Re: [atomic-devel] We have a bugzilla requesting that we change the default CMD to systemd for base images in RHEL

[Josh Berkus <jberkus redhat com>]

On 10/25/2016 10:54 AM, Daniel J Walsh wrote:
> 
> 
> On 10/25/2016 01:43 PM, Josh Berkus wrote:
>> On 10/21/2016 01:17 PM, Daniel J Walsh wrote:
>>>
>>> On 10/21/2016 01:14 PM, Matthew Miller wrote:
>>>> On Fri, Oct 21, 2016 at 01:03:58PM -0400, Joe Brockmeier wrote:
>>>>> I like Dan's proposal of rhel7-init (or fedora-init, centos-init).
>>>> For whatever it's worth, I don't. It makes sense if you're steeped in
>>>> the distro world where init systems have been a hot topic for the last
>>>> few years, but without that context it sounds like it's something about
>>>> inititalizing RHEL/Fedora/CentOS, not that there's a process manager
>>>> running. Or even an initial attempt at making a rhel7 container. 
>>>>
>>> Do you prefer
>>> rhel7-systemd, fedora-systemd, centos-systemd
>>> or
>>> rhel7-system, fedora-system, centos-system
>>>
>>> Or don't like the idea at all?
>> Given that we already get flak ... which leads to reduced adoption ...
>> of our base images due to their large size, this seems like the wrong
>> direction to move into.
>>
>> Systemd and its dependencies are quite large, sometimes larger than
>> everything else in the container combined, both in storage space and in
>> memory footprint.  If we want to push the idea of having systemd inside
>> containers, then we need a new tiny version of systemd designed for
>> containers.  And by "tiny" I mean "less than 10MB, with few or no
>> dependancies which are liable to create extra update churn".
>>
>> A tiny systemd would be a great selling point.  We'd have integrated
>> journaling, more sophisticated PID1 scripts, zombie harvesting, etc.
>> Those things are benefits to the developer, and we could promote them.
>> But nobody's going to buy those at the cost of 100MB+ of bloat.
>>
> Well splitting it into two container images rhel7 and rhel7-systemd,
> would allow you to
> get closer to your deal.  We attempted to do fakesytemd and
> systemd-container, but both
> ended causing more trouble then they were worth.  We currently ship real
> systemd inside
> of the RHEL7/Fedora/Centos container images, I  believe it has been
> shrunk as far as its dependencies
> but removing it would be better, although every daemon package that you
> install will pull it
> in, since they all currently require systemctl. 
> 
> Colin's work on this should improve the situation, and I think we have
> shrunk the size of Fedora
> images with some of the newer versions of rpm/dnf.

Well, there's no reason why a systemd designed for just container init
couldn't be tiny; it's requirements would be so much less than one
running outside a container.  But it would be essentially a new piece of
software.


-- 
--
Josh Berkus
Project Atomic
Red Hat OSAS