engineering / UCL / 2025–2026
Distributed ML inference platform
A 5-node Kubernetes inference cluster provisioned as infrastructure-as-code.
Kubernetes (K3s) / Terraform / Ansible / Docker / MinIO
Question
How can an inference cluster be provisioned reproducibly and remain operable under resource pressure?
Why it matters
Infrastructure-as-code makes the intended cluster configuration explicit. Diagnosing the storage failure connects the design to operating behavior.
Method and findings
- Designed and deployed a 5-node inference cluster (22 cores, 140 GB RAM), fully provisioned as infrastructure-as-code.
- Added distributed object storage (MinIO) and a message-queue task layer for distributing inference jobs.
- Diagnosed live failures under resource pressure, traced them to the storage layer, and re-architected it to remove the root cause.
Limitations
The cluster specification describes the deployed system; this page does not present a throughput, latency or availability benchmark.
Next questions
How would the revised storage design behave under a documented, repeatable load test?
Sources
Project details supplied by the author. A public technical report is not linked here.