Transform data and engineer features

• DataBrew recipes preserve an ordered sequence of preparation steps
• A Data Wrangler flow must be executed on the full dataset for production output
• Glue job bookmarks support incremental processing of supported sources
• Rewinding a Glue bookmark does not remove existing target output
• Glue bookmark pause processes data without advancing saved bookmark state
• Spark on EMR supports distributed preprocessing in an existing batch environment
• Inference preprocessing must reuse the transformations fitted during training
• Feature metadata enables reuse without embedding sensitive values in descriptions
• Feature Store TTL is based on event time and applies to online records
• Record-level TTL overrides a feature group's default retention duration
• A Firehose transformation must finish within Firehose's invocation limit
• A Firehose transform must identify the disposition of each record
• Firehose transformation batches must fit Lambda request and response limits
• Glue streaming jobs use checkpoints for recovery rather than batch bookmarks
• Windowed feature aggregation needs a stateful streaming processor
• Firehose partition-key extraction needs Lambda when the payload is not directly parseable JSON
• Flink checkpoints and snapshots preserve application state across different restart paths
• Firehose must separate bundled events before dynamically partitioning them
• Flink Studio suits interactive exploration while IDE development supports maintained streaming applications
• Flink's DataStream API supports fine-grained state control and custom timers
• Flink asynchronous I/O overlaps external enrichment requests instead of waiting for each response
• Stable Flink operator UIDs preserve saved-state associations as application flow evolves
• Tumbling windows form disjoint intervals while hopping windows can overlap
• Unseen categories follow the fitted encoder's Invalid handling strategy
• Min-max scaling and standardization produce different numeric representations
• Robust scaling uses statistics resistant to outliers
• A log transformation can compress a positive long-tailed feature
• Binning represents a numeric measurement by its interval
• One-hot encoding represents nominal categories without imposing a numeric ranking
• Feature splitting exposes useful components of a compound value
• Unit-vector normalization rescales each vector rather than each feature column
• Custom tokenization can preserve domain-specific text distinctions
• Term-frequency and document-frequency filters remove different kinds of text features
• Embedding models produce numerical representations rather than generated answers
• Query and document embeddings must use a compatible embedding space
• Multimodal embeddings can place text and images in a shared retrieval space
• Embedding input limits require segmenting oversized documents
• TF-IDF combines local term frequency with corpus-wide term rarity
• Feature hashing avoids a stored vocabulary but can collide terms
• Knowledge-base CSV ingestion separates one content column from metadata columns
• A higher semantic breakpoint threshold generally produces fewer larger chunks
• Multimodal chunking depends on whether media is embedded directly or converted to text
• Overlapping fixed-size chunks preserve context across chunk boundaries
• Semantic chunking uses meaning to locate document boundaries
• Hierarchical chunking combines focused matching with broader returned context
• No chunking preserves each source document as a single retrieval unit
• A knowledge-base data source's chunking strategy cannot be changed in place
• Document metadata supplies retrieval filters separately from embedded content
• Comprehend PII detection locates sensitive spans in text
• PII redaction can conceal characters or replace spans with entity types
• Deterministic encryption preserves equality patterns in protected columns
• Reversible protection requires encryption rather than hashing or removal
• Substitution and shuffling preserve different properties of sensitive data
• Anonymization requires preventing likely identification rather than merely replacing identifiers
• Fine-tuning demonstrations and continued-pretraining corpora serve different purposes
• Bedrock customization data must satisfy both JSONL syntax and the model's schema
• Training conversations should match the structure used at inference
• Distillation prompts should represent the workload the student must perform
• Multimodal fine-tuning examples must keep media linked to the correct conversation
• Domain-specific text augmentation must preserve the intended task and answer

Unlock with Premium — includes all practice exams and the complete study guide.