Hyperscale Prep

Public study site for operating large-scale compute at hyperscale — the kind of role that asks for 100+ clusters / 10K+ nodes, deep Kubernetes internals, cloud & cluster networking, security, and safe change systems (Terraform/Atlantis, Temporal/Argo).

Maintained as daily operator-deep lessons plus diagrams: written like an engineer who ran this in production — not a certification cram sheet.

Live docs: caretak3r.github.io/hyperscale-prep (opens in a new tab)

Practice problems for LLMs

AI Inference & Hosting Infrastructure Benchmarks (12-page PDF): 21 production-grade prompts, ordered by difficulty from L1 (a single-replica OpenAI-compatible server) to L9 (hierarchical KV cache and encode-prefill-decode on mixed accelerators), with multi-region failover, Kubernetes-native inference (KServe, llm-d, DRA, Kueue) and Slurm hybrids in between. Each prompt has an objective, staged artifacts to produce, and a mechanical grader contract, so you can run them against an LLM (or yourself) and grade the artifacts, not the prose.

Agent infrastructure track

The next stretch of lessons goes one layer up, into the agent runtime infrastructure from the Agent Infra Landscape 2026 map: memory and context stores, MCP/A2A gateways, sandboxes, browser and computer-use fleets, and agent observability. Same house rules as every other lesson (operator depth, Security & hardening, diagrams, failures & fixes), written from the view of the team that has to run these for thousands of agents on Kubernetes and GPU cells.

Upcoming (Lessons 52–54 are done):

What this JD is really asking

Interviewers saying “100+ clusters / 10K+ nodes” are not asking if you can kubectl get nodes on a big managed cluster. They want evidence you have run a fleet under partial failure: many independent control planes, continuous node death, and automation that treats clusters as cattle.

This site builds that mental model lesson by lesson.

How daily lessons work

1. One lesson per day on a JD-aligned topic.
2. Each lesson targets production depth:
   • Diagram (PNG + Mermaid)
   • Concrete commands / config
   • Failures & fixes (symptoms → causes → mitigations)
   • Security & hardening (operator controls; AI/ML weights/images/deployments when relevant)
   • Real reference URLs
   • Interview soundbites
   • Self-check prompt
   • Watch — YouTube talks/demos when applicable
3. Also covers AI/ML platform engineering (GPU fleets / inference serving) using wafer-ai/gpu-perf-engineering-resources (opens in a new tab) as curriculum fuel — operator depth, not CUDA tutorials — including model/weight hardening, private registries, signing/provenance, and secured inference deployments.
4. Future lessons live as MDX under lessons/ — see Conventions.
5. The original hyperscale-one-pager.md (opens in a new tab) is kept as an archive; the docs site is the canonical reading path.

Curriculum

Progress metadata: curriculum.json (opens in a new tab).

How to read

1. Interview crunch: newest lessons first; use soundbites + Failures & fixes as flashcards.
2. Deep prep: Lesson 1 → 2 → … in order — later lessons assume cell/fleet vocabulary.
3. Open linked PNGs / Mermaid while reading; each lesson says how to read the diagram top-down.
4. Do the Quick self-check out loud before the next lesson.

Start here

→ Lesson 1 — Hyperscale mental model