
On 03/16/2018 12:28 PM, Nathanaël Blanchet wrote:
Thanks for precious advices!
So it means that people who thought about hosted engine feature didn't get into your philosophy of running the engine into a second datacenter
Again, strictly a "risk" thing. Hosted engine is by definition a "chicken and egg" thing. It's great for learning and for lab... but if you're going to run production, I'd at least consider the latter option I presented. With that said, we run dedicated engines today, not hosted. Remember, ovirt nodes run even while the engine is down. So you can tolerate an engine outage for a time period, just can't have reliability in case of node failures, etc. So for us, most of the risk is in rebuilding a new engine if we have to... but certainly considered a "rare" case. Putting key infrastructure inside the very thing that needs the key infrastructure to run is just fraught with problems. Everything has costs and typically, the more robust/reliable your setup, the more it's going to cost. I just wanted to present an "in between" style setup that gives you more reliability, but perhaps not the "best", while keeping costs way down. To me, if you're running any datacenter cluster (for example), you need to have a minimum of 3 nodes. People might not like that, but it's my minimum for reliability and flexibility. So... if wanted to use VMs for core infrastructure, that's 3 nodes. That core infrastructure datacenter might have a hosted engine, but likely also has "static definitions". It's part of the "core", at least several parts of it are. But the idea is it could hold: DNS, DHCP, Active Directory/LDAP, files shares (even storage domains for other datacenters), etc. Obviously a "core" failure is a "core" failure and thus needs the same treatment as whatever you consider to be "core" today. (thus on total "outage" bring up, you bring up the core, which now includes this core infrastructure datacenter... your core "tests" are run to verify, and then the rest is brought up) Then each general production datacenter cluster would have 3 nodes with the engine(s) being a VM(s) off the infrastructure datacenter using core infrastructure off that infrastructure datacenter as well. Again, this is very much like most cloud service providers today. Again, just ideas, mainly thinking on the "cheap", though some might not think so (you'll just have to trust me, what I'm presenting here is incredibly cheap for the reliability and flexibility it provides). Just my opinion.