Open Stack

On Wed, May 4, 2016, at 08:55 AM, Flavio Percoco wrote:
> On 04/05/16 15:05 +0000, Amrith Kumar wrote:
> >I'm emailing the ML on the subject of a review ongoing in the Trove project regarding image building[1].
> >
> >TL;DR
> >
> >One of the most frequent questions that new users of Trove ask is how and where to get guest images with which to experiment with Trove, and how to build these images for themselves. While documentation about this exists in multiple places (including [2], [3]) this is still something that can do with some improvement.
> >
> >Trove currently uses diskimage-builder for building images used in testing the product and these can serve as a good basis for anyone wishing to build an image for their own use of Trove. The review [1] makes the argument for the libguestfs based approach to building images and advocates that Trove should use this instead of diskimage-builder.
> 
> At the summit we discussed the possibility of providing an implementation
> that
> would allow for both DIB and libguestfs to be used but to give priority
> to DIB.
> Since there's no real intention of just switching tools at this point, I
> believe
> it'd be good to amend the spec so that it doesn't mention libguestfs
> should be
> used instead of DiB.
> 
> The goal at this stage is to provide both and help these move forward.

I disagree, as someone that has tried and failed in the past to build
trove images the goal should be to make that process easy for the user.
Maybe we can finally get rid of redstack years after it was supposed to
go away.

> >I believe that a broader discussion of this is required and I appreciate Greg Haynes' proposal at the design summit to have this discussion on the ML. I took the action item to bring this discussion to the ML.
> >
> >Details follow ...
> >
> >Before going further, I will state my views on these matters.
> >
> >1. It is important for the Trove project to do things quickly to make it easier for end users who wish to use Trove and who wish to build their own images. I am not concerned what tool or tools a person will use to build these images.
> >
> >2. If we provide multiple alternatives to image building as part of the Trove project, we should make sure that images built with all sets of tools are equivalent and usable interchangeably. Failing to do this will make it harder for users to use Trove because we will be providing them with a false choice (i.e. the alternatives aren't really alternatives). This is harder than it sounds given the combination of tools, operating systems, and the source(s) from which you can get database software.
> 
> Maintaining both in the long run will be harder especially because, as
> you
> mentioned, the output must be usable interchangeably. However, I think
> we're at
> a point, based on the comments in [1] made by Pino Toscano, Luigi Toscano
> and
> some other folks that it'd be beneficial for us to have this discussion
> and to
> also experiment/test other options.
> 
> The Sahara team seems to be going in a direction that differs with the
> one used
> by the infra team and the one we're headed to (although they overlap in
> some
> areas).

Has Sahara expressed their needs somewhere too? This is the first I am
hearing of them having trouble with image builds (I had thought they
were using DIB successfully).

> >3. Trove already has elements for all supported databases using DIB in the trove-integration project but these elements are not packaged for customer use. Making them usable by customers is a relatively small effort including providing a wrapper script (derived from redstack[4]) and providing an element to install the guest agent software from a fixed location in addition to the development and testing version that is better suited to Trove development [5] and [6].
> >
> >4. My comments on various patch sets in the review[1].
> >
> >I agree with Monty and Greg Haynes that we should understand the deficiencies if any in DIB, and if it is in fact the case that they are "intractable/unsolvable", we should switch toolchains. This discussion should include issues faced by the Trove team as well as other teams that may have faced problems with DIB (such as the sahara team who described some of them in the past).
> 
> ++
> 
> Agreed with the above. I'm think collaboration should be the preferred
> way. I
> don't think I've enough technical insight on this topic to provide a
> detailed
> list of things that are good/bad on either of these tools but I wanted to
> mention that I believe providing support for both in the short run is
> good for
> us and it helps to make a better decision on what tool works best for the
> project.
> 
> There's someone willing to do the job and spend sometime doing the
> research.
> This same person will provide feedback in addition to the one already
> provided
> in [1].
> 
> Sorry for not providing much technical details now but I did want to
> share the
> above. Thanks for starting this thread, I believe this discussion in the
> ML will
> be beneficial.

This is the problem. We need technical details and we should fix the
problems. It is easy to say X is better than Y because. Unfortunately it
doing so doesn't address the actual problems in using either X or Y.

The infra team has been using DIB successfully for a while now and it is
a very flexible tool. On the flip side Nova disables file injection by
default [1] because it created problems and there are better ways of
accomplishing this use case (the inject_partition default of -2 remains
today). On top of that devstack seems to have decided that nova's file
injection should not be configurable and hard sets it off [2].

What are the problems with DIB and how do they prevent trove users from
building an image today?

On the positive side using DIB it is very flexible. As has been pointed
out in the spec review you can bootstrap using tools like yum and
debootstrap if you want complete control of your image build from the
ground up, but you can also start from previously published images if
you don't need that level of control. One recent benefit we have run
into is DIB allows us to pick the filesystems we want which starting
from existing images with prebaked filesystems does not make easy (/me
shakes fist at silly defaults chosen by particular distros).

As for security as I understand it using libvirt/qemu on a machine with
virt capabilities is not very different from having root? But with DIB
you can run DIB within a virtual machine without access to the
hypervisor and without nested virt and achieve a high level of
isolation.

My suggestion would be to fix problems in DIB (at the very least please
file bugs and let the appropriate parties know what problems you are
having), and see if Trove user needs can be addressed quickly without
reinventing wheels. And in the mean time see if the devstack and nova
teams have any input based on their experiences which led to turning
these sorts of features off.

[1] https://review.openstack.org/#/c/70239/
[2] https://review.openstack.org/#/c/194586/

Clark