Gradual installation of debian packages

[Elazar Leibovich]

All real servers, with custom hardware attached, geographically distributed
across the planet.

Real people actually use the hardware attached to this computers, and it's
not obvious to test whether or not it failed.

The strategy therefor is, deploy randomly to small percentage of the
machines, wait to see if you get complains from those customers using these
hardware devices, and if everything went well, update the rest of the
servers.

The provisioning solution is chef, but I'm open to changing it. As I said,
I don't think it makes too much difference.

As of immutable server images, I'd do it with ZFS/brtfs snapshots
(+docker/machinectl/systemd-nspawn if you must have some sort of virtual
environment), but it's probably a better idea than apt-get install
pkg=oldversion. Immutable filesystem for execution is of course not enough,
since you might have migrations for the mutable part, etc. In this
particular case, I don't think it's a big deal.

You see, not everything is a web startup with customer facing website ;-)

Thanks,
Appreciate you sharing your experience.
I'm not disagreeing with your points, but in this particular case, where
testing is expensive, not all of them seems valid.

On Fri, Aug 5, 2016 at 3:15 PM, Amos Shapira <amos.shapira at gmail.com> wrote:

> What provisioning tools do you use to manage these servers? Please tell me
> you aren't doing all of this manually.
> Also what's your environment? All hardware servers? Any virtualisation
> involved? Cloud servers?
>
> Reading your question it feels like you are setting yourself up to fail
> instead of minimising the failure altogether.
>
> What I suggest is that you test your package automatically in a test
> environment (to me, Vagrant + Rspec/ServerSpec would be first candidates to
> check) then rollout the package to the repository for the servers to pick
> it up.
>
> As for "roll-back" - with comprehensive automatic testing this concept is
> becoming obsolete, there is no such thing as "roll-back" only
> "roll-forward", i.e. since the testing and rolling out are small and
> "cheap", it should be feasible to fix whatever problem was found instead of
> having to revert the change altogether.
>
> If you are in a properly supported virtual environment then I'd even go
> for immutable server images (e.g. Packer building AMI's, or Docker
> containers), then it's a matter of just firing up an instance of the new
> image both when testing and in production.
>
> --Amos
>
> On 3 August 2016 at 16:55, Elazar Leibovich <elazarl at gmail.com> wrote:
>
>> How exactly you connect to the server is not in the scope of the
>> discussion, and I agree that ansible is a sensible solution.
>>
>> But what you're proposing is to manually update the package on a small
>> percent of the machines.
>>
>> Manual solution is fine, but I would like to hear experience of people
>> who actually did that on many servers.
>>
>> There are many other issues, for example, how to you roll back?
>>
>> apt-get remove exposes you to the risk that the uninstallation script
>> would be buggy. There are other solutions, e.g., btrfs snapshots on root
>> partitions, but I'm curious to hear someone experienced with it to expose
>> issues I didn't even thought of.
>>
>> Another issue is, how do you select the servers you try it?
>>
>> You suggested a static "beta" list, and I think it's better to select the
>> candidates randomly on each update.
>>
>> Anyhow, how exactly you connect to the server is not the essence of the
>> issue.
>>
>> On Wed, Aug 3, 2016 at 9:30 AM, Evgeniy Ginzburg <nad.oby at gmail.com>
>> wrote:
>>
>>> Hello.
>>> I'm assuming that you have paswordless ssh to the servers in question as
>>> root.
>>> Also I assume that you don't use central management/deployment software
>>> (ansible/puppet/chef)
>>> In similar cases I usully use parallel-ssh (gnu-parallel is another
>>> alternative).
>>> First stage install the package manually on one server to see that
>>> configuration is OK, daemons restart, etc...
>>> If this stage is ok second step will be creating list of servers for
>>> "complain" list and install package on them trough parallel-ssh.
>>> Instead of waiting for complains, one can define metrics to check and
>>> use some monitoring appliance for verification.
>>> I case of failure remove package from repository and remove-install
>>> again.
>>> Third will be parallel-ssh install on all the servers.
>>>
>>> P. S. In case of few tens of servers I'd prefer to work with ansible or
>>> alternative, it's worh it in most cases/
>>>
>>> Best Regards, Evgeniy.
>>>
>>>
>>> On Tue, Aug 2, 2016 at 8:50 PM, Elazar Leibovich <elazarl at gmail.com>
>>> wrote:
>>>
>>>> Hi,
>>>>
>>>> I'm having a few (say, a few tens) Debian machines, with a local
>>>> repository defined.
>>>>
>>>> In the local repository I have some home made packages I'm building and
>>>> pushing to the local repository.
>>>>
>>>> When I'm upgrading my package, I want to be sure the update wouldn't
>>>> cause a problem.
>>>>
>>>> So I wish to install them on a few percentage of the machines, wait for
>>>> complaints.
>>>>
>>>> If complaints arrive - roll back.
>>>> Otherwise keep upgrading the whole machines.
>>>>
>>>> I'll appreciate your advice and experience of similar situation,
>>>> I'll appreciate if someone who had actual real life experience with
>>>> this situation would mention it in the comments.
>>>>
>>>> Thanks,
>>>>
>>>> _______________________________________________
>>>> Linux-il mailing list
>>>> Linux-il at cs.huji.ac.il
>>>> http://mailman.cs.huji.ac.il/mailman/listinfo/linux-il
>>>>
>>>>
>>>
>>>
>>> --
>>> So long, and thanks for all the fish.
>>>
>>
>>
>> _______________________________________________
>> Linux-il mailing list
>> Linux-il at cs.huji.ac.il
>> http://mailman.cs.huji.ac.il/mailman/listinfo/linux-il
>>
>>
>
>
> --
> <http://au.linkedin.com/in/gliderflyer>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.cs.huji.ac.il/pipermail/linux-il/attachments/20160805/29326cc9/attachment.html>