2015-12-11 Compute Requirements for Images and Kernels¶
It has become clear that the myriad selections in both compute operating system and underlying technology such as bare metal or containers mean we should be clear in the intent to provide a unified user experience across many OpenStack clouds.
It is the opinion of the OpenStack Technical Committee that the following statements should be true in the definition of an OpenStack Powered Compute cloud.
The following use key words as found in RFC 2119.
An OpenStack Powered Compute Cloud MUST allow end user Image Uploads
An OpenStack Powered Compute Cloud MUST be able to boot a Linux Guest
An OpenStack Powered Compute Cloud MUST be able to boot arbitrary operating system kernels the user uploads, given matching processor architecture
An OpenStack Powered Compute Cloud MUST be able to provide those capabilities using VMs
An OpenStack Powered Compute Cloud MAY additionally provide users the ability to boot uploaded images on Bare Metal machines
An OpenStack Powered Compute Cloud MAY additionally provide users the ability to boot using more restrictive technologies such as containers, zones or jails
Images as provided by cloud deployers diverge, and cause pain for our end users from an interoperability standpoint. While cloud providers may choose to provide value with special images, OpenStack should spend effort in facilitating end user choice in what base images they run. Direct image upload is an example of a way to facilitate end user choice.
We take a strong stand on the ability to test the inbound interfaces a cloud provides to a running server instance. In order to test the interfaces that a user of an OpenStack cloud should be able to expect (for instance, does config-drive show up, does the instance have the IP that neutron says it has) we have to be able to have tests in the gate. Since OpenStack follows The Four Opens that means that our tests must be Open Source. Alternate versions of the tests that might test the same interfaces but use Close Source Operating Systems are not possible to be validated since we can’t run those Operating Systems in the gate, so cannot be accepted.
Linux is not an undue burden to ask someone to run. It runs on x86, power, sparc, mips, alpha, atom and arm processors and even IBM z-Series mainframes. It runs on phones, TVs, and watches. It is free and anyone can get an operational image. The utilities the tests require of a Linux guest are standard and available.
Asking someone to boot a Linux guest in order to test inbound interfaces is, therefore, not only completely reasonable, it’s the most reasonable and most obvious thing possible.
An OpenStack cloud that has LXC, Docker, Rocket, Solaris-zones or other lightweight container technology sounds pretty neat at first pass, but it turns out that an LXC cloud would not allow a Solaris image and a Solaris Zones cloud would not allow a Linux image. That is too much deployer choice bleeding through the abstraction layer, and as such is bad for interoperability. While there is nothing preventing deployers from using the OpenStack software to create clouds that are solely based on such technologies, those clouds are divergent from what OpenStack clouds are.