[arch-general] Question about automated builder

[Thomas Dziedzic]

On Thu, Jan 27, 2011 at 2:06 PM, Thomas S Hatch <thatch45 at gmail.com> wrote:
> On Thu, Jan 27, 2011 at 12:57 PM, C Anthony Risinger <anthony at extof.me>wrote:
>
>> On Thu, Jan 27, 2011 at 1:48 PM, Thomas S Hatch <thatch45 at gmail.com>
>> wrote:
>> >
>> > Awesome, I actually have a few servers I will use, and since it will be
>> > distributed, we will be able to use a lot of servers as builders all
>> > reporting to a single master.
>>
>> why all the distributed goodness, then a crippling single master?  it
>> would not be much more difficult to use 1-to-1 queues-to-server.
>> nodes that fill up return a redirect, or denial.  a simple DNS scheme
>> can facilitate auto-discovery... nodes register for an ID under the
>> domain, any simply move down the list.  TXT records can provide the
>> lists of root nodes if you really want.  with more work, this can be
>> made to work through NAT, a la p2-esque, but that's more ambitious.
>>
>> > As for the web end, I was thinking of having the web frontend just act as
>> a
>> > notification area about the queue and the builds, so people could check
>> on
>> > build stats and download experimental pkgs etc. Then the queue would be
>> > managed via the scms.
>>
>> no way man, go big or go home!  web interface is full out AUR replacement.
>>
>> C Anthony
>>
>
> Hmm, a pure peer setup? My thought on the master has more to do with central
> job control, and the fact that distributing can be easily done, the master
> is a lightwieght server compared tot he builders, and the master will just
> present what needs to be build and pull whats built into repos, this way the
> master can scale out to a ton of builders and all of the decisions about who
> builds is done by the builders talking to each other.
>
> Of course.... we could have separate master dedicated to specific repos and
> configurations but share the builder pool, so that we have a simple many to
> many relationship.....
>
I prefer the central server idea rather then a purely distributed
system simply because we can't distribute workload well with a purely
distributed one. Imagine serving openoffice and libreoffice to one
server and some python module to another server. At least with the
central server, we have more control over this.