Lambda system design interview

Round 5 of 7 · 4 questions reported

Length: 60 min
Format: Design discussion
Heads-up: Reported: provisioning and destroying VMs across data centers whose hypervisors run one operation at a time, with slow boots, user feedback, scaling and disaster recovery.

Questions Lambda asks in this round 4

L
Design a GPU scheduling platformSystem designLockedDesign a platform that shares a large GPU cluster among research and production teams. Teams submit training jobs (from a single GPU to thousands, running for days) and inference…
Last asked Sep 2026
L
Design a high-throughput object store for ML datasets and checkpointsSystem designLockedA GPU cloud needs storage for machine learning. Datasets hold millions of small files and large immutable shards, and thousands of GPUs read the same shards at the same time when…
Last asked Sep 2026
L
Design a Kubernetes control plane for GPU workloads with topology and health-aware drainingSystem designLockedCustomers run distributed training and inference on Kubernetes clusters of GPU machines. Design the control plane that makes Kubernetes understand GPUs well: which GPUs and GPU…
Last asked Sep 2026
L
Design a VM provisioning service across data centers and hypervisorsSystem designLockedDesign the backend that launches and destroys virtual machines for a GPU cloud. Customers ask for a VM and later delete it; leave the web frontend out.
Possibly retired