MLA-B01 Study Guide

Key Topics

Must-Know Concepts

• Feature Store architecture: feature entities (business objects), feature views (transformation logic), managed vs external feature tables, and automated incremental refresh
• Point-in-time feature lookups: how Feature Store ensures training data does not leak future information when generating historical feature sets
• Online vs offline feature serving: online for low-latency real-time inference, offline for training dataset generation and batch scoring
• Dynamic Tables for feature pipelines: declarative SQL transformations with automatic refresh, target lag configuration, and DAG chaining for multi-step feature computation
• Streams and Tasks for complex feature pipelines: when to use CDC-based streams with task scheduling instead of Dynamic Tables (MERGE operations, stored procedures, custom error handling)
• Snowpark ML preprocessing: distributed scalers (StandardScaler, MinMaxScaler), encoders (OneHotEncoder, OrdinalEncoder), and Pipeline objects for chaining transformations
• Data quality monitoring with Data Metric Functions: freshness, completeness, and accuracy checks on feature tables
• ML Lineage for feature traceability: tracking data flow from source tables through feature definitions to training datasets and models
• Feature engineering patterns: group-by aggregations, window functions, pivot operations, and join strategies using Snowpark DataFrames
• Incremental feature computation: how Dynamic Tables and Streams enable efficient processing of only new or changed data

Common Exam Traps

Key Topics

Must-Know Concepts

• SPCS compute pool architecture: instance families (CPU_X64_S/M/L for CPU, GPU_NV_S/M for A10G workloads, GPU_NV_L for A100 workloads), min/max node autoscaling, and workload-to-pool matching
• GPU support in SPCS: A10G and A100 NVIDIA GPUs for model training, fine-tuning, and inference. Know when GPU compute is necessary vs CPU-sufficient
• Model Registry as first-class schema objects: model versioning, default version designation, metadata storage, and deployment to warehouse or SPCS
• Experiment Tracking: logging hyperparameters, metrics, and artifacts during training runs for comparison and model selection
• Snowpark ML model training: built-in wrappers for scikit-learn, XGBoost, and LightGBM that run natively in Snowflake without UDF creation
• Distributed hyperparameter optimization: GridSearchCV and RandomSearchCV execution across multiple warehouse or SPCS nodes using UDTFs
• Container Runtime for Notebooks: preconfigured ML software stacks (PyTorch, TensorFlow, scikit-learn) on CPU or GPU compute pools, extensible with additional packages
• ML Jobs for production training: scheduling model retraining, integrating with external IDEs, and running on Container Runtime
• Warehouse compute vs SPCS: warehouse for SQL inference and supported framework training; SPCS for custom containers, arbitrary packages, GPU workloads, and HTTP serving endpoints
• Model artifact management: storing, versioning, and comparing trained models and their associated metadata in the Model Registry

Common Exam Traps

Key Topics

Must-Know Concepts

• Model Serving architecture: Model Registry model deployed to SPCS as a managed HTTP endpoint. Snowflake automates container image building, deployment, and endpoint setup
• Batch vs real-time inference: batch runs on warehouse or SPCS jobs for bulk processing; real-time deploys as HTTP endpoints on SPCS for individual predictions with autoscaling
• Deployment process: register model in Model Registry, configure compute pool, deploy to SPCS, verify endpoint health, route traffic
• Autoscaling for Model Serving: configure min/max nodes in SPCS compute pools. Snowflake scales based on request load automatically
• GPU selection for inference: A10G GPUs for standard inference workloads, A100 GPUs for large models requiring more memory and compute
• Model version management: default version designation in Model Registry, deploying specific versions, and rolling back to previous versions
• Inference on warehouse vs SPCS: warehouse for SQL-integrated batch predictions with supported models; SPCS for custom containers, arbitrary packages, and HTTP endpoints
• Multi-modal batch inference: SPCS job-based batch inference supporting GPU acceleration across multimodal datasets
• Deployment patterns: blue-green deployments using multiple Model Serving endpoints, canary releases by splitting traffic between model versions
• Endpoint observability: monitoring inference latency, throughput, error rates, and resource utilization for deployed models

Common Exam Traps

Key Topics

Must-Know Concepts

• ML Jobs for pipeline orchestration: scheduling retraining pipelines, connecting multiple steps, and running on Container Runtime with support for external IDE development
• Task Graphs (DAGs): chaining Snowflake Tasks using AFTER clauses to create multi-step workflows with CRON scheduling and dependency management
• CI/CD with Git integration: connecting Snowflake Notebooks and code to Git repositories (GitHub, GitLab) for version control and collaborative development
• Snowflake CLI GitHub Actions: automating deployment pipelines triggered by Git events (merge to main, release tags) to deploy updated DAGs and ML artifacts
• External orchestrator integration: using Airflow, Prefect, or Dagster with ML Jobs when organizations have existing orchestration infrastructure
• Pipeline stages: data ingestion, feature engineering, model training, model validation, model registration, model deployment, and monitoring — each as an orchestrated step
• Version control for ML artifacts: tracking notebook code, model configurations, feature definitions, and pipeline definitions in Git
• Automated model retraining: scheduling periodic retraining pipelines triggered by data drift detection or calendar schedules
• Deployment automation: from code merge to production deployment using CI/CD pipelines that validate, test, and deploy ML models
• Error handling in pipelines: task retry policies, failure notifications, and graceful degradation patterns in production ML pipelines

Common Exam Traps

Key Topics

Must-Know Concepts

• RBAC for ML artifacts: securing models, feature tables, compute pools, and inference endpoints using Snowflake's role hierarchy (access roles + functional roles)
• Dynamic data masking for ML: applying column-level masking policies to ensure sensitive training data is not exposed to unauthorized roles
• Snowflake Horizon governance suite: universal data discovery, Trust Center for security posture, data classification for sensitive data identification
• ML Observability: configuring model monitors for regression and binary classification models, tracking performance metrics, detecting data drift using Difference of Means, and setting alerts
• Drift detection: understanding data drift (input distribution changes) vs concept drift (relationship between inputs and outputs changes), and how Snowflake detects each
• ML Lineage for compliance: tracing data from source through features, datasets, and models for audit trails, reproducibility, and regulatory requirements
• ML Explainability with Shapley values: computing feature importance scores for model interpretability and transparency requirements
• Cost management for ML workloads: monitoring compute costs across warehouses, SPCS compute pools, and serverless functions. Understanding credit consumption patterns
• Audit logging: tracking who accessed which models, when inference was run, and what data was used for training — critical for regulated industries
• Data Clean Rooms: privacy-preserving collaboration environments for secure cross-organizational ML without exposing raw data

Common Exam Traps