ICCP CDP cheat sheet

DAMA-DMBOK2 defines 11 data management knowledge areas; data management is the umbrella discipline, data governance is one of the 11, not a superset.

Data lifecycle stages: create/acquire, maintain/use, archive, purge. Retention rules attach to the record, not the storage medium.

Metadata types: business (definitions, owners), technical (schema, lineage, ETL jobs), operational (logs, job stats, access counts).

Data quality dimensions commonly tested: completeness, accuracy, consistency, timeliness, uniqueness, validity. Each needs a measurable rule.

Master data vs reference data: master data describes core entities (customer, product, party); reference data is controlled code sets (country codes, status codes).

MDM implementation styles: registry (index only), consolidation (read-only warehouse), coexistence (bidirectional sync), centralized (single transactional hub).

ACID applies to transactional stores; BASE (basically available, soft state, eventual consistency) describes many distributed NoSQL designs.

Normalization: 1NF atomic values, 2NF no partial dependency on part of a composite key, 3NF no transitive dependency on non-key attributes.

Data lineage answers 'where did this value come from and what consumes it'; impact analysis is the reverse direction (what breaks if I change it).

Kimball dimensional model: fact tables hold measures plus foreign keys at a declared grain; dimension tables hold descriptive context. Declare grain first.

Star schema = denormalized dimensions around one fact; snowflake schema = normalized dimension hierarchies joined outward.

Slowly changing dimensions: Type 1 overwrites (no history), Type 2 adds a new row with version/effective dates and a current flag, Type 3 adds a prior-value column.

Grain mismatch is the most common fact-table defect: every measure must be true at the declared grain (e.g., order line, not order).

Additive measures sum across all dimensions; semi-additive (balances, inventory) sum across some but not time; non-additive (ratios, percentages) must be recomputed from components.

Inmon approach: enterprise data warehouse built normalized (3NF) first, then departmental data marts derived from it.

Factless fact tables capture events or coverage (who attended, which product was on promotion) with no numeric measure.

Conformed dimensions are shared across fact tables with identical keys and attribute definitions, enabling drill-across.

Aggregate navigation and materialized summaries trade storage for query latency; they must reconcile exactly to base-grain totals.

Operating model components: decision rights, accountability assignments, policies and standards, controls, and escalation paths.

Roles: data owner (accountable, funds and approves), data steward (executes quality and definition work), data custodian (technical storage and security), CDO (enterprise program lead).

Data steward is typically a business role; custodian is typically IT. Confusing the two is a frequent exam distractor.

Policy hierarchy: policy (why/what, mandatory) → standard (specific requirement) → procedure (how) → guideline (recommended).

Data classification tiers (e.g., public, internal, confidential, restricted) drive handling, encryption, retention, and access-approval requirements.

GDPR principles: lawfulness/fairness/transparency, purpose limitation, data minimization, accuracy, storage limitation, integrity and confidentiality, accountability.

Data subject rights under GDPR include access, rectification, erasure, restriction, portability, and objection; response generally within one month.

A business glossary defines business terms with owners and approved definitions; a data dictionary documents physical structures. Glossary ≠ dictionary.

Governance metrics should be outcome-based (issue closure rate, certified-data coverage, stewardship coverage) rather than activity counts.

Analytics maturity ladder: descriptive (what happened) → diagnostic (why) → predictive (what will happen) → prescriptive (what should be done).

Supervised learning needs labeled targets; unsupervised learning (clustering, association rules) does not. Classification predicts a class, regression predicts a continuous value.

Bias-variance: high bias underfits (too simple), high variance overfits (too sensitive to training noise). Regularization trades variance for bias.

Precision = TP/(TP+FP); recall = TP/(TP+FN); F1 = harmonic mean of precision and recall. Choose the metric that matches the cost of errors.

Confusion matrix cells: true positive, true negative, false positive (Type I), false negative (Type II).

Train/validation/test split: fit on train, tune on validation, report once on test. Leakage occurs when test information influences training.

Correlation measures linear association and does not establish causation; confounding variables can create or hide apparent effects.

k-fold cross-validation averages performance over k held-out folds; stratified folds preserve class proportions for imbalanced data.

A p-value is the probability of data at least as extreme assuming the null hypothesis is true; it is not the probability the null is true.