On Tue, May 28, 2013 at 08:33:22AM -0400, Matthew Miller wrote:
I was asked (with my Red Hat hat on) to put together a little report
on
areas in Fedora which could be improved with an investment in better
automation. From what I'm working on myself, I'm aware of the gigantic need
in the cloud images production process, and I've been keeping an eye on Tim
Flink's autoqa revamp ideas. I expect there are others, because I know from
my previous jobs that there's always a balance between building
condiment-passing machines and just _passing the salt_. [1] Are there other
things which could be made better if only someone came up with the spare
time and resources to do the work?
Hi!
Some observations/ideas:
1) The packager workflow is pretty tedious. There has been some
improvement to it, but more can be done. Things like
fedora-review and fedora-create-review (and bodhi!) are a huge
help. But there are plenty of inefficient "blocking" points in the
process.
For instance, once a new package is approved, only then does the
submitter declare what branches they want with an scm admin
request. They then wait for an scm admin to declare that they
have created their branches, and then wait for a cronjob to run
that gives them permission to push on those branches (manually).
They then wait for their koji builds to finish to (manually) submit
bodhi updates.
It would be nice if we could automate that whole process -- once a
package is approved, if there were a "make-it-so" button that
required no further intervention from the packager (but still
required the keen eye of an scm admin).
There are further sequences down the pipeline like requesting that
packages in testing be pushed to stable, but there are good
arguments against automating that.
2) Continuous deployment for infrastructure. It has been tossed
around in IRC, possibly at FUDCon as well. If application
developers could "git push" on the develop branch and have those
changes automatically roll out to our staging infrastructure --
that would save a lot of time. Packaging our apps, building rpms,
signing them, copying them to our infra yum repos, rebuilding those
repos, clearing the cache on the target machines, performing a yum
update <-- that process is cumbersome.
I suspect that the "release only when we have accumulated enough
changes to warrant enduring the burdensome release process" mode
of deployment (as opposed to "release early, release often") also
poses somewhat of a barrier to new contributors. They contribute
a patch.. nice! When does it go live? When one of our
overstretched sysadmin-mains can get around to it (it is required
that one of them sign the package).
Caveat #2.1:
There are some ways around this. Individuals can get around the
requirement of having a sysadmin-main touch their test release by
installing their rpm directly on the target machine. They still
have to jump through some hoops to make it happen.
Caveat #2.2:
This is one of the reasons we put so much work into our private
cloud (dev nodes). There is no barrier there for teams to set up
their own continuous deployment mechanism. This meets most needs,
but we don't have a way to iterate rapidly on some of the more
important pieces of our infrastructure. Apps/services that
interact with each other don't quite work out on isolated cloud
nodes. The bodhi masher? Koji? fedmsg? mirror manager? We
can't necessarily test those on dev nodes (and some we can't test
in staging -- resolving this down the road would save some
headaches).