bug#51472: substitute servers should be preferred according to their coverage rate

[Maxim Cournoyer <maxim.cournoyer@gmail.com>]

merge 48808 51472
thanks

Hello Ludovic,

Ludovic Courtès <ludo@gnu.org> writes:

> Hi,
>
> Maxim Cournoyer <maxim.cournoyer@gmail.com> skribis:
>
>> When using substitute servers discovery, I've noticed that if one of the
>> substitute servers doesn't have any substitutes available, it'll keep
>> getting tried instead of others, leading to a slide-show of substitutes
>> updates such as:
>>
>> normalized load on machine '127.0.0.1' is 0.04
>> building /gnu/store/ajd0hx104702jpz2ycdwgrnyrv8jsp6d-xorg-server-21.1.0.tar.xz.drv...
>> process 9195 acquired build slot '/var/guix/offload/127.0.0.1:6666/1'
>> normalized load on machine '127.0.0.1' is 0.04
>> building /gnu/store/49rqi3wpvdm5pv6in9pamzdvg0wscrl8-xorgproto-2021.5.drv...
>> substitute: updating substitutes from 'http://192.168.10.102:80'...   0.0%
>> substitute: updating substitutes from 'http://192.168.10.102:80'...   0.0%
>> substitute: updating substitutes from 'http://192.168.10.102:80'...   0.0%
>> substitute: updating substitutes from 'http://192.168.10.102:80'...   0.0%
>> substitute: updating substitutes from 'http://192.168.10.102:80'...   0.0%
>
> We’d need to check why this particular server is checked repeatedly.
> The fact that it displays “0.0%” doesn’t mean that the server lacks
> substitutes, but that it does not reply to ‘GET /xyz.narinfo’ requests,
> for example because it’s off-line (see
> <https://issues.guix.gnu.org/48808>.)
>
>> We should implement some scheme to prefer querying high-substitute
>> servers first, instead of wasting time querying servers always failed
>> queries; this would greatly improve performance when using substitute
>> discovery for example combined with low coverage.
>
> There are several problems with that.  First one is that you can’t tell
> what substitute coverage is until you’ve actually made those GET
> requests.  Second one is that substitute coverage varies and it’s not an
> absolute measure; for example, if a server provides substitutes for only
> 0.1% of all the packages, but that’s precisely the 0.1% you care about,
> it’s more valuable than the one that has 99% of the packages but lacks
> those you want.
>
> There are other issues such as the fact that current semantics is to
> respect the order of substitute URLs, which is presumably chosen by the
> user according to their own criteria: download speed, bandwidth usage,
> etc.
>
> I hope this makes sense!

It does!  I agree that it'd be tricky to get this right; makes me
realize that my problem is probably due to #48808, and fixing that one
would probably have avoided that bug report :-).

I'm merging this one with 48808.

Thank you!

Maxim