[08:44:32] <_joe_>	 FYI, I'm starting to look into gvisor/kata containers (but mostly the former) to run high risk payloads in production. For now I won't focus at all on GPU support because I'm focusing on wikikube, where the risk is higher; but potentially this is interesting also for e.g. ML payloads where we allow insecure input from the internet
[09:52:23] <btullis>	 _joe_: Nice. Please CC me on stuff that I can read to follow along, if convenient. Thanks.
[09:56:48] <_joe_>	 btullis: subscribe to https://phabricator.wikimedia.org/T435044 I'd say
[12:27:54] <btullis>	 Thank you.
[13:03:39] <inflatador>	 NICE!
[13:04:26] <inflatador>	 And as far as my comments about Openstack, I'm not a hater. I say that with love from years of supporting a very odd variant (XenServer/XAPI instead of KVM/Libvirt) at a scale of tens of thousands
[13:13:29] <inflatador>	 Re: Ganeti storage. Users have to pay a large I/O penalty (software RAID-5 -> network RAID-1 -> software RAID-5) to get storage redundancy. I think there are better ways to get these storage guarantees. I know virtualization isn't sexy in a k8s world, but anytime someone wants to kick it around LMK. I've done the VMWare/Openstack/XenServer thing at scale for many years.