What's preventing k8s from being useful in these larger scale deployments? GP was not very clear about that, they just said that it's not scaling enough.
K8s doesn't do anything to help you with managing the fleet of k8s hosts. At 10^5 host machines you have a much bigger problem on your hands than the one that k8s is solving for.
If you have and are solving that hyperscale infrastructure problem then whether or not you're running kubernetes doesn't actually make a damn bit of difference to how you allocate your human/financial resources.
Or summarized, at the scale where Kubernetes is advertised to make the biggest difference, it's actually irrelevant/an implementation detail.
A lot of k8s is just not optimized for beyond a relatively small scale (10k services etc.). In particular many of the components rely on a primary node to handle all the traffic rather than sharding work across a group of nodes.
Completely true. Woe unto those who learned this the hard way and have suddenly found themselves in the fulltime position of "tuning etcd for kubernetes".
What's preventing k8s from being useful in these larger scale deployments? GP was not very clear about that, they just said that it's not scaling enough.
K8s doesn't do anything to help you with managing the fleet of k8s hosts. At 10^5 host machines you have a much bigger problem on your hands than the one that k8s is solving for.
If you have and are solving that hyperscale infrastructure problem then whether or not you're running kubernetes doesn't actually make a damn bit of difference to how you allocate your human/financial resources.
Or summarized, at the scale where Kubernetes is advertised to make the biggest difference, it's actually irrelevant/an implementation detail.
10^5 machines cost $100M, at which point you are a huge enterprise who can afford to pay people to customize your system.
2 replies →
Is that a proper job for k8s? Isn't it supposed to be strictly for orchestrating workloads on top of that fleet?
2 replies →
A lot of k8s is just not optimized for beyond a relatively small scale (10k services etc.). In particular many of the components rely on a primary node to handle all the traffic rather than sharding work across a group of nodes.
Completely true. Woe unto those who learned this the hard way and have suddenly found themselves in the fulltime position of "tuning etcd for kubernetes".