Preparing your design workspace

Loading the problem and your latest saved attempt…

Burst-Aware Serverless Task Scheduler

Design a fair scheduler that places 10,000 short heterogeneous tasks arriving in one second.

Problem statement

For this exercise, assume tenants submit short serverless tasks with CPU, memory, accelerator, image, locality, deadline, and priority requirements. Bursts can reach 10,000 tasks in one second.

The scheduler must admit work durably, preserve per-tenant fairness, avoid resource fragmentation, exploit cached images and data, cancel queued or running work, and recover from scheduler or node failure. Exclude kernel scheduling and task code execution.

What to build

  1. Durably admit, validate, quota-check, and classify bursty task submissions.
  2. Choose eligible capacity using fairness, locality, deadline, and fragmentation signals.
  3. Fence assignments, process cancellation, and reschedule lost work without duplicate committed results.

At scale

  • Absorb 10,000 task submissions in one second and 100,000 queued tasks per region.
  • Begin 95 percent of tasks with available capacity within 2 seconds.
  • Tasks range from 100 milliseconds to 10 minutes and may request CPU, memory, or one to eight GPUs.
  • No tenant may consume more than its burst quota or starve a continuously eligible peer.