roadmap
- 01
- 02
- 03
- 04
- 05
- 06
- 07
- 08
Takes you through data analyst interviews, from SQL and Python screens to statistics and case questions. Covers querying, data modeling, analysis libraries, BI dashboards, and machine learning basics.
7660 of 7660 questions in your plan: 0 known, 0 in progress, 0 due, 0 skipped, 7660 new. 0 cut, outside your plan.
A data analyst loop is organised around one artifact, a number, and how far you can be trusted with it. It usually opens with a recruiter call and a SQL screen, live or as a timed take-home, then a conversation on statistics and experimentation, a case round on metrics, often a take-home analysis presented back to a panel, and a behavioral or stakeholder round to close. Python or pandas live coding is common in product analytics loops, Excel in business-facing ones. From senior level up, a round on designing a metric or an experiment is added. The SQL screen rewards predicting the result set, not typing the query: a `LEFT JOIN` whose right-hand table is filtered in `WHERE` quietly becomes an inner join, a total taken after a one-to-many join is inflated by the duplicates, and the second-latest order per customer wants `ROW_NUMBER()` over a partition rather than a self-join. The statistics conversation does the same to an experiment: why a small p-value is not the chance that the change did nothing, why a test read every morning and called on its first good day finds lifts that are not there, and how a metric that moved in every segment can move the other way in aggregate. The case round hands you a drop in a metric and watches you narrow it, by segment, by platform, by a release or a logging change on the same date, before you reach for a story. Levels separate on the questions you put to the number before reporting it: what grain the table is at, which denominator the rate uses, whether the comparison is confounded by who was exposed, and what decision the answer will drive, said out loud, in business terms, to an interviewer who may not read SQL.
The plan starts where the coding screen does, and that screen counts for less here than on an engineering roadmap: [computer science fundamentals](/topics/found-cs-fundamentals) for the vocabulary of cost and logic the later modules borrow, and [the structures and algorithms](/topics/found-dsa) behind the Python exercise some loops still grade on time and space. The languages follow because most of what comes after is exercised in one of them: [Python](/topics/lang-python), comprehensions and generators included, [R](/topics/lang-r) for teams that grew out of statistics, and [SQL](/topics/lang-sql), where [window functions](/topics/lang-sql-window-functions) mark the line between a junior and a middle screen. [Statistics and experimentation](/topics/found-math-statistics) comes straight after the languages because it is the question that follows a query: once you can compute a difference, the interview asks whether it is real, and [A/B testing](/topics/found-math-statistics-ab-testing) is where that gets decided. The plan then turns to where the data lives. [Relational concepts](/topics/db-relational-concepts) explain what the engine does with your statement; [analytical database concepts](/topics/db-analytical-concepts) explain why a warehouse reads a column rather than a row, and so why a predicate cheap in one is expensive in the other. [Data modeling for analytics](/topics/data-modeling) sits on top of both: the grain of a fact table decides whether a join doubles your rows, and a [slowly changing dimension](/topics/data-modeling-scd) is why a customer's region changed underneath last year's revenue. From the warehouse the plan moves into one process. [The dataframe layer](/topics/data-libraries-concepts) is placed after SQL because grouping, merging and reshaping are each a SQL idea with the memory now in your hands, and [BI and visualization](/topics/data-bi) is the output side of the same work. [Orchestration](/topics/data-orchestration) and [ingestion and product analytics](/topics/data-ingestion) look upstream from the consumer's side: enough of scheduling, event collection and [data quality](/topics/data-quality-contracts) to say why a table is a day late or an event count halved after a release. [Git](/topics/dev-git) follows because models, notebooks and dashboard definitions are versioned like any other code. The last technical stretch widens the toolkit. [Machine learning fundamentals](/topics/found-ml-fundamentals) and [scikit-learn](/topics/ai-scikit-learn) cover the models an analyst is asked about and when one beats a rule; [processing engine concepts](/topics/data-processing-concepts) cover what changes when the table no longer fits one machine, so that a slow job is a mechanism rather than a mystery. The plan closes with the [behavioral round](/topics/found-behavioral) and the anatomy of the loop itself, where the stakeholder and presentation rounds analyst interviews lean on are prepared.
Take the SQL module whole before you touch any control, whatever your deadline: the SQL screen is the round almost every analyst loop runs, and the one that ends candidacies early. After it the controls do the triage. A week out, cut level down to the range on the job description, set importance to must know only, and leave study order alone, because by frequency is what a week needs; the statistics, modeling and dataframe questions asked most then come first. With a quarter in hand, move study order to by topics and read the plan in its printed order, writing your case-round answers down as you go, because you will be asked to say them aloud. An analyst arriving from Excel or a business role already has the case instinct and should spend the time on window functions, statistics and the dataframe layer, which turn that instinct into an analysis someone else can rerun. A software engineer moving sideways has the languages and Git and should spend it on experimentation and data modeling, where the vocabulary is new. Senior analysts get most from the experimentation, modeling and ingestion stretches, where their rounds live: designing the metric and defending the pipeline behind it. The plan approaches the data platform as a reader of it, not a builder: enough of orchestration, ingestion and processing engines to explain a late table or a doubled count, not enough to build and run those pipelines as the job. That job is the [Data Engineer roadmap](/roadmaps/data-engineer). The models here stop at scikit-learn; the [AI & Data Scientist roadmap](/roadmaps/ai-data-scientist) carries on from there, into deep learning and into putting a model in front of users.
