FDST-BDS cheat sheet

Data governance is the exercise of decision rights and accountability for data; stewardship is the operational execution of those decisions by named individuals.

A RACI matrix assigns governance roles: Accountable owns the outcome, Responsible does the work, Consulted is asked before decisions, Informed is notified after.

A data owner holds accountability for a domain's data (policy, access approval); a data steward handles day-to-day quality, definitions, and issue triage.

A data governance council is a cross-functional decision body; a data stewardship committee resolves escalations the council delegates.

Policy hierarchy: policy (mandatory intent) > standard (measurable requirement) > procedure (step-by-step) > guideline (recommended).

Data classification tiers (public, internal, confidential, restricted) drive handling rules; classification is a governance decision, not an IT default.

A data issue log records issue, owner, priority, root cause, and resolution status; unresolved issues escalate to the governance council.

Governance maturity models typically progress through initial, developing, defined, managed, and optimizing stages.

Regulatory drivers such as GDPR, CCPA, HIPAA, and SOX map to governance controls: lawful basis, consent, retention, and auditability.

Three-schema architecture: conceptual (business concepts), logical (entities, attributes, relationships, no DBMS specifics), physical (tables, indexes, partitions, storage).

Normalization: 1NF removes repeating groups; 2NF removes partial dependencies on part of a composite key; 3NF removes transitive dependencies on non-key attributes.

Denormalization trades write complexity and storage for read performance; typical in star schemas and reporting layers.

Star schema: one fact table (measures, foreign keys) surrounded by denormalized dimension tables; snowflake schema normalizes dimensions into hierarchies.

Fact table grain is the business meaning of one row; declare grain before choosing measures or dimensions.

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

Kimball (dimensional, bottom-up) versus Inmon (normalized enterprise warehouse, top-down) are complementary patterns, not mutually exclusive.

Data Vault models use hubs (business keys), links (relationships), and satellites (descriptive history) for auditable, parallel-load integration.

Cardinality notation: 1:1, 1:M, M:N; an M:N relationship resolves into an associative entity or bridge table.

Six core dimensions: completeness, accuracy, consistency, timeliness, uniqueness, validity (some frameworks add conformity and integrity).

Completeness = non-null required values / total required values; accuracy requires comparison against a trusted source of truth.

Data profiling measures null rates, distinct counts, min/max, frequency distributions, and pattern conformance before rules are written.

A data quality rule is a testable assertion (e.g., 'ship date >= order date'); a threshold defines pass, warn, and fail bands.

Root cause analysis separates data entry errors, process defects, integration mapping errors, and system design gaps.

Data cleansing is remediation; prevention via validation at entry and constraints at the source is cheaper and more durable.

DQ scorecards aggregate rule results by dimension and domain; trend lines distinguish chronic defects from spikes.

A data quality SLA states the target (e.g., 99.5% completeness), measurement method, reporting cadence, and remediation owner.

Golden record selection for duplicates uses survivorship rules (most recent, most complete, trusted source priority).

ETL extracts, transforms, then loads; ELT loads raw first and transforms in the target (common with cloud warehouses and columnar engines).

Batch integration runs on schedules; CDC (change data capture) streams inserts, updates, and deletes from logs for near-real-time sync.

Idempotent loads are re-runnable without duplicating rows; achieved via merge/upsert on business keys or partition overwrite.

Full load replaces the target; incremental load applies only changed records and requires a reliable high-water mark.

Data lineage traces source-to-target flow; impact analysis uses the same graph in reverse to find downstream consumers.

Interoperability layers: syntactic (format/protocol), semantic (shared meaning), and organizational (process and policy agreement).

API styles: REST (resources, HTTP verbs), SOAP (XML contract, WS-* standards), GraphQL (client-specified queries), and event streaming (pub/sub).

Message delivery semantics: at-most-once (may lose), at-least-once (may duplicate), exactly-once (requires idempotency or transactions).

