Mostly harmless, conspicuously useful
The Hitchhiker's Guide to Becoming a Data Engineer
Data engineers arrange for records produced by applications, devices, and business systems to reach people who need them in a state fit for use. This sounds like moving facts from one place to another, which is how the profession introduces itself before revealing timestamps, ownership, storage, cost, privacy, and the duplicate order that has already arrived twice. They build routes into warehouses and lakes, transform awkward source data, test completeness and freshness, and make failures recoverable. Along the way, schemas acquire opinions, late events disregard calendars, and orchestration reports that all retries were performed impeccably while the same task failed six times for the same reason. This guide travels from questioning one table to choosing organisation-wide architecture and governance, with practical stops at pipelines, batch and streaming work, lineage, replay, quality, and operational responsibility. The grand objective is dependable data; the daily evidence is usually a count that reconciles and a pager that has, for once, nothing further to add.
Level 1 · Novice
Ask one table what it means before counting everything it says
Inspect and query one dataset, establish what each row represents, and test whether its values support a bounded answer before the resulting number applies for executive stationery.
Your first assignment is to inspect one dataset and produce a readable SQL query, notes for its important columns, and checks for missing or duplicate records. A dataset is a related collection of records, often displayed as table rows. You work with someone who understands the business question and, when possible, the developer whose application produced the records. Together you establish what the data can honestly answer. The table will otherwise permit almost any count with the calm professionalism of a clerk stamping the wrong form.
A support manager asks how many customers opened tickets yesterday. You count distinct ticket IDs, inspect the newest timestamp, and compare a sample with source records. Twelve rows turn out to be enthusiastic copies created when the application retried a request; counting rows would therefore promote persistence into customer demand. You provide the query, duplicate IDs, corrected total, and freshness evidence. The manager gets a defensible workload figure, while the application team receives a precise failure to investigate rather than a chart with excellent posture.
Words from the spaceship manual, translated
- Dataset
- A related collection of records used for a purpose, such as rows about tickets, orders, or device readings; its usefulness depends on knowing what each record actually represents, a detail datasets rarely volunteer unprompted.
- SQL query
- A written request that selects, filters, joins, groups, or calculates relational data. It preserves the logic behind an answer, unlike a number copied into a message and thereafter treated as ancestral wisdom.
- Duplicate record
- An extra record describing an event or entity already present, often because a source resent data or a load ran twice. It looks harmless until revenue, customers, or sheep are counted with equal confidence.
