The Data Engineer's Role

Where you fit in the data team

What does a data engineer actually do?

WhatWhyHow
What it is

A data engineer designs, builds, operates, and evolves the systems that move data from where it is created (operational sources) to where it is needed (analytics, ML, AI, products, and decisions). The role sits between data producers (software engineers, DevOps, external partners) and data consumers (analysts, data scientists, ML engineers, business leaders).

Why it matters

Without reliable data pipelines, analysts query stale or incorrect data, ML models train on garbage, dashboards show contradictory numbers, and leadership loses trust in data-driven decisions. The data engineer ensures dependable flow: not just movement, but quality, governance, timeliness, and cost control.

Core responsibilities
  • Ingestion: extract data from operational systems, APIs, files, streams
  • Storage: choose and manage data lakes, lakehouses, and warehouses
  • Transformation: clean, validate, conform, and model data for consumption
  • Orchestration: schedule, monitor, retry, and evolve pipeline workflows
  • Serving: publish data through APIs, BI tools, feature stores, and ML platforms
  • Governance: enforce security, privacy, quality, lineage, and cost policies

Data engineer's place within the data team

Software Engineers
DevOps Engineers
External Partners
Data Engineers
Build the trusted flow
Data Analysts
Data Scientists
ML Engineers
Business Leaders

OLTP vs. OLAP

Two fundamentally different purposes

OLTP: Online Transaction Processing

WhatWhy
What it is

OLTP systems are the operational backbone of applications. They handle day-to-day transactions: bank transfers, e-commerce orders, user registrations, inventory updates. Think PostgreSQL, MySQL, SQL Server, Oracle: systems optimized for fast, concurrent, small-transaction CRUD (Create, Read, Update, Delete) operations.

Key characteristics
  • Row-based storage: entire records stored together for fast single-row access
  • Normalized schema: typically 3NF to minimize redundancy and ensure consistency
  • High concurrency: hundreds or thousands of simultaneous small transactions
  • Low latency: millisecond response times for point lookups
  • ACID guarantees: transactions are atomic, consistent, isolated, and durable

OLAP: Online Analytical Processing

WhatWhy
What it is

OLAP systems serve analytical workloads: aggregations, trends, comparisons, reports, and machine learning features. They process large volumes of historical data to answer questions like "What was total revenue by region last quarter?" Think Snowflake, BigQuery, Redshift, ClickHouse: systems optimized for scanning billions of rows across a few columns.

Key characteristics
  • Columnar storage: columns stored separately for efficient scans and compression
  • Denormalized or star schema: fewer joins, faster analytical queries
  • Large scan volume: queries touch millions or billions of rows
  • Fewer concurrent users: analysts, not thousands of app users
  • High compression: 5 to 30x compression ratios on homogeneous column data
DimensionOLTPOLAP
PurposeRun the businessAnalyze the business
Storage layoutRow-basedColumnar
SchemaNormalized (3NF)Denormalized / Star
Query patternPoint lookups, small transactionsFull-table scans, aggregations
LatencyMillisecondsSeconds to minutes
Data volumeGBs (current state)TBs to PBs (historical)
UsersApplications, microservicesAnalysts, data scientists, BI tools
ExamplesPostgreSQL, MySQL, SQL ServerSnowflake, BigQuery, Redshift, ClickHouse

The OLTP to OLAP journey is the core of data engineering

Key insight

Data engineering is fundamentally about moving data from OLTP systems (where it is created) to OLAP systems (where it is analyzed). This is the core pipeline: extract from operational sources, transform to make it trustworthy, and load into analytical storage. Every tool, technique, and architecture decision serves this journey.

