---
source_url: "https://alexanderkhundoev.substack.com/p/mlops-in-practice-docker-skypilot"
title: "MLOps in Practice: Docker, SkyPilot, Kubernetes, and H100 GPUs"
mirrored_at: 2026-08-13T15:38:34.629Z
host: alexanderkhundoev.substack.com
cited_in_42a: true
mirror_canonical: "https://index.42a.ai/alexanderkhundoev.substack.com/p/mlops-in-practice-docker-skypilot"
---

> **Original source:** https://alexanderkhundoev.substack.com/p/mlops-in-practice-docker-skypilot

I’ve just completed Homework 1 for Module 3: MLOps in the Nebius Academy AI Performance Engineering course in London.

The course is a practical, industry-focused programme for experienced developers, covering LLM internals, training, scalable deployment, MLOps, and performance engineering. Module 3 focuses on MLOps topics such as Docker, Kubernetes, production deployment, scaling, monitoring, experiment management, and cloud infrastructure.

For this homework, I set up a distributed training workflow on Nebius Cloud:

\- built a Docker image for a PyTorch training workload

\- pushed it to Nebius Container Registry

\- created a Managed Kubernetes cluster with 2 H100 GPU nodes

\- configured local SkyPilot to use the Kubernetes cluster

\- launched a 2-node PyTorch DDP training job

\- collected NCCL initialization logs and training output for submission

The most valuable part was dealing with the real infrastructure details around the training code. I ran into and learned from issues with Docker image upload time, Apple Silicon vs \`linux/amd64\` image builds, registry paths, Kubernetes public endpoints, SkyPilot local setup, kubeconfig contexts, and cloud capacity/permission limits.

It felt close to a production-like environment: the code mattered, but so did the container, registry, network, GPU nodes, orchestration layer, logs, and cleanup process.

I also put together notes from the Discord discussions and my own troubleshooting here:

[https://gitlab.com/jamesbond00/nebius-academy-ddp](https://gitlab.com/jamesbond00/nebius-academy-ddp/-/blob/discord-notes/DISCORD_NOTES.md?ref_type=heads)

I’m also collecting some of my AI experiments and course-related work on GitHub:

[https://github.com/jamesbond00](https://github.com/jamesbond00)

Still a lot to learn, but this was a great hands-on step into running AI workloads on real cloud infrastructure.

Course details:

[https://academy.nebius.com/ai-engineering-uk](https://academy.nebius.com/ai-engineering-uk)

No posts