repospanner and our Ansible repo

[Randy Barlow]

Greetings!

Kevin asked me last week whether we are ready to move our
infrastructure Ansible repository into repospanner. The benefit of
moving it into repospanner is that it is one way to enable us to allow
pull requests into the repository, which I think would be nice.

repospanner seems to work correctly as a git server, but it does need
improvements in its performance, so I offered to do a little
benchmarking with our Ansible repo and repospanner to see what kind of
performance we might see.

I deployed a 3-node repospanner cluster today on fairly high
performance hardware (SSD storage). It was three VMs on the same
physical machine. Note that due to my test setup, network latency was
about as good as it could get, and so was storage iops. I believe the
performance bottlenecks will depend heavily on storage iops. Thus, this
hardware is not really a great way to predict how the performance might
be if we deployed into our infra, but it was easy for me to do and get
a "best case" performance benchmark. I am willing to attempt to
replicate this test on more realistic hardware in our infra if we want
more realistic data for our own use case.

I pushed the Ansible repository into it. This took a very long time:
298m2.157s! If we were to deploy nodes in different geos and use NAS
storage, I believe this would take longer. The good thing is that we'd
only need to do this operation once, if we were to decide to proceed.

The next test was to see how long it takes to clone our repo. I did
this on another machine on the same LAN (so again, ideal network
latency) and it took 2m27.433s. That's a pretty long time too I'd say,
but maybe liveable? This would impact every contributor who wanted to
clone us, so I'll let the list debate whether that is acceptable.

Next, I made a small commit (just added/deleted some lines) and pushed
it into the cluster. This went reasonably quick at 0.366s, which I
think we would be OK with.

The last test I performed was to see how quickly another checkout could
pull that commit, and this was again a speed I might consider to be a
bit slow at 4.931s, especially considering that the commit was small
and was only one. I would expect this to be somewhat proportional to
the amount of change that has happened since the user last fetched, and
this repo does see a lot of activity. So I might expect git pull to
take 10's of seconds for contributors who are fairly active and pull
once every few days or so, and maybe longer for users who pull less
frequently.

The repo copy I tested with has 199717 objects and 132918 deltas in it.
repospanner performance seems to be fairly proportionally correlated
with these numbers, as the bodhi repo pushed into it in about an hour
and has 50kish objects, iirc (didn't write it down, so from memory).

I personally am on the fence about whether we should proceed at this
time. I am certain that people will notice the speed issues, and I also
expect that it will be slower than the numbers I listed above since my
tests were done on consumer hardware. But it would also be pretty sweet
if we had pull requests on the repo.

Improving repospanner's performance is a goal I am focusing on, so if
we deployed it now I would hopefully be able to get it into better
shape soon. Alternatively, we hopefully wouldn't have to wait that long
if we wanted to wait for performance fixes before proceeding. I could
see either decision being reasonable.

To reiterate, I'd be willing to replicate the tests I did above on
infra hardware if we are on the fence about the numbers I've reported
here and want to see more realistic numbers to make a final decision. I
think that would give us more realistic numbers since the tests I did
here were on a much more ideal situation, performance wise.

What do others think?