Deep Battle Cards A–F
A — Big Data vs Data Science
Definitions: Big Data = management/environment pressure from scale, speed, variety and related V's. Data Science = iterative analytical practice using statistics, mining, ML, integration and modeling.
Deciding distinction: data/environment vs analytical reasoning/model lifecycle.
Inputs: varied sources/platform needs vs business question + source data + hypotheses.
Outputs: managed data environment vs findings/models/predictions/prescriptions.
Scenario pair: cannot store mixed sensor feed efficiently → Big Data architecture; ask whether feed predicts failure → Data Science.
Trap: large data set automatically means a Data Science problem.
Hook: environment vs analysis.
Source: pp. 471–474.
B — Descriptive vs Predictive vs Prescriptive
Definitions: historical explanation vs probable future outcome vs recommended intervention.
Deciding question: what happened/why? what likely? what should we do?
Scenario pair: churn probability → predictive; best retention offer → prescriptive.
Trap: forecast = prescription.
Hook: explain → predict → act.
Source: pp. 474–475, 481–482.
C — ETL vs ELT
Definitions: transform/integrate before load vs ingest/load first and transform/integrate later for use.
Purpose: different assumptions about when structure is imposed.
Scenario pair: stable finance warehouse load → ETL; evolving raw sources for exploratory models → ELT.
Nearest confusion: ELT still needs transformation, Metadata, DQ, and governance.
Hook: when does transformation happen?
Source: pp. 471–472, 476–478.
D — Data Lake vs Data Swamp
Definitions: flexible varied-data environment vs lake whose context/management is missing.
Deciding controls: inventory, Metadata, provenance, relationships, DQ, ownership, discoverability.
Scenario pair: raw files with source/structure/lineage/quality tags → lake; randomly named files with no provenance → swamp.
Trap: raw = unmanaged.
Hook: lake + context = asset; lake − context = swamp.
Source: pp. 478–480, 499–500.
E — Batch vs Speed vs Serving
Definitions: complete historical processing vs current/low-latency processing vs merged/exposed consumer view.
Purpose: balance completeness and latency.
Scenario: fraud alert seconds after event → speed; full nightly recomputation → batch; dashboard joining both → serving.
Trap: Big Data should always be real-time.
Hook: history, now, view.
Source: pp. 479–480.
F — Supervised vs Unsupervised vs Reinforcement
Definitions: labeled examples vs unlabeled discovery vs goal/reward feedback.
Deciding signal: known target, no target, or state/action feedback.
Scenario: spam labels → supervised; natural customer segments → unsupervised; game agent learning from rewards → reinforcement.
Trap: “no human in the loop” = unsupervised.
Hook: label, pattern, reward.
Source: pp. 480–481.