[LLVMdev] [Zorg] Reorganisation, documentation and other issues

[Daniel Dunbar]

My $.02


On Wed, Aug 6, 2014 at 3:52 PM, Dan Liew <dan at su-root.co.uk> wrote:

> Hi All,
>
> Recent conversations [1][2] would suggest we need to rethink zorg or
> at the very least improve the documentation.
>
> I've CC'ed Daniel Dunbar, Galina Kistanova and Duncan Sands because it
> seemed that they have contributed the most code to Zorg
>
> Organisation
> =========
>
> The current organisation of the Zorg repository doesn't make a huge
> amount of sense from my perspective. There are two things that we
> should be able to do with this code
>
> - Run the currently in use buildbot master.
> - Run a builtbot master and slave on a local machine for testing.
>

Realistically, these are pretty challenging goals I think. Running buildbot
is very challenging and the tool itself just isn't designed well for this
use case. My feeling on this is that if people are motivated enough to
actually do this, then figuring out zorg isn't going to be the limited
factor. Of course if you have improvements in mind I would encourage you to
commit them, I just don't think it is worth making it a priority.

>
> The current project layout doesn't seem to facilitate either of these
> needs particularly well. If we take a look at
> ``/buildbot/osuosl/master`` the ``zorg`` folder isn't in here so the
> ``master.cfg`` adds a non portable hack to load the zorg code (assumes
> zorg is in user's home directory).
>
> Unless there is a good reason for doing so I think the zorg modules
> should be moved into the master buildbot directory so this hack isn't
> necessary. If there is a really good reason for doing this we should
> document this so people can actually run the buildbot master on their
> own machine.
>
> The ``master.cfg`` file is also full of stuff that is not suited for
> testing a master and slaves hosted on the same machine.
>
> * Buildbot master should only listen on localhost (no out going
> connections)
> * Buildbot master should have no authorisation and show the web
> interface so manual builds are easy to trigger
> * Buildbot master should not use the large list of slaves because a
> locally hosted master cannot connect to these because they shouldn't
> be contacted for testing (it doesn't have the passwords for these
> buildbot slaves either).
> * Buildbot should load slaves from a file not under version control
> for local testing.
>

This is all fine with me.


> Either we need to make it so that the buildbot master
> (/buildbot/osuosl/master) can operate in two modes (normal and
> localhost testing) or we need to provide another buildbot master
> configuration for localhost testing. I think this is what
> ``buildbot/smooshlab`` is for but that code has not been touched in
> ages and is probably dead (if it is dead, we should delete it).
>

smooshlab was a different configuration, not just one for testing. It can
be removed.


>
> Other issues
> ==========
>
> * Ancient version of buildbot.
>
> Why do we use buildbot 0.8.5? Newer versions are available in Pypi
> which anyone using Python can install. I have issues with 0.8.5 and I
> have to patch it before it will actually work. I don't think we should
> have in our documentation that you need to patch buildbot, we should
> just upgrade to a newer version.
>

I have no opinion on this. In the past I know we have hit issues moving to
newer versions of buildbot, so it just means someone taking responsibility
for testing and updating.


>
> * reload()
>
> I see lots of reload() all over the place. Why do we need these?
> Surely if you do
>
> $ buildbot stop master
> $ buildbot start master
>
> This would reload everything making uses of ``reload()`` pointless.
>

These are necessary and useful for when you run "buildbot reconfig", which
is a very useful way of working esp. when dealing with a production setup
where you don't want builders to stop.

The reloads are necessary because reconfig just reloads the data inside the
running instance, so any modules need to be reload()ed, and because of how
imports work, there is a dependency issues between every module you want
reloaded. There are cleaner ways to solve this problem (like trying to auto
reload by inferring the appropriate dependencies and scanning sys.modules).


>
> * KLEE stuff
>
> KLEE is no longer actively developed as part of LLVM so all this stuff
> should be dropped from Zorg.
>

Sure, I don't know of anyone using this.


>
> * Lack of documentation
> This can be addressed once we fix the other issues.
>
> * Builders
>
> There are lots of python modules in ``zorg/zorg/buildbot/builders``.
> These are Factories, not builders so why do they have ``Builder`` in
> their name and are in the "builders" folder? This is unnecessarily
> confusing.
>

Yes, renaming this would make sense.


> * llvmlab, osuosl, smooshlab?
>
> Not sure those names are particularly helpful. I suspect ``smooshlab``
> will probably be removed entirely and I suspect osuosl ought to be
> renamed. I'm not sure if we need to keep any of the llvmlab stuff.
>

llvmlab -- instance on lab.llvm.org
osuosl -- old place we used to run buildbots
smooshlab -- an internal setup that doesn't need to remain in the zorg repo



> This is probably an incomplete list of issues but theses are issues I
> have observed in my brief time playing with Zorg and buildbot.
>

Generally speaking, there just isn't anyone who is really owning and
maintaining the buildbot setup. If you want to take it over and clean it
up, I don't think anyone is going to fight with you over it. :)

My experience after maintaining the buildbot stuff for a long time was that
I wasn't particularly happy with buildbot itself and how it all played out.
If I had to do it over again, I would probably move more in the direction
of having the CI system run build scripts that are checked in to the
individual project repositories. If people needed to reproduce a CI build,
the recommended advice would then be to run those scripts directly, and
those scripts could be written in languages people are comfortable with,
not the complicated buildbot scripting-by-instantating-factory-steps-setup.
This seems to be the direction Duncan went with the Dragonegg scripts.

HTH,
 - Daniel




> [1] http://lists.cs.uiuc.edu/pipermail/llvmdev/2014-August/075287.html
> [2] http://lists.cs.uiuc.edu/pipermail/llvmdev/2014-August/075348.html
>
>
> Thanks,
> Dan.
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20140807/2f1aab6a/attachment.html>