OLTP
PostgreSQL
Extract
CDC / Batch
Transform
Clean & Model
OLAP
Snowflake

SQL & Data Profiling

The engineer's first diagnostic tool

Why SQL is non-negotiable

What you need

SQL is the universal language of data. Every database, warehouse, lakehouse, and most processing engines speak SQL. Data engineers use SQL for extraction, transformation (dbt is SQL-first), quality checks, profiling, and debugging. You need fluency in:

  • Window functions: ROW_NUMBER, RANK, LAG, LEAD, running totals
  • CTEs: Common Table Expressions for readable, modular queries
  • Aggregations: GROUP BY, HAVING, ROLLUP, CUBE
  • Joins: INNER, LEFT, FULL OUTER, CROSS, self-joins, anti-joins
  • Subqueries: correlated and uncorrelated
  • Data types: dates, timestamps, strings, numerics, JSON

Data profiling: know your data before building

WhatWhyHow
What it is

Data profiling is the systematic examination of data to understand its structure, content, quality, and relationships. It is the diagnostic step before building any pipeline, like a doctor examining a patient before prescribing treatment.

Why it matters

Building a pipeline without profiling leads to silent failures: unexpected NULLs crash transformations, wrong data types cause casting errors, undiscovered duplicates inflate metrics, and missing values create misleading dashboards. Profile first. Always.

What to check
  • Completeness: percentage of NULL values per column
  • Uniqueness: duplicate counts, candidate keys
  • Distribution: min, max, mean, median, standard deviation
  • Cardinality: count of distinct values
  • Patterns: date formats, email formats, phone numbers
  • Referential integrity: orphaned foreign keys
  • Freshness: latest timestamp, gaps in time series

Framing Requirements

Start with the decision, not the tool

The five questions every project starts with

Before choosing any tool, answer these
  • 1. What decision does this data support? Know the end use case: dashboard, ML model, operational alert, or product feature.
  • 2. What is the latency requirement? Does the consumer need data within seconds (streaming), minutes (micro-batch), hours (scheduled batch), or next day?
  • 3. What is the data volume? GBs per day? TBs? This shapes storage format, compute engine, and cost model.
  • 4. Who owns the source data? Understanding ownership, contracts, and schema stability determines the ingestion approach.
  • 5. What are the compliance constraints? PII, GDPR, CCPA, data residency, retention policies, and audit requirements.

Interactive: Match the requirement to the approach

Select a scenario to see which data engineering approach fits best.

Executive dashboard: "What happened last quarter?"

Latency: Next-day is fine (batch). Volume: Moderate, aggregated. Approach: Scheduled ELT into a warehouse with a star schema model and BI tool on top. Focus on data quality, freshness SLAs, and consistent metric definitions. dbt is ideal for the transformation layer.

Understanding Data Consumers

Design for who uses the data

Different consumers, different needs

ConsumerNeedsData shapeLatency
Business analystReports, KPIs, ad-hoc queriesStar schema, aggregatesDaily / hourly
Data scientistFeatures, training data, experimentsWide tables, historyBatch is often fine
ML engineerFeature store, model servingKey-value, versionedNear real-time
Product teamEmbedded analytics, recommendationsAPIs, pre-computedLow latency
Data engineerMonitoring, lineage, qualityMetadata, logsReal-time alerts
Key insight

The same underlying data may need to be served in multiple shapes, at different latencies, through different interfaces. This is why the pipeline architecture matters: a well-designed system serves all consumers without building separate pipelines for each.

Knowledge Check: Module 1

1. What is the primary purpose of an OLTP system?

OLTP systems are designed for operational workloads: fast CRUD operations, high concurrency, and ACID guarantees. They run the business day-to-day (e.g., processing bank transfers, e-commerce orders).

2. Which is the best first question when starting a data engineering project?

Architecture is a response to requirements. Start with the business decision, SLA, risk, and constraints. Tool selection comes after the job-to-be-done is understood.

3. Why is data profiling done before building a pipeline?

Profiling reveals data quality issues (nulls, duplicates, unexpected formats) before they become pipeline bugs. Building without profiling leads to silent failures and incorrect downstream data.

4. Where does the data engineer sit within the data team?

Data engineers are the bridge between systems that create data and people/systems that consume it. They build the trusted flow that connects producers to consumers.

5. An analyst needs last quarter's revenue by region on a Monday morning dashboard. What approach fits?

Next-day aggregated metrics fit a scheduled batch pattern. ELT into a warehouse with star schema modeling gives the analyst fast queries and consistent metrics. Streaming is overkill; querying OLTP directly would harm production performance.