A data analyst loop is organised around one artifact, a number, and how far you can be trusted with it. It usually opens with a recruiter call and a SQL screen, live or as a timed take-home, then a conversation on statistics and experimentation, a case round on metrics, often a take-home analysis presented back to a panel, and a behavioral or stakeholder round to close. Python or pandas live coding is common in product analytics loops, Excel in business-facing ones. From senior level up, a round on designing a metric or an experiment is added.
The SQL screen rewards predicting the result set, not typing the query: a LEFT JOIN whose right-hand table is filtered in WHERE quietly becomes an inner join, a total taken after a one-to-many join is inflated by the duplicates, and the second-latest order per customer wants ROW_NUMBER() over a partition rather than a self-join. The statistics conversation does the same to an experiment: why a small p-value is not the chance that the change did nothing, why a test read every morning and called on its first good day finds lifts that are not there, and how a metric that moved in every segment can move the other way in aggregate. The case round hands you a drop in a metric and watches you narrow it, by segment, by platform, by a release or a logging change on the same date, before you reach for a story.
Levels separate on the questions you put to the number before reporting it: what grain the table is at, which denominator the rate uses, whether the comparison is confounded by who was exposed, and what decision the answer will drive, said out loud, in business terms, to an interviewer who may not read SQL.
The plan starts where the coding screen does, and that screen counts for less here than on an engineering roadmap: computer science fundamentals for the vocabulary of cost and logic the later modules borrow, and the structures and algorithms behind the Python exercise some loops still grade on time and space. The languages follow because most of what comes after is exercised in one of them: Python, comprehensions and generators included, R for teams that grew out of statistics, and SQL, where window functions mark the line between a junior and a middle screen.
Statistics and experimentation comes straight after the languages because it is the question that follows a query: once you can compute a difference, the interview asks whether it is real, and A/B testing is where that gets decided. The plan then turns to where the data lives. Relational concepts explain what the engine does with your statement; analytical database concepts explain why a warehouse reads a column rather than a row, and so why a predicate cheap in one is expensive in the other. Data modeling for analytics sits on top of both: the grain of a fact table decides whether a join doubles your rows, and a slowly changing dimension is why a customer's region changed underneath last year's revenue.
From the warehouse the plan moves into one process. The dataframe layer is placed after SQL because grouping, merging and reshaping are each a SQL idea with the memory now in your hands, and BI and visualization is the output side of the same work. Orchestration and ingestion and product analytics look upstream from the consumer's side: enough of scheduling, event collection and data quality to say why a table is a day late or an event count halved after a release. Git follows because models, notebooks and dashboard definitions are versioned like any other code.
The last technical stretch widens the toolkit. Machine learning fundamentals and scikit-learn cover the models an analyst is asked about and when one beats a rule; processing engine concepts cover what changes when the table no longer fits one machine, so that a slow job is a mechanism rather than a mystery. The plan closes with the behavioral round and the anatomy of the loop itself, where the stakeholder and presentation rounds analyst interviews lean on are prepared.
Take the SQL module whole before you touch any control, whatever your deadline: the SQL screen is the round almost every analyst loop runs, and the one that ends candidacies early. After it the controls do the triage. A week out, cut level down to the range on the job description, set importance to must know only, and leave study order alone, because by frequency is what a week needs; the statistics, modeling and dataframe questions asked most then come first. With a quarter in hand, move study order to by topics and read the plan in its printed order, writing your case-round answers down as you go, because you will be asked to say them aloud.
An analyst arriving from Excel or a business role already has the case instinct and should spend the time on window functions, statistics and the dataframe layer, which turn that instinct into an analysis someone else can rerun. A software engineer moving sideways has the languages and Git and should spend it on experimentation and data modeling, where the vocabulary is new. Senior analysts get most from the experimentation, modeling and ingestion stretches, where their rounds live: designing the metric and defending the pipeline behind it.
The plan approaches the data platform as a reader of it, not a builder: enough of orchestration, ingestion and processing engines to explain a late table or a doubled count, not enough to build and run those pipelines as the job. That job is the Data Engineer roadmap. The models here stop at scikit-learn; the AI & Data Scientist roadmap carries on from there, into deep learning and into putting a model in front of users.