4 ms·

CRIU looks amazing, live-migration of LXC containers :)

jacques_chester 9y ago
There are rough edges to be worked out before we arrive at solidly productionised systems, but I believe Red Hat and Docker are both heading in that direction. Mind you, ask yourself: why do you need it? If it's for apps, isn't the point of these platforms to not need to care about individual instances? If it's for data, don't you have an existing way to manage HA or parallelism?
- cpitman 9y ago
  Yep, every time I hear people get excited about live migration of containers I have to give them a sideways look. If you are running only one instance, than availability must not have mattered that much to you.
- snuxoll 9y ago
  You’re making assumptions that I’m thinking of Docker, which couldn’t be further from the case. I’m looking at using LXC/LXD to replace full-fat VM’s for highly stateful, single-instance services - game servers. Being able to live-migrate a LXC container to drain a host for maintenance is indeed a useful feature to have.
  - jacques_chester 9y ago
    I usually assume all sorts of things[0]; I've become accustomed to living in a world where strict division between state and logic is assumed, even for relatively intensive stuff like trading platforms and very large ticketing systems. Kubernetes has been more open to absorbing sticky, less-factored workloads than more opinionated platforms are. Red Hat have definitely made that point to folk whenever we and they are competing for business. I had occasion to look at CRIU relatively closely in recent times. It is, as I said, early days, with annoying corner cases that need someone to patiently fix and polish (we ran into problems with networking, environment variables and process IDs). In your case it will probably become attractive in the next year or two, because both Red Hat and Docker Inc (which is why I mentioned Docker) are touting checkpoint migration as a feature and I expect both will devote engineering effort to productionise it. As an aside, if you're using the JVM, there was a really clever paper on coordinating the JVM G1GC and CRIU to greatly improve migration time: http://www.gsd.inesc-id.pt/~rbruno/publications/rbruno-middleware16.pdf http://www.gsd.inesc-id.pt/~rbruno/publications/rbruno-middl... edit: [0] Which is why as usual I am a dimwit for starting with answers instead of questions.

There are rough edges to be worked out before we arrive at solidly productionised systems, but I believe Red Hat and Docker are both heading in that direction. Mind you, ask yourself: why do you need it? If it's for apps, isn't the point of these platforms to not need to care about individual instances? If it's for data, don't you have an existing way to manage HA or parallelism?

cpitman 9y ago
Yep, every time I hear people get excited about live migration of containers I have to give them a sideways look. If you are running only one instance, than availability must not have mattered that much to you.
snuxoll 9y ago
You’re making assumptions that I’m thinking of Docker, which couldn’t be further from the case. I’m looking at using LXC/LXD to replace full-fat VM’s for highly stateful, single-instance services - game servers. Being able to live-migrate a LXC container to drain a host for maintenance is indeed a useful feature to have.
- jacques_chester 9y ago
  I usually assume all sorts of things[0]; I've become accustomed to living in a world where strict division between state and logic is assumed, even for relatively intensive stuff like trading platforms and very large ticketing systems. Kubernetes has been more open to absorbing sticky, less-factored workloads than more opinionated platforms are. Red Hat have definitely made that point to folk whenever we and they are competing for business. I had occasion to look at CRIU relatively closely in recent times. It is, as I said, early days, with annoying corner cases that need someone to patiently fix and polish (we ran into problems with networking, environment variables and process IDs). In your case it will probably become attractive in the next year or two, because both Red Hat and Docker Inc (which is why I mentioned Docker) are touting checkpoint migration as a feature and I expect both will devote engineering effort to productionise it. As an aside, if you're using the JVM, there was a really clever paper on coordinating the JVM G1GC and CRIU to greatly improve migration time: http://www.gsd.inesc-id.pt/~rbruno/publications/rbruno-middleware16.pdf http://www.gsd.inesc-id.pt/~rbruno/publications/rbruno-middl... edit: [0] Which is why as usual I am a dimwit for starting with answers instead of questions.