StrataScratch vs DataDriven for Data Engineer Interviews

StrataScratch was founded in 2017 and sells itself as data science and AI interview preparation. Its core is more than 1,000 coding questions labeled by the company that asked them. You can solve them in several SQL dialects and in R, and the Python versions run in pandas and Polars as well as PySpark. Several hundred concept questions cover statistics and probability alongside product sense and machine learning system design, and there are take-home projects in hosted notebooks and an AI mock interviewer. Data engineer is one of more than a dozen roles it lists. Data analyst is another, as are data scientist and ML engineer.

That breadth shapes the questions. A typical StrataScratch question asks for a business metric from a few tables: monthly revenue growth, the top 3 products per category, the share of users active in consecutive months. Those questions are exactly what an analyst or data scientist faces, and they overlap with the easier half of a data engineering SQL round.

DataDriven covers only data engineering. Its {COUNT:total} problems are split between code and design: the code problems are in SQL and Python, with PySpark alongside, and the design problems cover data modeling and pipeline design. Every one of them has a mock interview variant. The difference shows less in any single question than in the rounds it leaves out: none of DataDriven is spent on statistics or A/B tests, and all of it is spent on the work a data engineering loop examines.

StrataScratch is strongest in the SQL round. Its company-labeled questions drill joins and aggregations at speed, along with window functions and date logic, and a phone screen at most companies asks for exactly that: a correct query on a small schema in 20 to 30 minutes.

A data engineering SQL round goes further in 2 directions. The first is the kind of problem. Some are about messy events, such as deduplicating events that arrive twice or sessionizing clickstreams by the gap between events. Others ask you to find gaps and islands in daily activity or build a slowly changing dimension, and some want a load that can rerun without double counting. These are pipeline problems written in SQL, and they're what separates a data engineering round from an analytics one.

The second is correctness under data you haven't seen. An interviewer will ask what happens when 2 rows tie, when a key is NULL or when a day has no events. DataDriven runs every SQL answer against 10 randomized datasets, so a query that only works on the sample, because it skipped a tiebreaker or assumed every day has a row, fails the way it would in review. The SQL interview questions for data engineers cover these patterns, and the SQL practice problems let you run them.

StrataScratch's Python questions are mostly its SQL questions solved with a DataFrame, filtering and grouping in pandas and then merging and ranking. That is real Python practice, and some data engineering screens do ask for pandas, but it trains analysis on data that already fits in memory and already has a clean shape.

A data engineering Python round usually hands you raw records and asks you to make them reliable: parse a log file with malformed lines, deduplicate a stream of events by key and timestamp, merge overlapping intervals, flatten nested JSON, retry a flaky API call with backoff, or process a file too large to load at once with a generator. The interviewer is watching how you handle bad input and memory, and whether you catch the edge cases a pipeline meets every night.

DataDriven's {COUNT:python} Python problems are written in that shape, with plain Python first and libraries second. If your target companies run a pandas screen, StrataScratch's DataFrame questions are good extra practice on top.

Both sites run PySpark in the browser, so either one will make the DataFrame API familiar. On StrataScratch, PySpark is another language for the same analytics questions: you write the revenue or ranking query as groupBy and agg calls with a window on top, instead of SQL.

A Spark interview for a data engineer rarely stops at a correct DataFrame. The follow-up questions are about how the job runs: why a join shuffles both sides and when a broadcast join avoids it, what a skewed key does to 1 task while the others sit idle, and how partitioning on write decides how much a later query has to read. DataDriven's {COUNT:spark} Spark problems include the SQL problems written for Spark, and its mock interviews ask those follow-up questions after the code runs.

The data modeling round is where the 2 sites differ most. StrataScratch covers data modeling as concept questions: you read a prompt and think through or type an answer. Then you compare it with a written solution. That helps you learn the vocabulary, but the round itself asks for something else.

