On Sun, 2021-03-14 at 13:47 -0700, Kevin Fenzi wrote:
> The eln composes (at least as far as I know) are done via ODCS
> (on demand compose service) and are already available on the master
> mirrors:
>
> https://dl.fedoraproject.org/odcs/production/
> (Although not via rsync currently).
>
> The big problem with mirroring them back in the past was that they
> changed too quick. There's a compose every 3 hours I think. That
> wouldn't be nearly enough time for our mirror network to keep up.
>
> Given that I expect the number of people who would sync this content is
> so small, perhaps we could just leave them on master mirrors?
> (we can enable rsync if you want... just put in a ticket).
> If that proves to be too much load, we could perhaps try and sync them
> to a s3 bucket or some other location? I just dont think our normal
> mirror network would be a good fit here.
Thanks! I'm ok with consuming these from the master mirror. I've filed
https://pagure.io/fedora-infrastructure/issue/9730 to get rsync enabled
for that endpoint. Once that's sorted out I'll setup a periodic sync,
(daily or weekly, at least at the beginning), and republish the
composes on https://mirror.facebook.net.
setting up rsync for this point will be a difficult problem and may need to be delayed for a long time.
1. ELN recomposes every 3-4 hours. The /srv/odcs is 4.8 TB and I think it is on SATA disks. This means that an rsync may take longer than the compose is alive, and rsync's will fail a lot with stale files most of the time.
2. If you are only doing it every day, then you are going to be 21 to 20 hours off from 'new' at multiple times during the week.
3. The existing http/rsync traffic for
dl.fedoraproject.org has already maxed out our 1 gigabit link already. We are actually having to look to making all our rsyncs a 'hidden' mirror for only tier-1 mirrors.
We are supposed to get an increased bandwidth to the site later this year but it is not certain when or how much we will have. Our download/mirror system also probably needs a rethought as it was essentially designed for a 1999 world and we are now in a 2020 network schema.