← DMML vault

BITS · DMML · cheatsheet · CS1-L5

Data Management for Machine Learning -- cheatsheet.

dmml data-engineering pipelines dataops

One-card revision for DSE/AIML ZG529. The course is not "databases only." It is the full path from raw data to reliable ML: representation, storage, governance, architecture, movement, transformation, serving, automation, monitoring and feedback into production systems.

Course map

DeckThemeMust remember
CS1Data representationsData -> information -> knowledge -> wisdom; formats, models, layouts, serialization, OLTP/OLAP, ACID/BASE.
CS2Data management fundamentalsData as asset/liability, data sensitivity, reliability, governance/security, platform components and data life cycle.
L3Data architecturesCentralized/decentralized/hybrid; warehouse, lake, lakehouse, marts, mesh, fabric, Lambda and Kappa.
L4Data pipelinesFlow through databases, services and message brokers; modern data stack; ETL/ELT, batch, stream, CDC, Beam/Dataflow.
L5Infrastructure and DataOpsIngestion, transformation, orchestration, storage systems, serving, automation, cloud/security, CT/CD and observability.

CS1: data representation basics

CS1

  • Data is raw material. Information is organized data. Knowledge is actionable information. Wisdom is integrated understanding used for judgement.
  • Data management = ingesting, processing, securing and storing organizational data so it can improve business outcomes and support ML systems.
  • ML system as pipeline: source data -> transformation -> features/training data -> model -> predictions -> feedback. Bad data can break ML while ordinary data movement still "looks green."
  • Structured data: fixed schema, tables, rows, constraints. Semi-structured: partial schema, JSON/XML/logs. Unstructured: text, images, audio, video, emails.
  • Data models: relational for strong structure; hierarchical for tree-like parent-child data; graph for relationship-first queries; document for flexible JSON/XML objects; key-value for lookup/cache patterns.
  • Declarative query: say what result you want; optimizer decides how. Imperative/procedural: specify step-by-step instructions; easier to reason locally, harder for systems to optimize globally.
  • Row layout: faster when reading/writing full records. Column layout: faster for analytical scans over selected columns, compression and feature-heavy tables.
  • Serialization: convert in-memory data into a storable/transmittable format. CSV/JSON/XML are text-friendly; Protobuf/Avro are compact binary formats with schema support.
  • OLTP: many small transactions, current state, strict consistency. OLAP: analytical queries, aggregations, drill-down, roll-up, slice-and-dice and trend analysis.
  • ACID: atomicity, consistency, isolation, durability. BASE: basically available, soft state, eventual consistency; common in highly available distributed systems.

CS2: data as asset, liability and platform

CS2

  • Data as asset: valuable when accessible, accurate, governed and reusable for decisions or models.
  • Data as liability: dangerous when it contains PII, violates consent boundaries, is stale, unneeded, poorly secured or expensive to retain/delete.
  • More data != better ML. More wrong, biased, missing, delayed, duplicated or legally risky data makes systems worse.
  • Data reliability minimums: durability (not lost), consistency (same truth across copies), version/time travel where needed, availability and freshness.
  • Data management components: integration/processing, storage, governance and security.
  • Data governance: policies, stewardship, taxonomy, metadata, standards, quality rules, lineage and ownership.
  • Data security: roles, least privilege, access control, PII controls, encryption, audit and separation of duties.
  • Data platform: central repository + processing house + pipeline system that ingests, normalizes, transforms and delivers data to consumers.
  • Life cycle: generation -> ingestion -> validation -> cleaning -> enrichment -> storage -> transformation -> serving/visualization -> feedback/retention.
  • Data engineering role: design, build and maintain systems that collect, clean, move, store and serve data reliably.

L3: architecture patterns

L3

PatternBest whenRisk
CentralizedNeed one control plane for governance, auditing and enterprise reporting.Can bottleneck domain teams and become slow to change.
DecentralizedBusiness units move fast with their own data and reporting.Silos, duplication and inconsistent definitions.
HybridDomains need ownership but enterprise still needs golden sources and shared standards.Requires clear domain boundaries and governance.
WarehouseStructured analytics, BI, curated schema, dimensional models.Less flexible for raw/semi-structured high-velocity data.
LakeRaw/semi-structured/unstructured data, cheap storage, ELT and data science exploration.Can become a data swamp without metadata and governance.
LakehouseNeed lake flexibility plus warehouse governance/performance.Operational complexity and table-format discipline.
Data meshLarge org with domain-owned data products and self-serve platform.Needs mature ownership, standards and product thinking.
Data fabricNeed integrated access across lakes, warehouses, databases and SaaS systems.Can hide complexity rather than solve data-quality problems.
  • Star schema: central fact table + denormalized dimensions. Simple and fast for BI.
  • Snowflake schema: normalized dimensions split into sub-dimensions. Saves storage and improves organization but adds joins.
  • Data mart: department/line-of-business subset of the warehouse, optimized for a specific consumer group.
  • Lambda architecture: batch path for historical correctness + speed path for low-latency results + serving layer. Powerful but duplicates logic.
  • Kappa architecture: one streaming backbone for ingestion, processing and serving. Simpler logic; depends on robust replayable logs.