Schema evolution strategies: backward compatible (new consumers read old data), forward compatible (old consumers read new data), full compatibility.

CIA triad: confidentiality (only authorized access), integrity (unmodified and accurate), availability (accessible when needed).

Encryption at rest protects stored data; encryption in transit (TLS) protects the wire; key management (KMS, rotation, separation of duties) is the real control.

Tokenization replaces a value with a non-reversible surrogate mapped in a vault; masking partially obscures values for non-production or display.

Authentication verifies identity; authorization grants permissions. RBAC uses roles, ABAC evaluates attributes (user, resource, context).

Least privilege and separation of duties limit blast radius; privileged access management adds just-in-time elevation and session logging.

Privacy principles: purpose limitation, data minimization, storage limitation, accuracy, integrity and confidentiality, accountability.

PII is any data that can identify a person directly or indirectly; pseudonymization reduces risk but is not anonymization.

Data subject rights under GDPR include access, rectification, erasure, restriction, portability, and objection; response timelines are typically one month.

Defense in depth layers controls: perimeter, network segmentation, identity, application, data, and monitoring/detection.

A data warehouse is subject-oriented, integrated, time-variant, and non-volatile; a data mart is a subset scoped to one business area.

OLTP optimizes short transactions with normalized schemas; OLAP optimizes large scans and aggregation with columnar storage and denormalized models.

Columnar storage improves analytical scan performance via compression and column pruning; row storage favors single-row transactional access.

Partitioning prunes scanned data by a key (usually date); clustering or sort keys co-locate related rows within partitions.

Aggregates and materialized views precompute results; refresh strategy (full, incremental, on-query) determines staleness.

BI semantic layer defines measures, dimensions, and hierarchies once so reports share consistent definitions.

Additive measures sum across all dimensions; semi-additive measures (balances) sum across some dimensions but not time; non-additive (ratios) cannot be summed.

Conformed dimensions are shared across fact tables and enable drill-across; without them, cross-process comparison is unreliable.

Dashboard design: match chart to question (trend = line, comparison = bar, distribution = histogram, part-to-whole = stacked or treemap).

Metadata types: technical (schemas, types, lineage), business (definitions, owners, glossary), operational (job logs, row counts, runtimes).

A business glossary defines business terms; a data dictionary defines technical attributes; a data catalog indexes and connects both.

Active metadata is machine-readable and used by tools (lineage, policy tags); passive metadata is documentation for humans.

MDM domains: customer, product, supplier, employee, location, and reference data (code sets such as ISO country codes).

MDM implementation styles: registry (index only), consolidation (aggregate for reporting), coexistence (bidirectional sync), centralized (single authoritative system).

Entity resolution matches records via deterministic rules (exact keys) or probabilistic scoring (similarity thresholds), then merges into a golden record.

Reference data is a controlled value set shared across systems; it changes rarely and must be synchronized, not re-created per application.

Data lineage supports impact analysis, root cause debugging, and regulatory evidence; it is metadata, not documentation of intent.

Analytics types: descriptive (what happened), diagnostic (why), predictive (what will happen), prescriptive (what should be done).

Reporting modes: operational/real-time, tactical/periodic, and strategic/ad hoc; each implies different latency and aggregation needs.

A KPI needs an explicit formula, data source, owner, target, and measurement frequency; without these it is a vanity metric.

Correlation does not imply causation; establishing causation requires controlled experiments or quasi-experimental designs.

Statistical significance depends on effect size and sample size; a p-value below alpha does not measure business importance.

Sampling bias (selection, survivorship, response) distorts conclusions more often than small samples do.

Time-series decomposition splits a series into trend, seasonality, and residual; seasonality must be removed before comparing periods.

A/B testing requires random assignment, a pre-declared primary metric, sufficient power, and guardrail metrics to catch harm.

Data visualization integrity: consistent axis scales, no truncated bar axes, and denominators shown for percentage charts.