Technology & Digital Tools Flashcards

Which Python library provides the 'Pipeline' abstraction that chains preprocessing steps and a final estimator, ensuring consistent transformations during training and inference?

Answer: scikit-learn

scikit-learn's Pipeline chains transformers and an estimator so that fit() and predict() apply the same sequence of steps, preventing data leakage and simplifying deployment.

What does 'canary deployment' mean in the context of ML model serving?

Answer: Routing a small percentage of live traffic to a new model version while the old version handles the rest

Canary deployment gradually shifts a small traffic slice (e.g., 5%) to the new model version, allowing real-world validation before a full rollout to reduce deployment risk.

In Kubernetes-based ML serving, what is the purpose of a Horizontal Pod Autoscaler (HPA) configured on a model serving deployment?

Answer: It automatically scales the number of inference pods up or down based on CPU/memory or custom metrics

HPA monitors resource utilization or custom metrics (e.g., requests per second) and adjusts the pod replica count to handle inference load fluctuations.

What is 'data versioning' in the context of ML tools like DVC (Data Version Control)?

Answer: Tracking changes to large datasets and model artifacts using Git-like version control

DVC adds Git-like version control for large data files and ML artifacts by storing metadata in Git while the actual data lives in remote storage (S3, GCS, etc.).

Which serving framework was developed by NVIDIA specifically for high-performance, multi-model GPU inference with dynamic batching support?

Answer: Triton Inference Server

NVIDIA Triton Inference Server supports multiple frameworks (TensorRT, ONNX, PyTorch, TensorFlow), GPU/CPU backends, dynamic batching, and model ensembles for production-scale inference.

In distributed training using the 'Ring AllReduce' algorithm, what is communicated between nodes?

Answer: Gradients computed on each worker's data shard

Ring AllReduce passes gradients around a logical ring of workers, where each node sends and receives partial gradient sums until all nodes converge on the globally averaged gradient.

What is the primary use case of 'SHAP (SHapley Additive exPlanations)' in ML tooling?

Answer: Explaining individual model predictions by attributing contribution scores to each feature

SHAP uses game-theoretic Shapley values to assign each feature a contribution score for a specific prediction, providing consistent and locally accurate model explanations.