This is the purpose of symbolic versions to avoid propagation.

>
> In Moose we  did not encounter a situation where a month-old numbered
> version configuration was still loadable in the latest Pharo.
???

>
> There is a use case for snapshots of exact versions. They don't belong
> in configurations.
I do not get what you mean.
Now if I want to deliver exactly the same version I do not see how we
can avoid versions.
And this "nothing observable changes" is a nice idea but what I see is
that often pharo breaks
because of a nothing observable change.

>
> Focussing improvements in Versioner on releasing more often numbered
> versions is counter productive. It just generates more work to keep
> configurations working.
I do not think so
     - in dev you use baseline for you and development for foreign code
     - when releasing you freeze.
With nice tool support I do not see why this would not work.

>
> I have been fixing lots of configurations on ci, and nearly all problems
> have to do with not using #'releasex.x' style symbolic versioning.
I did not get it so you should explain it better :)

> There are two problems:
> - Configurations using numbered versions. They break as soon as a
>    dependent project changes functionality or does a bug fix, often as
>    a result of underlying code in Pharo changing.
> - Configurations using #stable, where the dependent project does
>    a new release.
Most of the projects on Jenkins are not there to work against release
but day to day changes
in this case they should load latest.

>
> Now Seaside, Grease and Magritte have switched to this style of
> versioning, the number of problems has reduced drastically.
>
> Stephan
>
>
>
>

BTW an important point.
We cannot load any moose versions of the past (except the ones we
snapshotted), to me this is a sign of problems.
never be able to get back in a previous state is a problem.

Stef

I'll be in Lille tomorrow. We could have a beer instead of a blackboard ;)

Stephan