Re: Shutting down qa.guix?

[Christopher Baines]

Maxim Cournoyer <maxim.cournoyer@gmail.com> writes:

> Hi,
>
> Christopher Baines <mail@cbaines.net> writes:
>
>> Tobias Geerinckx-Rice <me@tobias.gr> writes:
>>
>>> Christopher Baines 写道：
>>>> it's not the most cost effective setup
>>>
>>> Has this been explained in more detail before?
>>
>> Probably not, beid is currently a CPX51 Hetzner cloud server costing
>> €65.33 a month. This has been useful as it's enabled scaling the
>> resources dynamically, but it would be possible to reduce the costs and
>> still have sufficient RAM/disk space by using a Hetzner server auction
>> machine for example.
>>
>> It's not all about cost though, given the data service is one of the
>> slow points of QA, if we want QA to get faster at giving feedback, it's
>> probably important to not try and cut costs on this part of the system.
>
> Isn't QA mostly slow because of the lack of x86 build machines?  Does
> the head node needs to be powerful itself?  What kind of resources does
> it likes having the most?  CPU?  RAM?  Storage?

There are two key bottlenecks, processing the revisions in the data
service, then the build coordinator performing the builds.

For the data service, lots of RAM helps as computing and building the
derivations for Guix (similar to pull, time-machine, ...) is quite
expensive in CPU and RAM. Also computing all the derivations for each
revision takes a lot of RAM.

Storage is also an issue as beid currently is working with 340G of total
storage and that's almost full, and this doesn't leave any space for
maintenance. More storage means being able to store data about more
patch series at once.

For the build coordinator, the machine doesn't need to be powerful, it
has quite low requirements. While bayfront's storage isn't particularly
fast, it's more than sufficient in terms of hardware. More build
machines, including x86 ones would speed up the test results for patches
and branches though.

signature.asc
Description: PGP signature