Part 1 · Q3
Why taint GPU nodes, and how do GPU workloads land on them?
What they test
scheduling primitives applied to expensive nodes.
Strong answer
A NoSchedule taint such as nvidia.com/gpu keeps ordinary pods off nodes that cost far more per hour. GPU workloads add a matching toleration, plus a nodeSelector or node affinity on labels like nvidia.com/gpu.product. A toleration only allows a pod onto the node; the selector is what attracts it. Some clusters enable the ExtendedResourceToleration admission plugin, which adds the toleration automatically to any pod that requests a GPU.
Follow-up to expect
“A GPU pod is Pending with ‘untolerated taint’. Walk me through it.” (Walkthrough 1.)
Part 5 · Q23
Streaming responses get cut after exactly 60 seconds. Where do you look?
What they test
tracing a request through every hop.
Strong answer
A round number means a timeout. Every hop has one: the client, the cloud load balancer's idle timeout, the ingress controller (NGINX's proxy read timeout defaults to 60 seconds), any service mesh, and the gateway itself. I find the hop whose timeout matches and align them all for streaming routes. I also check that response buffering is off for server-sent events; otherwise tokens arrive all at once at the end instead of streaming.
Follow-up to expect
“Why not set every timeout to an hour?” Stuck connections then hold resources. Use long timeouts only on streaming routes, plus heartbeats.