Design a Feedback-Aware Cluster Workload Scheduler
Place concurrent application workloads on cluster nodes with fenced ownership, fair admission, and versioned node feedback.
Problem statement
Design a feedback-aware scheduler for a container cluster. The reported exercise has four applications requesting scheduling at once and asks how a scheduler plug-in can use node feedback.
For this exercise, assume each application submits retryable tasks with CPU, memory, optional accelerator, topology, priority, and deadline hints. The scheduler admits bounded work, filters eligible nodes from authoritative capacity and policy, scores feasible placements, and commits one fenced binding per task.
Assume a versioned plug-in may contribute recent node-performance feedback, but stale, unavailable, invalid, or slow feedback cannot violate hard constraints or stop core scheduling. Also assume application fairness bounds starvation while higher priority receives bounded preference.
Design submission, queueing, filtering, scoring, reservation and binding, node feedback, retries, preemption, autoscaling signals, failover, and observability. Building the container runtime, application code, or cloud autoscaler is outside scope.
What to build
- Validate and admit bounded application workloads with resource, topology, priority, and deadline constraints.
- Filter and score nodes, reserve capacity, and commit one fenced task binding.
- Consume versioned bounded-freshness node feedback as an optimization input without weakening hard constraints.
- Recover scheduler or node failure while preserving per-application fairness and task retry identity.
At scale
- For this exercise, assume 10,000 nodes, 100,000 running tasks, 5,000 pending tasks, and bursts of 2,000 new tasks per second.
- For this exercise, target five seconds p95 from admission to first placement attempt for an eligible task.
- For this exercise, node feedback may be 30 seconds old; older or unavailable feedback is excluded while capacity and policy remain authoritative.
- For this exercise, scheduler failover must not create two committed bindings for one task, and an eligible backlogged application receives placement consideration within 30 seconds.