Proposal for new Linux image registration policies
Executive Summary
Currently, for each Linux image, there are two series of images, one
with a datestamp (e,g, _20180106_0) and one with a _latest name
which gets renamed to _prevN and hidden upon a new _latest image
registration. We have followed a roughly monthly update schedule
(aligning with the Windows world) for the datestamp images and the
_latest images while pushing out additional _latest images in
between occasionally based on major improvements, bug fixes or security
issues.
The rationale was to both offer images that can be used long-term (for customers that may not yet have adopted continuous deployment models) as well as a possibility for customers to have current images (which don't need a lot of online updates to be secure).
The disadvantage of the current approach is that a lot of old images are
stored on the platform and (more importantly) cause the image list to
become larger and larger, confusing customers and causing a
glance image-list to require more than 10s of time.
We found a lot more customers using the _latest images and thus
propose to completely drop the images with a date stamp. We have created
a possibility to hide older images from the _prevN series in a way
that they can be continued to be used IF referenced by image UUID.
The proposal in detail
We stop creating new images with a date stamp. Only _latest images
will be created, previous _latest get renamed.
New cycling of previous _latest images
When we register a new _latest image, we suggest to now rename the
previous _latest image to carry a date stamp e.g. _20180129_0
instead of _prevN.
These date stamp images get hidden immediately upon rename.
The hidden images do NOT show up in the image catalog in the Web
Interface ("Service Console") any more, nor are they visible to normal
tenant users via the glance image-list invoked API call.
However, if a user has stored the image UUID, (s)he can still query
details on the image details using glance image-show and can still
create new VMs/BareMetal ("ECS") Instances from this image. Likewise, a
nova rebuild (redeploying the image to an existing VM) still works as
well. The web interface and nova also display the image name correctly
for existing VMs with (meanwhile) hidden images.
Option: If we need to keep older images visible for a while, we insert a
step with a _prev label before hiding: Sequence then could be:
_latest (visible) -> _prev (visible) -> _DATE (hidden). This
way no _DATE image can ever be referenced by name, which would lead to
breakage.
Customer recommendations
After implementation of the new policies, customer have the same choices as before: Useing latest or fixed images. The choice can be done by
- Always using our latest images by referencing it by the name
_latest. - Staying on the same image by referencing it by UUID.
We expect most cloud users to have mechanisms in place where they can
handle change -- they are typically best served by always using our
_latest images. However, we acknowledge that there are valid cases
where users can not deal with OS maintenance updates well and thus we
offer the ability to reference images (current ones as well as old ones)
by UUID. While we hide the old images after some time, they remain
available by UUID for a long time.
Migration
To achieve the goal of cleaning up our public image catalog, we need to hide or remove old images.
Here is our proposal for this:
- We only delete ancient images (2016) which are not in use any more or don't work well any longer.
- We hide the images that still carry minor versions which we stopped using.
- All other old images will be hidden, not deleted, meaning that the images can still be used if the customer has taken a note of the image UUID (or requests that information from our support organisation).
- We announce the new policy to customers and tell them that they need
to remember UUIDs if they chose to not use the
_latestimages. - We can start with the new image registration policy immediately, BUT wait with hiding all old images only after some weeks to give customers a bit of time to adjust.
Benefits
- The amount of (new) images that we register and store on the platform is cut by a factor of two.
- The images showing up in the image catalog is reduced by an order of magnitude, avoiding confusion and speeding up the catalog operations.
- The Linux images are handled more similarly to the Windows images.
- Communication is easy: Images by name
_latestare always latest and Images by UUID are stable and aging (similar to content-based addresing as used by git).
Timeline
We suggest to align this within T-Systems and with key customers and then implement this within February.