Hi,
On Tue, Jul 2, 2019 at 1:21 AM Miroslav Vadkerti <mvadkert(a)redhat.com>
wrote:
Hi,
This is awesome! Looking forward meeting you on Flock to discuss where
this is going.
Sure.
On Mon, Jul 1, 2019 at 2:53 PM Fabien Boucher
<fboucher(a)redhat.com> wrote:
> Hello,
>
> I'm writing to this list to share about a proof of concept me and my team
> are setting up. The goal of the POC is to show that Zuul, OpenStack's
> code gating system [1], can be used to automate RPM-related jobs for
> Fedora's CI.
>
> We've managed to implement the following workflows:
>
> * Run jobs when PRs are opened/changed on a Pagure instance. Recently
> Zuul added support for Pagure events so we were able to set up some
> packaging jobs on Zuul triggered by PRs on
src.fedoraproject.org. There
> are three jobs attached:
> ** The first job runs a scratch build on Koji and shares artifacts
> (rpms) with two child jobs.
>
Is this really build on Koji, or you build it in mock on your nodes?
https://fedora.softwarefactory-project.io/logs/5/5/616ac07a47e38483d6bbf6...
The SRPM is built on the job's node in a mock then the built SRPM is passed
to koji. See below:
https://fedora.softwarefactory-project.io/logs/5/5/616ac07a47e38483d6bbf6...
The ARA report is great to get an overview of the job flow:
https://fedora.softwarefactory-project.io/logs/5/5/616ac07a47e38483d6bbf6...
** The first child job runs a rpm lint.
> ** The second child job runs the functional tests included in the package
> .
>
> Here [2] you can see a PR where those Zuul jobs ran. More details can be
> found here [3].
>
> * Run jobs when PRs are opened/changed that depend on others PRs. A
> packager is able to tell Zuul that a PR depends on others PRs by using
> the "Depends-On:" keyword in the PR's initial message. The artifacts
> from dependent PRs are available on a PR's job's workspace as well; in [4]
> the python-mock RPM from the dependent PR is used when building
> python-redis.
>
Very cool feature! I am glad to see Zuul support for Pagure.
>
> * Run jobs when a package is built on Koji. The idea is to listen to the
> fedmsg queue, waiting for the topic "buildsys.build.state.change", run
> the configured Zuul jobs, and finally publish the jobs' results on
> fedmsg. At the moment we are able to run a linter job each time a
> package is built, but we don't publish the results on fedmsg. Here is
> the list of the most recent linter job results [5]. More details can be
> found here [6].
>
Zuul-gateway seems to be quite a workaround, is the long term plan to
change to support event based triggers "directly" ?
Yes it is a workaround. We have started a discussion about an AMQP trigger
with Zuul's maintainers
http://lists.zuul-ci.org/pipermail/zuul-discuss/2019-May/000929.html. But
no real plan for the moment.
> We started that POC because we think Zuul offers a lot of useful features
> [1] like handling PR dependencies that could be of interest to the
> Fedora project. The POC has shown some working workflows but now we need
> insights from people who know what the needs of Fedora's CI are. Let us
> know your thoughts, and we will be happy to extend the POC and onboard anybody
> wanting to experiment with Zuul in that context.
>
My Testing Farm team would like to experiment using Zuul for our workflows.
Sure, let's us know how we could help.
Does Nodepool support prioritization of different drivers? We would
like
to offload some of our workloads to AWS in case of spikes, but would like
to use it only if non-paid resources (Openstack, Openshift) had been
exhausted.
AFAIK not out of the box. But with an external process you might be able to
detect your are out of quota on the OpenStack provider and push in the
Nodepool providers configuration the setting AWS:max-server value from 0 to
something else. Then Nodepool will start to use the AWS provider. When
spikes stop you revert the max-server setting to 0 then Nodepool will stop
spawning nodes on AWS.
How is the driver selected for a job? Our testing jobs describe an
environment which we would like the CI system to translate to a specific
driver, according to an predefined priority.
A job definition in Zuul needs to specify a nodeset. A nodeset is a
collection of nodes for the job's Ansible inventory. Each node must match a
Nodepool node's label. In other words in Nodepool you define the nodes
details (provider type (openstack/aws/...), base image, VM flavor) and set
labels to nodes. In Zuul, via a nodeset, you define the available nodes
(using the Nodepool labels) for the job's inventory. For instance you could
have the same label accross multiple providers. However there is not any
priority system in that mechanics.
https://zuul-ci.org/docs/zuul/user/config.html#nodeset
How hard would it be to support non-Ansible based jobs please?
Difficult to respond w/o more details, Zuul only understand a pre-defined
job format
https://zuul-ci.org/docs/zuul/user/config.html#job so a
migration tool would need to be written.