Open Stack

On 11/08/2016 12:13 PM, Matt Riedemann wrote:

> I'd also like to say that I dislike the constant comparisons to the kernel. If
> people are going to make that comparison, then let's say the kernel overall is
> all of OpenStack, and there are subsystems, like nova/cinder/glance/etc, with
> their own subsystem maintainers, e.g. nova-core.

While there is some merit to that argument, it's not entirely accurate.  Nova is 
roughly 1/45th the size of the whole kernel source, but the kernel has roughly 
1600 separately-maintained subsystems.  The "arch/x86" subdirectory in the 
kernel source is almost half as big as nova, but there are also 
separately-maintained drivers with only a few hundred lines of code.

Something that was mentioned in the summit discussion (but I think it bears 
repeating) is that the thing that enables this sort of fine-grained 
maintainership in the kernel is that the contracts are better-defined between 
the various subsystems.  Generally a driver has a limited set of things that it 
needs to worry about getting right to interface with the rest of the system, and 
other than that it's got free rein in its own sandbox.

That said, I don't know that there's an easy solution to this in nova due to the 
fact that it's a distributed system with a central shared data store.  Enhancing 
a sched filter might require new data, which means modifying the DB model and 
the DB migrations, and updating the InstanceNUMATopology object and tweaking 
nova-api to request additional attrs, and it might have implications for 
upgrade, etc.

Chris