Kubestronaut / Current
KCNA
Use the associate-level cloud-native exam as the cheap, evidence-backed on-ramp into Kubernetes and CNCF fundamentals.
2026—2029 / learning system
Become an inference infra engineer by mid-2029.
The route is deliberate: earn the Kubestronaut through five CNCF certifications by December 2027, then build depth in MLOps and GPU serving before studying inference engine internals.
Live ledger / evidence-gated
The route, rendered as dependencies.
Every box is a milestone. A box changes state only when its evidence exists; dates keep the horizon visible, but they never unlock work.
Horizon / declared windows
Year themes
- Year 1 Credential and foundation Primary: Kubestronaut · Secondary: Systems foundations
- Year 2 Bridge to GPU serving Primary: MLOps and GPU serving · Secondary: Inference infrastructure
- Year 3 Inference infra Primary: Inference infrastructure
Roadmap graph / dependencies
Milestone lanes
State key
- Locked Dependencies remain incomplete.
- Available Dependencies are complete; evidence can unlock it.
- Current The next declared milestone for this track.
- Unlocked Evidence has unlocked this milestone.
Track
Kubestronaut
Track
Systems foundations
Track
MLOps and GPU serving
Track
Inference infrastructure
Kubestronaut / Locked
KCSA
Build the associate-level security baseline needed before operating production Kubernetes workloads.
Depends on: KCNA
Kubestronaut / Locked
CKA
Execution is delegated to the existing first-principles road engine at cka-vim-first-principles/; that engine owns the CKA lab sequence and evidence validation rather than duplicating it here.
Depends on: KCSA
Kubestronaut / Locked
CKAD
Add application-design and workload-development depth after the KCNA cloud-native foundation is proven, creating an early confidence win independent of CKA.
Depends on: KCNA
Kubestronaut / Locked
CKS
Complete the advanced Kubernetes security credential on top of the CKA administration foundation; CNCF requires an ACTIVE CKA to sit CKS.
Depends on: CKA
Systems foundations / Current
Modern C++ CPU tensor library
Build a deliberately small Tensor2D<float> library with contiguous row-major storage, shape and stride handling, elementwise addition, transpose, matrix multiplication, softmax, and RMSNorm.
Systems foundations / Locked
Optimized matmul report
Explain the performance effects of loop ordering, cache blocking, compiler optimization, SIMD, and multithreading on the tensor library's matrix multiplication.
Depends on: Modern C++ CPU tensor library
Systems foundations / Locked
Linux perf profiling study
Use Linux perf and compiler output to locate the hot loop and connect cache, vectorization, and instruction-level evidence to the optimized CPU result.
Depends on: Optimized matmul report
MLOps and GPU serving / Current
GPU Operator cluster
Operate a Kubernetes cluster with the GPU Operator, device plugin, monitoring components, and a reproducible GPU workload.
MLOps and GPU serving / Locked
Autoscaled model serving deployment
Deploy a model-serving workload on the GPU cluster and make scaling respond to measured demand rather than a fixed replica count.
Depends on: GPU Operator cluster
MLOps and GPU serving / Locked
Observability and cost-per-token dashboard
Make serving behavior legible by connecting request metrics, token counts, GPU utilization, and accelerator cost in one operational view.
Depends on: Autoscaled model serving deployment
MLOps and GPU serving / Locked
Model CI/CD
Automate model artifact validation, image or package publication, staged serving rollout, and a tested rollback path.
Depends on: Autoscaled model serving deployment
Inference infrastructure / Current
CUDA kernel ladder
Implement and measure vector add, reduction, tiled matmul, softmax, and RMSNorm kernels from first principles before relying on higher-level kernel DSLs.
Inference infrastructure / Locked
Quantization and KV-cache study
Measure the bridge-role vocabulary across FP8 and INT4 weight and activation quantization, AWQ and GPTQ, paged attention, KV-cache management, and chunked prefill.
Depends on: CUDA kernel ladder
Inference infrastructure / Locked
Non-trivial Triton kernel
Build a non-trivial fused Triton kernel, such as RMSNorm plus quantization, and show where it wins and loses across realistic inputs.
Depends on: CUDA kernel ladder
Inference infrastructure / Locked
vLLM or SGLang internals patch
Change a real inference-runtime mechanism—such as scheduling, KV-cache allocation, batching, quantized execution, or CUDA-graph behavior—instead of only wrapping an endpoint.
Depends on: Non-trivial Triton kernel
Inference infrastructure / Locked
Serving optimization report
Prove an end-to-end serving improvement from a runtime or kernel change under a stated workload and realistic concurrency.
Depends on: Quantization and KV-cache study
Inference infrastructure / Locked
Multi-GPU scaling-efficiency study
Measure how distributed inference loses efficiency through collectives, topology, synchronization, or stragglers rather than reporting only aggregate throughput.
Depends on: Serving optimization report
Inference infrastructure / Locked
Submitted upstream contribution
Contribute a production-quality inference, kernel, compiler, or portability improvement and engage maintainer review; merge is a documented bonus, not an unlock condition.
Depends on: vLLM or SGLang internals patch
A coarse plan with a hard gate.
This roadmap keeps the horizon visible without pretending a calendar can prove competence. Evidence from certificates, artifacts, and working sessions unlocks the next milestone; the ledger will show that state when its renderer is wired in.