mlops-skills

Plugin mlops-skills from marketplace ops-engineering-skills in selvarajmurugesan90/ops-engineering-skills. Commands read from the repo's manifest — 1 month ago.

Skills (22)

• complete-mlops-platform-deployment-on-aws-from-scratch

  Sequences a complete, end-to-end MLOps platform deployment on AWS from a bare AWS Organization to a production-ready platform serving a first retrained model — AWS landing zone, the SageMaker-vs-EKS+Kubeflow platform decision (EKS+Kubeflow is the worked path here), GPU node pools via the NVIDIA GPU

• complete-mlops-platform-deployment-on-azure-from-scratch

  Sequences a complete, end-to-end MLOps platform deployment on Azure from a bare tenant to a production-ready platform serving a first retrained model — Azure landing zone, the Azure-ML-vs-AKS+Kubeflow platform decision (Azure ML is the worked path), GPU compute clusters/quota, MLflow-compatible expe

• complete-mlops-platform-deployment-on-gcp-from-scratch

  Sequences a complete, end-to-end MLOps platform deployment on GCP from a bare Google Cloud Organization to a production-ready platform serving a first retrained model — GCP landing zone, the Vertex-AI-vs-GKE+Kubeflow platform decision (Vertex AI is the worked path), GPU quota and accelerator-backed

• complete-mlops-platform-deployment-self-hosted-k8s-from-scratch

  Sequences a complete, end-to-end, self-hosted and cloud-agnostic MLOps platform on any Kubernetes cluster — no managed ML service from any vendor — from a bare cluster to production: GPU node pools via the NVIDIA GPU Operator, a Kubeflow-vs-Ray orchestration choice, self-hosted MLflow-style tracking

• data-and-model-lineage

  Guides tracking end-to-end lineage from raw data sources through feature computation, training runs, and model versions, to enable root-cause tracing, impact analysis, and compliance/audit answers. Use when the user asks to "trace where a feature/model came from", set up data/model lineage tracking,

• data-catalog-and-lineage-governance-datahub-amundsen

  Guides deploying and operating enterprise data catalog and governance tooling — DataHub or Amundsen — for discovering, tagging, classifying, and owning datasets, features, and models across an organization, including deprecation workflows that must check downstream consumers before deletion. Use whe

• experiment-tracking

  Guides setting up experiment tracking for ML training runs — logging params, metrics, artifacts, and code/data versions so results are comparable and reproducible. Use when the user asks to "track experiments", set up MLflow/Weights & Biases/Neptune/DVC, compare training runs, reproduce a past exper

• feature-pipeline-failure-investigation

  Guides diagnosing and mitigating an overnight/scheduled feature pipeline failure that is causing predictions to run on stale or missing features, focused on fast triage and interim mitigation before the full root-cause fix lands. Use when the user says a "feature pipeline failed overnight", "predict

• feature-store-design

  Guides designing and operating a feature store (offline + online layers) for ML, covering feature definitions, point-in-time-correct training data generation, entity keys, and low-latency online serving. Use when the user asks to "design a feature store", "avoid feature leakage", set up Feast/Tecton

• gpu-accelerator-configuration-validation

  Guides pre-flight validation of GPU scheduling configuration — device plugin health, resource requests/limits, tolerations/nodeSelector, MIG profile availability, and driver/CUDA compatibility — before a training or serving job silently falls back to CPU or fails deep into a run. Use when the user a

• gpu-accelerator-infrastructure-for-ml-training

  Guides standing up GPU accelerator infrastructure on Kubernetes for ML training and serving workloads — installing the NVIDIA GPU Operator, partitioning GPUs with Multi-Instance GPU (MIG), designing GPU node pools, and scheduling/bin-packing training and serving pods onto that capacity without fragm

• kubeflow-ml-pipeline-orchestration

  Guides authoring, compiling, and operating ML pipelines specifically with Kubeflow Pipelines (KFP) — the Kubernetes-native pipeline SDK, its container-per-component execution model, caching, Katib hyperparameter tuning integration, and multi-tenant Profiles/namespaces. Use when the user asks to "wri

• llmops-fine-tuning-and-deployment

  Guides deciding whether fine-tuning is the right approach for an LLM use case (versus prompt engineering or RAG), and if so, executing it safely — covering LoRA/QLoRA/full fine-tuning configs, dataset curation, evaluation, and deployment. Use when the user asks whether to "fine-tune a model", "shoul

• mlflow-experiment-tracking-and-model-registry

  Sets up and operates MLflow specifically — tracking server deployment and backend/artifact store choices, Model Registry stage transitions (Staging/Production/Archived) and the newer alias-based promotion model, MLproject definitions for reproducible runs, framework autologging, and self-hosted vs.

• model-drift-alert-triage

  Guides the fast, time-boxed triage of a drift-monitoring alert that has already fired: deciding whether it is genuine data/concept drift, a monitoring code bug, a stale/broken upstream feature pipeline, or normal seasonality — before deciding whether to retrain, roll back, silence, or escalate. Use

• model-monitoring-and-drift-detection

  Guides setting up production model monitoring for data drift, concept drift, prediction drift, and model quality decay, with concrete statistical tests, thresholds, and alerting/retraining triggers. Use when the user asks to "monitor a model in production", "detect data drift", "detect concept drift

• model-packaging-and-versioning

  Guides packaging trained ML/LLM models into versioned, reproducible artifacts (model registry entries, container images, ONNX/TorchScript bundles) with a clear semantic versioning and promotion scheme (dev → staging → production). Use when the user asks to "register a model", "version a model", "pro

• model-serving-and-scaling

  Guides deploying ML/LLM models for online inference with autoscaling, latency budgets, batching, canary/shadow rollout, and cost control. Use when the user asks to "serve a model", "deploy a model for inference", set up autoscaling for a model endpoint, reduce inference latency or cost, do a canary/

• production-model-rollback-procedure

  Guides the incident-time procedure for rolling back a production model to a previous version after a bad deploy, with feature-schema compatibility verification against the rollback target as a required, non-skippable step. Use when the user says a new model deploy is "causing errors" or "bad predict

• ray-distributed-ml-orchestration

  Guides using Ray (Ray Core, Ray Train, Ray Tune, Ray Serve) via the KubeRay operator as a Python-native distributed computing framework for ML training, hyperparameter search, and serving — as an alternative orchestration paradigm to graph-based tools like Kubeflow Pipelines. Use when the user asks

• training-pipeline-orchestration

  Guides designing reproducible, orchestrated ML training pipelines (DAGs) covering data ingestion, feature computation, training, evaluation, and conditional registration/promotion steps. Use when the user asks to "orchestrate a training pipeline", set up Airflow/Kubeflow Pipelines/Argo Workflows/Ver

• weights-and-biases-experiment-tracking

  Sets up and operates Weights & Biases (W&B) specifically — the run/sweep/ artifact object model, hyperparameter sweep configuration and agents, report generation for stakeholder communication, and how W&B compares to MLflow (SaaS-first vs. self-hostable, sweep automation depth) to help a team choose

Similar plugins