In a modeling round the interviewer gives you a business process to model, rides for instance, and you design the tables live; orders and subscriptions are other typical prompts. The Kimball Group's 4-step dimensional design process is the order most interviewers expect you to reason in: select the business process, declare the grain, identify the dimensions, then identify the facts. Most mistakes happen at step 2, where a fact table with a vague grain double counts the moment someone joins it to a second fact.

DataDriven's {COUNT:data_modeling} data modeling problems are answered on a schema canvas. You place tables and their columns, then mark the primary and foreign keys that connect them. The answer is checked for grain and keys, and also for how it handles change, such as a customer who moves city. The dimensional modeling interview questions cover the follow-ups interviewers ask once the diagram is drawn.

StrataScratch's system design questions are concept questions tagged mostly for machine learning and AI roles, and they're answered in text. A data engineering design round asks a different question: how data reliably gets from where it is produced to where it is used. It also expects a diagram.

A typical prompt is to ingest clickstream events from a mobile app and give analysts hourly session metrics. A strong answer names each stage and why it is there. Events land on a queue such as Kafka so producers never wait on consumers. A stream or batch job sessionizes them, and a quality gate checks volume and nulls before anything is published. The results land in a warehouse table partitioned by date. The interviewer then pushes on the edges: events that arrive 2 days late, a backfill after a bug, a consumer that falls behind.

DataDriven's {COUNT:pipeline_architecture} pipeline design problems are answered on a canvas that uses the vocabulary of that round: sources and queues feed transforms and quality gates, which write to storage that consumers read. Practising this round in prose alone tends to produce answers that list tools without saying what each one guarantees.

Both sites now offer an AI mock interviewer, and a mock is the closest thing to the real round because it adds the part silent practice skips. You have to clarify the question and explain your approach, then answer a follow-up you didn't expect.

The difference is which rounds they simulate. StrataScratch's mock interviews are built around the data science and AI roles it serves, machine learning included. DataDriven runs mock interviews for every round of a data engineering loop, the design rounds included, from {COUNT:total} problems, and each one ends with a hire or no hire verdict and the reasons for it, so you can see which part of your answer cost you.

StrataScratch's best-known feature is its company labels: questions marked with the company that asked them, so before an onsite at Meta you can filter for Meta's questions, and the same works for Amazon or Airbnb. Because the library serves data science and analyst roles too, check that a labeled question fits the round you face before you treat it as data engineering preparation.

DataDriven's problems are real data engineering interview questions, tagged with {COUNT:companies} employers and drawn from {COUNT:interviews} interview reports. Use company tags on either site as a guide to a company's style and topics rather than a forecast of your exact question: loops change, and the engineer who can solve unseen problems of the same kind is the one who passes.

StrataScratch is the better fit when the role is closer to analytics than to pipelines: an analytics engineer or product data role whose loop is SQL and metrics plus a case study. It is also the place to practise if your target company interviews in R, or in a SQL dialect such as SQL Server or Oracle that you do not use every day.

Its take-home projects are another reason to keep it. Some data teams send a take-home instead of a live coding round, and StrataScratch's hosted notebooks with real datasets are good rehearsal for that format. And if you already know StrataScratch well, its SQL questions remain useful warm-ups, the same way LeetCode's are. The LeetCode SQL guide for data engineers and the HackerRank SQL guide explain where those 2 sites fit in the same plan.

If you have 6 weeks before a data engineering loop, spend the first 2 on SQL until medium problems take under 20 minutes. Either site works for that; StrataScratch's company filter is useful if you know where you're interviewing, and DataDriven's randomized datasets catch the queries that only work on the sample.

Spend weeks 3 and 4 on the rounds StrataScratch doesn't reach. Do data modeling and pipeline design on DataDriven's canvases, aiming for 2 or 3 designs a week, and add Python problems that parse and deduplicate records or stream them. Use the last 2 weeks for mock interviews of every round, and go back to StrataScratch only for SQL warm-ups on the days before the onsite.