Open Stack

On 9/23/2017 8:58 AM, LIU Yulong wrote:
> Hi nova developers,
> 
> This mail is proposed to reconsider the key pair resetting of instance. 
> The nova queens PTG discuss is here: 
> https://etherpad.openstack.org/p/nova-ptg-queens 
> <https://etherpad.openstack.org/p/nova-ptg-queens> L498. And there are 
> now two proposals.
> 
> 1. SPEC 1: https://review.openstack.org/#/c/375221/ 
> <https://review.openstack.org/#/c/375221/> started by me (liuyulong) 
> since sep 2016.
> 
>     This spec will allow setting the new key_name for the instance 
> during rebuild API. That’s a very simple and well-understood approach:
> 
>   * Make consistent with rebuild API properties, such as name, imageRef,
>     metadata, adminPass etc.
>   * Rebuild API is something like `recreating`, this is the right way to
>     do key pair updating. For keypair-login-only VM, this is the key point.
>   * This does not involve to other APIs like reboot/unshelve etc.

This was one of the issues I brought up in IRC, is that if we just 
implemented this for the rebuild API, then someone could also ask that 
we do it for things like reboot, cold migrate/resize, unshelve, etc. 
Anything that involves re-creating the guest.

>   * Easy to use, only one API.

Until someone says we should also do it for the other APIs, as noted above.

> 
> By the way, here is the patch (https://review.openstack.org/#/c/379128/ 
> <https://review.openstack.org/#/c/379128/>) which has implemented this 
> spec. And it stays there more than one year too.

It's been open because the spec was never approved. Just a procedural issue.

> 
> 2. SPEC 2 : https://review.openstack.org/#/c/506552/ 
> <https://review.openstack.org/#/c/506552/> propose by Kevin_zheng.
> 
> This spec supposed to add a new updating API for one instance’s key 
> pair. This one has one foreseeable advantage for this is to do instance 
> running key injection.
> 
> But it may cause some issues:
> 
>   * This approach needs to update the instance key pair first (one step,
>     API call). And then do a reboot/rebuild or any actions causing the
>     vm restart (second step, another API call). Firstly, this is waste,
>     it use two API calls. Secondly, if updating key pair was done, and
>     the reboot was not. That may result an inconsistent between instance
>     DB key pair and guest VM inside key. Cloud user may confused to
>     choose which key should be used to login.

1. I don't think multiple API calls is a problem. Any GUI or 
orchestration tool can stitch these APIs together for what appears to be 
a single operation for the end user. Furthermore, with multiple options 
about what to do after the instance.key_name is updated, something like 
a GUI could present the user with the option to picking if they want to 
reboot or rebuild after the key is updated.

2. An orchestrator or GUI would make sure that both APIs are called. For 
a user that is updating the key_name, they should realize they need to 
make another API call to enable it. This would all be in the API 
reference documentation, CLI help, etc, that anyone doing this should 
read and understand.

>   * For the second step (reboot), there is a strong constraint is that
>     cloud-init config needs to be set to running-per-booting. But if a
>     cloud platform all images are set cloud-init to per-deployment. In
>     order to achieve this new API goal, the entire cloud platform images
>     need updating. This will cause a huge upgrading work for entire
>     cloud platform images. They need to change all the images cloud-init
>     config from running-per-deployment to running-every-boot. But that
>     still can not solve the inconsistent between DB keypair and
>     guest-key. For instance, if the running VM is based on a
>     running-once-cloud-init image: 1. create image from this VM; 2.
>     change the keypair of the new VM; 3. reboot still can not work
>     because of the old config of per-deployment-running.

This is per-cloud configuration, and as such it should be documented 
with the cloud documentation. I can't say what is more "normal" for 
OpenStack clouds to be doing with cloud-init here, that would be a good 
question for the operators community (I've cross posted to the ops ML).

>   * For another second step (rebuild), if they have to rebuild, or
>     rebuild is the only one way to deployment new key, we are going back
>     to rebuild SPEC 1. Two steps for keypair updating are not good. Why
>     don’t directly using SPEC 1?

Because of the points made above, which are if I can simply reboot my 
instance to use the new keypair rather than rebuild it, that is much 
better. Plus then it doesn't just limit us to rebuild, but the new key 
could also be used after unshelving or cold migrating the instance.

>   * Another perspective to think about this is that SPEC 2 is expanding
>     the functionality of reboot. What about one day user want to change
>     password/name/personality at a reboot?

Kevin's spec does not propose any change to the reboot (or rebuild) APIs.

>   * Cloud user may still have questions: why key pair can not be updated
>     during rebuilding ? Why name/image/personality can?
>   * If new API does not supporting running injection, the DB keypair and
>     guest keypair are inconsistent if cloud user forget a rebuiding,
>     reboot, unshelving API call.

This is the same as the point made above, which I replied to, but anyone 
calling the native APIs should be reading the docs and understanding 
what they do. If they don't want that level of detail, use a GUI or 
orchestration tool.

> 
> 
> In conclusion, IMHO SPEC 1 is the reasonable way to set key pair for 
> rebuild API. It’s simple and easy.
> 
> SPEC 2 can be used for future running key injection. And it is still a 
> way for reboot  API to deploy the new key. But the its disadvantages are 
> as stated above.
> 
> 
> There already has some discuss [1] about this with Matt and Kevin.
> 
> Sincerely hope to receive your opinion. Feel free to ping me at IRC 
> openstack-nova, my nick is liuyulong. Thank you very much.
> 
> 
> 
> [1] 
> http://eavesdrop.openstack.org/irclogs/%23openstack-nova/%23openstack-nova.2017-09-22.log.html#t2017-09-22T14:05:07 
> <http://eavesdrop.openstack.org/irclogs/%23openstack-nova/%23openstack-nova.2017-09-22.log.html#t2017-09-22T14:05:07>
> 
> 
> 
> 
> __________________________________________________________________________
> OpenStack Development Mailing List (not for usage questions)
> Unsubscribe: OpenStack-dev-request at lists.openstack.org?subject:unsubscribe
> http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
> 

Thanks for at starting this discussion in the mailing list and 
summarizing the concerns with both approaches.

-- 

Thanks,

Matt