{ "item_title" : "Production AI on Kubernetes", "item_author" : [" Alira Vexel "], "item_description" : "Building an AI demo is easy. Operating AI reliably in production is not.PRODUCTION AI ON KUBERNETES is a practical, production-focused guide to building, scaling, securing, and operating modern LLM inference and agentic AI platforms on Kubernetes.Rather than treating Kubernetes as a simple place to run model containers, this book shows you how to design the complete production stack-from GPU infrastructure and accelerator scheduling to distributed inference, intelligent routing, GitOps, security, observability, and cost control.You will build a working platform using technologies such as vLLM, KServe, KubeRay, Ray Serve, Kueue, NVIDIA GPU Operator, Gateway API, InferencePool, llm-d, Argo CD, OpenTelemetry, Prometheus, Grafana, MCP, and Kubernetes 1.37.Inside, you will learn how to: - Build production-ready CPU and GPU node pools with NVIDIA GPU Operator, MIG, DRA, Kueue, quotas, and topology-aware scheduling- Deploy open-source LLMs with vLLM and expose OpenAI-compatible APIs- Serve models Kubernetes-natively with KServe and implement practical autoscaling and scale-to-zero strategies- Run multi-GPU and multi-node inference with KubeRay, Ray Serve, tensor parallelism, and distributed serving patterns- Build multi-model and Multi-LoRA platforms while protecting GPU capacity from noisy neighbors- Implement intelligent inference routing with Gateway API, InferencePool, queue-aware scheduling, prefix locality, and KV-cache-aware endpoint selection- Secure production AI APIs with TLS, authentication, authorization, rate limits, tenant isolation, and workload identity- Build agentic AI applications with MCP tools, sandboxed execution, RBAC, NetworkPolicy, and controlled Kubernetes access- Manage AI infrastructure declaratively with Helm, Argo CD, and GitOps- Monitor GPU utilization, TTFT, TPOT, tokens per second, KV-cache pressure, queue depth, errors, and cost using DCGM, Prometheus, Grafana, and OpenTelemetry- Diagnose failed model Pods, exhausted GPU capacity, broken routing, GitOps drift, high latency, and security violations- Measure GPU-hours, cost per million tokens, scaling efficiency, and performance per GPUThe book concludes with a complete zero-to-production capstone that brings every layer together into one deployable platform and then deliberately breaks it to prove recovery, security, scaling, and observability.Whether you are a Kubernetes administrator, DevOps engineer, SRE, platform engineer, cloud engineer, MLOps engineer, or AI infrastructure engineer, this book will help you move beyond experimental model serving and build AI infrastructure that is secure, observable, scalable, recoverable, and operationally disciplined.If your goal is to run serious LLM inference and agentic AI workloads on Kubernetes-not just deploy them-this book gives you the production blueprint.", "item_img_path" : "https://covers1.booksamillion.com/covers/bam/9/79/816/072/9798160723396_b.jpg", "price_data" : { "retail_price" : "31.00", "online_price" : "31.00", "our_price" : "31.00", "club_price" : "31.00", "savings_pct" : "0", "savings_amt" : "0.00", "club_savings_pct" : "0", "club_savings_amt" : "0.00", "discount_pct" : "10", "store_price" : "" } }

Overview

Customers Also Bought

Details