[libvirt PATCH 7/7] ci: integration.sh: Define the SCRATCH_DIR variable for local execution

[Erik Skultety]

On Mon, Jan 23, 2023 at 11:42:52AM +0100, Martin Kletzander wrote:
> On Mon, Jan 23, 2023 at 10:50:31AM +0100, Erik Skultety wrote:
> > On Mon, Jan 23, 2023 at 09:41:10AM +0100, Martin Kletzander wrote:
> > > On Fri, Jan 20, 2023 at 04:06:03PM +0100, Erik Skultety wrote:
> > > > Running outside of GitLab will likely not have the variable set and
> > > > hence the execution would fail.
> > > >
> > > > Signed-off-by: Erik Skultety <eskultet at redhat.com>
> > > > ---
> > > > ci/integration.sh | 8 ++++++++
> > > > 1 file changed, 8 insertions(+)
> > > >
> > > > diff --git a/ci/integration.sh b/ci/integration.sh
> > > > index 41326d6e40..ac04c46d8e 100644
> > > > --- a/ci/integration.sh
> > > > +++ b/ci/integration.sh
> > > > @@ -37,6 +37,14 @@ virsh net-start default &>/dev/null || true
> > > > # END AS ROOT
> > > > exit
> > > >
> > > > +# If we're running outside of GitLab, this variable will likely not exist, so
> > > > +# we need to define it and create the scratch directory
> > > > +if [ -z "$SCRATCH_DIR" ]
> > > > +then
> > > > +    SCRATCH_DIR="/tmp/scratch"
> > > > +    mkdir "$SCRATCH_DIR" 2>/dev/null
> > > 
> > > This could fail if someone has this directory already.  Which is a good
> > > thing as otherwise it could override some of it.  But wouldn't it be
> > > nicer to use mktemp -d and print the result?
> > 
> > Although an option, the main motivation here to remain consistent with how it
> > works in GitLab. Since we define SCRATCH_DIR under the 'vars' section, IIRC you
> > can only use a scalar value, not a command (if we can, I retract my argument)
> > and hence we'd have to export and define the variable under each script,
> > before_script, after_script sections.
> > 
> 
> I don't really understand how that affects a change from:
> 
> SCRATCH_DIR="/tmp/scratch"
> mkdir "$SCRATCH_DIR"
> 
> to something like
> 
> SCRATCH_DIR=$(mktemp -d)

Simple, ^this is not consistent and results in a different environments.

> 
> or possibly
> 
> SCRATCH_DIR=$(mktemp -d "/tmp/scratch.XXX")

^This one is close enough, I'm fine doing that, but again, one expects that the
directory will be in /tmp/scratch and it isn't. We can keep arguing about "you
can just hit tab-tab in a shell", or "that association is obvious to anyone",
or "any engineer who wishes to debug libvirt must be able to figure out what
the correct directory is". My only argument was about consistent and uniform
user experience. However, the deal breaker here kinda supporting your
suggestion and where my original proposal fails is quite different actually -
not all platforms actually clean /tmp on reboots, e.g. CentOS Stream - in this
particular case it will be done with systemd-tmpfiles-clean timer and service,
other platforms might employ a different mechanism, but the point is, if it's
not mounted as tmpfs, the reboot guarantee isn't there and hence we could have
a left-over directory from a previous run.

However, having said that, rather than ending up with multiple /tmp/scratch.XXX
directories where you actually have to verify which one is it that you want to
interact with, I'd still suggest to preventively remove the directory and
re-create it. Why? Because the proposed scenario is to actually get a throwaway
VM with an "identical" (it won't be because of package versions) environment to
that running in GitLab and we can even document this behaviour so that these
VMs are created on-demand locally rather than re-using them which also leads to
other potential problems.

So, given that I document the recommendation wrt creating throwaway VMs, would
you agree to:

SCRATCH_DIR="/tmp/scratch"
if [ -d $SCRATCH_DIR ]
then
    rm -rf $SCRATCH_DIR
fi
mkdir "$SCRATCH_DIR"


Erik