On Thu, Oct 12, 2017 at 4:01 PM, Norbert Hartl <norbert@hartl.name> wrote:
> Am 12.10.2017 um 14:23 schrieb Torsten Bergmann <astares@gmx.de>:
>
> Hi,
>
> today there was an interesting effect visible for the Pharo build on CI for Pharo 7:
>
> Esteban merged three contributions for Pharo 7 into #development (see [1]). All three merges on git triggered the CI.
> For each a separate build for Pharo 7 started on CI independently.
>
> While usually one would expect the builds in regular build order 183, 184, 185 the CI job for building 185 job was faster
> (as you can see on the attached screenshot).
>
> So the 185 series of images was copied first to the file server [3] because the job was already finished.
> Afterwards 183 build completed on CI and these images were copied to the file server with a newer timestamp.
>
> So you have the order 185, 183 with 183 having the files with the newest timestamp (but not being the most recent state of Pharo).
>
> This leads to several unfortunate effects:
>�� 1. When you sort on file server [3] descending by date the build 183 images are newer than the build 185 images
>�� 2. When you use PharoLauncher to download "latest-32" you get an 183 image instead of the "real latest" 185 image
>�� 3. When you ask Zeroconf to download the latest using http://get.pharo.org/70+vm you also get 183 instead of the already
>�� �� ��available 185
>
> So as the CI build of an older "update" (here 183) finished later than the most recent version (here 185)
> you get 183 as "latest" instead of 185.
>
> Any ideas how we could avoid such a side effect? Maybe a chain on the builds so 183 has to finished before 184 and 185 could be
> built?
>
I don't understand why it is working like this. Usually only build for a job is active, others will be queued. So the problem here is that
the jenkins spawns more builds in parallel for the same job.
Yes, the thing is that the same job is being used to build the development branch and the pull requests.
If you think about it, this is the same setup as in travis.
The thing is that this does not bother me too much, considering the tradeoffs.
- I want pull request to run in parallel.
- I want to have less things to maintain :)
Now,��
��- we could split the job in two (integration vs pr validation). But this requires setting up a separate jenkins because otherwise they share the same slave configuration :/. And there is no way in the actual state of the plugins we use to limit the number of parallel builds...
��- we could try to make a build wait until a build with lower build number is building...
��- we could also forbid Esteban to push on the merge button while a build is running
��
But doing any of those is a lot of work...
- Moreover, I think the problem is that we are using the file server as something it is not. The file server is... a file server.
Ideally, I'd like to attach meta-data to a release/build artifact, and be able to browse/search/sort using that metadata. Something alike artifactory or bintray.
- Also, I don't know how the pharo launcher is getting the latest image, but if instead of using the file timestamp it uses the build number, then this may be fixed, isn't it?
��
Norbert
> Thx
> T.
>
>
>
> [1] https://github.com/pharo-project/pharo/commits/ development
> [2] https://ci.inria.fr/pharo-ci-jenkins2/job/Test%20pending% 20pull%20request%20and% 20branch%20Pipeline/job/ development
> [3] http://files.pharo.org/image/70/
> <buildOrder.png>
|
������ | Guille Polito Research Engineer Centre de Recherche en Informatique, Signal et Automatique de Lille CRIStAL - UMR 9189 French National Center for Scientific Research - http://www.cnrs.fr Phone: +33 06 52 70 66 13 |
