On Wed, 10 Jun 2020 at 05:10, Adrian Reber <adrian@lisas.de> wrote:
On Thu, Jun 04, 2020 at 07:02:41PM -0700, Kevin Fenzi wrote:
> On Mon, Jun 01, 2020 at 04:18:35PM +0200, Adrian Reber wrote:
> > Our MirrorManager setup exports the current state of all mirrors every
> > hour at :30 to a protobuf based file which is then used by the
> > mirrorlist servers to answer the requests from yum and dnf.
> >
> > The Python script requires up to 10GB of memory and takes between 35 and
> > 50 minutes. The script does a lot of SQL queries and also some really
> > big SQL queries joining up to 6 large MirrorManager tables.
> >
> > I have rewritten this Python script in Rust and now it only needs around
> > 1 minute instead of 35 to 50 minutes and only 600MB instead of 10GB.
>
> Wow. nice!
>
> > I think the biggest difference is that I am almost not doing any joins
> > in my SQL request. I download all the tables once and then I do a lot of
> > loops over the downloaded tables and this seems to be massively faster.
> >
> > As the mirrorlist-server in Rust has proven to be extremely stable over
> > the last months we have been using it I would also like to replace the
> > mirrorlist protbuf input generation with my new Rust based code.
> >
> > I am planing to try out the new protobuf file in staging in the next
> > days and would then try to get my new protobuf generation program into
> > Fedora. Once it is packaged I would discuss here how and if we want to
> > deploy in Fedora's infrastructure.
>
> Cool. You will need to hurry as staging goes off on monday, and back in
> a few weeks. :)

Then I just have to wait a bit. No problem.

> > Having the possibility to generate the mirrorlist input data in about a
> > minute would significantly reduce the load on the database server and
> > enable us to react much faster if broken protobuf data has been synced
> > to the mirrorlist servers on the proxies.
>
> Yeah, and I wonder if it would let us revisit the entire sequence from
> 'update push finished' to updated mirrorlist server.

Probably. As the new code will not run on the current RHEL 7 based
mm-backend01 would it make sense to run a short running service like
this on Fedora's OpenShift? We could also create a new read-only (SELECT
only) database account for this.


We can do that or we can make a mm-backend02 which is el8 for this


--
Stephen J Smoogen.