L4: data pipelines

L4

  • Data pipeline: sequence of steps that moves data from source to destination while applying ingestion, validation, transformation and delivery rules.
  • Through databases: shared DB acts as integration point. Watch schema evolution, rolling upgrades, old code reading new fields, and expensive migrations.
  • Through services: client-server, SOA, microservices, REST/SOAP/RPC. APIs reduce direct DB coupling but network calls are not local method calls.
  • Through message brokers: asynchronous buffer between producers and consumers. Improves resilience, decouples components and supports redelivery.
  • Modern data stack: cloud-based ingestion, storage, transformation, governance, analysis and serving tools designed for fast setup and self-service.
  • ETL: transform before loading; useful when target warehouse needs clean structured data upfront. ELT: load raw data first, transform inside scalable warehouse/lakehouse.
  • Batch pipeline: scheduled, bounded data. Good for monthly reports, daily aggregates and training snapshots.
  • Stream pipeline: continuous events. Good for monitoring, fraud, recommendation signals, operational analytics and low-latency ML features.
  • CDC: change data capture streams inserts/updates/deletes from source databases to downstream systems.
  • Micro-pipeline: break complex pipeline into small stages so requirements, ingestion, validation, transformation and delivery can evolve independently.

L5: infrastructure, DataOps and serving

L5

  • Ingestion infrastructure: connectors for databases, APIs, Kafka/streams, file shares and cloud buckets.
  • Transformation infrastructure: code or SQL that filters, joins, cleans, masks PII, aggregates and models data into usable shapes.
  • Orchestration: schedules and coordinates pipeline tasks, retries failures, manages dependencies, backfills data and tracks execution state.
  • Storage raw ingredients: HDD/SSD/RAM, network, CPU, serialization, compression and caching.
  • Storage systems: local file systems, NAS, block storage, object storage, memory caches, HDFS and streaming storage.
  • Object storage: immutable key-value object store for files/blobs; cheap and scalable but no random writes/append.
  • Serving: file exchange, databases, streaming systems, query federation, notebooks, dashboards, reverse ETL, embedded analytics and ML feature/model serving.
  • Reverse ETL: send processed data back into operational tools, e.g. lead scores from warehouse into CRM.
  • DataOps: DevOps-style discipline for data pipelines: versioning, testing, monitoring, observability, automation, governance and continuous improvement.
  • CT/CD for ML: code changes trigger tests and training; approved models move through deployment; data/model drift and data quality feed back into retraining.

Fast comparisons

PairDifference
OLTP vs OLAPOLTP optimizes transactional writes and current state; OLAP optimizes analytical reads, aggregates and historical slicing.
ETL vs ELTETL transforms before load; ELT loads raw data first and transforms inside the target platform.
Warehouse vs lakeWarehouse is curated/schema-first for analytics; lake is flexible/raw storage for varied data.
Lake vs lakehouseLakehouse adds table formats, governance and query performance controls on top of lake storage.
REST vs RPCREST models resources over HTTP; RPC models remote calls like functions. RPC can hide network failure semantics too much.
Batch vs streamBatch handles bounded chunks on a schedule; stream handles unbounded events continuously.
DataOps vs DevOpsDevOps optimizes code delivery; DataOps applies similar discipline to data pipelines, data quality and analytical delivery.

Exam gotchas

  • If a question says relationship-first query, think graph model.
  • If it says high-volume analytics over few columns, think columnar storage.
  • If it says strict transactions, think OLTP and ACID.
  • If it says enterprise BI and dashboards, think warehouse, dimensional schema, OLAP.
  • If it says raw mixed-format data for data science, think data lake/lakehouse.
  • If it says two separate paths for historical + realtime, think Lambda.
  • If it says one replayable stream path, think Kappa.
  • If it says CRM needs model scores back, think reverse ETL.
  • If it says pipeline green but model bad, think data sensitivity, drift, quality, missing segment or feature breakage.
One-line memory hook. DMML = make data reliable enough that ML can trust it: collect it legally, model it correctly, move it predictably, transform it reproducibly, serve it fast enough, and observe it continuously.

More in this vault

← DMML vault slides explained →
© cvam -- written in plaintext, served warm