Open Stack

Am 2016-07-25 19:05, schrieb Jeremy Stanley:
> On 2016-07-25 11:08:35 +0530 (+0530), Vipul Nayyar wrote:
>> Honestly, I was also thinking that using containers for implementing
>> blue/green deployment would be best for implementing minimal downtime. 
>> I
>> suggest having a basic run-through of this idea with the community 
>> over
>> tomorrow's irc meeting should be a good start.
> 
> Waving containers at the problem doesn't really solve the
> fundamental issue at hand (we could just as easily use DNS or an
> Apache redirect to switch between virtual machines, possibly more
> easily since we already have existing mechanisms for deploying and
> replacing virtual machines). The issue that needs addressing first,
> I think, is how to get new DevStack deployments from master branch
> tip of all projects to work consistently at each rebuild interval
> or, more likely, to design a pattern that avoids replacing a working
> deployment with a broken one along with some means to find out that
> redeployment is failing so that it can effectively be troubleshot
> post-mortem.

Hi Jeremy,

broken DevStack installation - that's the point. With LXD container you 
can take snapshot, run unstack or clean script, fetch new code and stack 
again. If it failed you can restore the snapshot and try new 
installation on another day. Without snapshot you can start new 
container with new code and shutdown the old one. So I like the idea 
with haproxy in front but wouldn't change any DNS entries because it 
takes time for end-users.
If you have enough resources then we can work with 3 VMs: 2 DevStack 
installation with translation check-site and one with haproxy hosting 
the public FQDN and a kind of trigger to refresh the installation on the 
DevStack VM _if_ the other VM is up. If the other DevStack service is 
down, the trigger should try an unstack/clean/stack after one day and 
switch over if the service is up. This could be done with lb-update 
(https://www.haproxy.com/doc/aloha/7.0/haproxy/lbupdate.html) or haproxy 
API.
The process should have a small monitoring about the status.

kind regards

Frank