[openstack-dev] [kolla] image bloat
Steven Dake (stdake)
stdake at cisco.com
Fri Jul 17 03:49:51 UTC 2015
From: Robert Collins <robertc at robertcollins.net<mailto:robertc at robertcollins.net>>
Reply-To: "OpenStack Development Mailing List (not for usage questions)" <openstack-dev at lists.openstack.org<mailto:openstack-dev at lists.openstack.org>>
Date: Thursday, July 16, 2015 at 6:48 PM
To: OpenStack Development Mailing List <openstack-dev at lists.openstack.org<mailto:openstack-dev at lists.openstack.org>>
Subject: Re: [openstack-dev] [kolla] image bloat
On 17 Jul 2015 12:30, "Steven Dake (stdake)" <stdake at cisco.com<mailto:stdake at cisco.com>> wrote:
> Hey folks,
> The bloat on the from-source comes from the C toolchain. The bloat on the binary images comes from our common dependency layer. We can probably just get rid of all the Python dependency installs that occur in the common layer and make things a whole lot faster when deploying at large node scales although it may negatively impact build speed.
I very much hope plain pip would be faster not slower:)
We don’t use pip during deployment as it results in an uncontrolled deployment mechanism. Instead we pre-build containers and rely on an atomic deployment (either the container deploys or doesn’t, vs using pip during deployment, where some could fail and some could succeed). Pre-building is more deterministic, but images must be build.
Using pip is slower then yum when installing packages. Most of this has to do with invocation of the c compiler for various modules (lxml is an example that takes awhile).
What I actually proposed in the above email was getting rid of our common dependency layer where yum is only used one time during the build. This would be replaced by using yum for each container (* 40 or so containers *100 or so dependencies). We already do this for from-source, but not from-binary container building.
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the OpenStack-dev