DEA Level 3 flashcards

32 cards. Read the term and the definition here. The deck at the top of this page is optional.

Data steward

Role accountable for data quality, definitions, and issue resolution within an assigned data domain, working under the domain data owner's risk and access decisions.

Critical data element

A field whose quality materially affects regulatory reporting, financial statements, or a key business process, and therefore requires mandatory stewardship, quality rules, and lineage.

Business glossary

A governed inventory of business terms with approved definitions, synonyms, and owners, used to align meaning across systems independently of physical column names.

Data quality score formula

Conforming records divided by total records assessed, multiplied by 100, computed per quality dimension and per critical data element rather than as one blended figure.

Third normal form

A table is in 3NF when it is in 2NF and no non-key attribute depends transitively on the primary key through another non-key attribute.

Fact table grain

The precise business meaning of one fact row, such as one order line per day; it must be declared before measures and dimensions are chosen.

Slowly changing dimension Type 2

A dimension design that preserves history by inserting a new row with new surrogate key, effective dates, and a current-record flag when an attribute changes.

Star versus snowflake schema

A star keeps dimensions denormalized for fewer joins and simpler queries; a snowflake normalizes dimension hierarchies, reducing redundancy at the cost of additional joins.

Data Vault components

Hubs store business keys, links store relationships among hubs, and satellites store time-stamped descriptive attributes, giving auditable, parallel-loadable warehouse structures.

CAP theorem

A distributed data system facing a network partition must sacrifice either consistency or availability; PACELC adds the latency-versus-consistency trade-off when no partition occurs.

Idempotent pipeline

A pipeline whose rerun produces the same target state as a single run, typically implemented with merge-upsert on a business key or partition overwrite.

Log-based change data capture

Reading the source database transaction log to emit inserts, updates, and deletes with low source impact and near-real-time latency, without polling full tables.

Exactly-once delivery

An end-to-end guarantee usually built from at-least-once delivery plus idempotent writes or deduplication keys, since true exactly-once requires transactional sinks.

Stream watermark

A progress marker in stream processing that tells the engine how long to wait for out-of-order events before finalizing a window result.

ELT versus ETL

ETL transforms data before loading into the target; ELT loads raw data first and transforms inside the target platform, preserving raw history and using target compute.

Small-file problem

Accumulating many tiny files in a distributed or columnar store inflates metadata overhead and slows scans; target files in the 128 MB to 1 GB range.

Precision and recall

Precision is true positives divided by predicted positives; recall is true positives divided by actual positives; F1 is their harmonic mean.

ROC-AUC versus PR-AUC

ROC-AUC measures ranking quality across thresholds and is insensitive to class balance; PR-AUC is more informative when the positive class is rare.

Bias-variance trade-off

High bias means underfitting with high train and validation error; high variance means overfitting with low train but high validation error.

L1 versus L2 regularization

L1 (lasso) penalizes absolute coefficients and can zero them out for feature selection; L2 (ridge) penalizes squares and shrinks coefficients smoothly.

Data leakage

Training on features that encode target information unavailable at prediction time, producing validation scores that will not hold in production.

Stratified k-fold cross-validation

A resampling scheme that preserves the class proportion in each fold, used for imbalanced classification; time-series data instead requires forward-chaining splits.

Population stability index

A drift metric comparing the distribution of a model input or score between a reference period and a current period; higher values indicate greater shift.

Data classification tiers

Labels such as public, internal, confidential, and restricted that determine required encryption, access approval, retention, and sharing controls for a dataset.

Pseudonymization versus anonymization

Pseudonymization replaces direct identifiers while re-identification remains possible with separately held data; anonymization must make re-identification not reasonably possible.

Static versus dynamic masking

Static masking irreversibly alters data in non-production copies; dynamic masking applies at query time and leaves the stored values unchanged.

Attribute-based access control

An authorization model granting access from attributes of user, resource, and context, such as department, classification, location, and time of request.

Least privilege

The principle that each identity receives only the access required for its task, granted for a defined scope and duration and reviewed periodically.

Data subject access request

A privacy right allowing an individual to obtain their personal data and processing details; it requires an identified system of record and a response deadline.

Data product

A managed data asset with an owner, documented schema or contract, quality service-level objective, and identified consumers, treated as a reusable deliverable.

Federated operating model

A governance structure where domain teams build and own data while a central function sets standards, policy, and shared platform services.

Data maturity assessment

Scoring of capabilities such as governance, quality, architecture, and analytics along a staged scale, used to sequence roadmap investment by weakest critical capability.