Design a production-grade ML inference pipeline capable of serving 50,000 requests per second with a P99 latency under 20 milliseconds. The pipeline involves data preprocessing (feature extraction), model inference (using a deep learning model), and post-processing. Your design should specify: 1) The infrastructure components (e.g., Kubernetes, load balancers, message queues) and their roles. 2) The model serving technology (e.g., TensorFlow Serving, TorchServe, Triton Inference Server) and justification for the choice. 3) Optimization techniques such as model quantization, batching strategies, or caching to meet latency requirements. 4) A strategy for Canary deployments and A/B testing new model versions without impacting the live traffic.