Move 200 TB of Media to GPU Training Hosts
You have 200 TB of media data — images and video — and you need to get it onto the compute hosts that will preprocess, run inference, and train on it. Design that, end to end: from the source, to storage, to GPUs that never sit idle waiting for data.
Cover: the questions you would ask first; the transfer arithmetic and the network-versus-shipping decision; where the data lands and why; how the transfer survives failure; how the training and inference hosts consume the data without starving the GPUs; and what changes if the transfer becomes recurring.
Hints
Hint 1
200 TB is 1.6 × 10^15 bits: about 18.5 days at 1 Gbps, 1.85 days at 10 Gbps, 4.4 hours at 100 Gbps — and roughly 40% worse at a realistic 70% of line rate. When the transfer takes longer than shipping disks, ship the disks.
Hint 2
The data does not go onto a processing host's local disk. It lands in object storage, which decouples transfer from consumption, survives a resumed upload, and can be read by the whole cluster.
Hint 3
Millions of small files are a metadata and IOPS problem, not a bandwidth one — preprocess offline into large sharded archives so training reads are sequential, and stream shards with prefetch instead of staging 200 TB locally.
Asked at
Sign in to talk to the interviewer — ask what's in scope, what scale to design for, or for a nudge when you're stuck. Conversations save to your account.