[Dev] [FYI] changes to [repo] HTTP
Luke Shumaker
lukeshu at sbcglobal.net
Fri Jul 29 21:24:34 GMT 2016
I've made some changes to how repo handles HTTP requests. I'm taking
bets on how long before someone complains about it.
TL;DR: repomirror is now prime-time; anyone who tries to hit
repo.parabola.nu will now be redirected to repomirror.parabola.nu by
default. repomirror will then redirect to a randomly-selected
suitable mirror.
What I did, in order:
- Notice that everything is slow and terrible.
- Performance-tune nginx a bit (use epoll, et c.)
- Observe that things are only slightly less terrible.
- Run iotop to see if it gives any insights to why disk wait is so
high.
- Notice that unionfs is dominating the system IO
- Configure nginx to bypass unionfs, only letting PHP fall back to it
for the indexes[^1].
- Notice that nothing really changed.
- Use `lsof` to try to figure out why.
- See that some asshole is downloading multiple ISOs from repo
directly (instead of a mirror)[^2], which means that nginx is
keeping the files open on unionfs until they finish.
- `sudo systemctl kill nginx; sudo systemctl restart nginx` to get
nginx to let go of the unionfs handles.
- Yay, things are way better!
- But still slow.
- htop says that now the CPU is being slammed by soft-IRQ, which is
at least better than disk wait.
- iotop says that nginx is now king of IO.
- Use lsof to see why.
- See that user(s) must have restarted their ISO downloads.
- Become frustrated that users are using repo instead of a mirror.
- Modify repomirror to append "?noredirect" to the URL when it sends
you to repo directly.
- Configure repo to redirect to repomirror unless it sees
"noredirect" in the query string.
- `sudo systemctl kill nginx; sudo systemctl restart nginx`
- Yay, proton.parabola.nu seems happy for the first time in a long
time. I don't remember the last time I saw the load average <5
except for immediately after boot.
- Notice that (somewhat ironically), this puts unionfs back into a
hot-path because repomirror uses it. But repomirror will only ever
use stat/lstat/readlink/readdir on it, so it's fine.
So, to recap:
- nginx: tune performance settings
- repo.parabola.nu: bypass unionfs if possible
- repo.parabola.nu: redirect to repomirror unless "noredirect" is in
the query string
- repomirror.parabola.nu: append "?noredirect" to the URL if
redirecting to repo.parabola.nu
[^1]: I'd tried to do this in the past, but couldn't figure out how.
I was dumb back then.
[^2]: Because we scrub client IPs, it's possible that it was actually
several users. They all just show up as 127.0.0.1[^3].
[^3]: https://www.xkcd.com/742/
--
Happy hacking,
~ Luke Shumaker
More information about the Dev
mailing list