Interview Question Arena
Practice questions for Product Management, Data Science, and Data Engineering interviews. 306 questions across 29 topics.
Product Management
Product Sense
- Jobs-to-be-done — How do you use the Jobs-to-be-done framework to identify what to build next?
- Problem framing — Walk me through how you would frame a product problem before jumping to solutions.
- User segments — How do you segment users to decide which group to focus on first?
- Accessibility — How do you ensure a product feature is accessible to users with disabilities?
- Trade-offs — Describe a product decision where you had to make a significant trade-off. How did you approach it?
- Hypothesis-driven product development — How do you decide what to build next, and how do you know when you're wrong?
- Solution ideation — How do you generate and narrow down solution ideas for a product problem?
- Launch criteria — How do you decide whether a feature is ready to launch?
- Edge cases — How do you identify and handle edge cases when designing a feature?
- Zero marginal cost products — When would you choose a local-first or zero-server-cost architecture, and how does that decision shape your business model?
Product Design
Metrics & Analytics
- North Star Metric — How would you define a North Star Metric for [product]? What makes a good NSM?
- AARRR funnel — Walk me through the AARRR framework and how you use it to diagnose a product problem.
- Guardrail metrics — What are guardrail metrics and how do you use them alongside your North Star?
- A/B test design — How would you design an A/B test for a new checkout flow?
- Counter metrics — You shipped a feature and your primary metric went up. What counter metrics would you check?
- Leading vs lagging indicators — Explain the difference between leading and lagging indicators with a product example.
- Statistical significance — Your A/B test shows a 3% lift with p=0.04. Do you ship it? Walk me through your reasoning.
- Experiment design — When would you not use a standard A/B test and what alternatives exist?
- Metric debugging — DAU dropped 15% overnight on Facebook Events. Walk me through your investigation.
- Competing metrics ship decision — You redesigned Facebook Watch. Watch time is up 4%, but likes and comments are down 8%. Do you ship?
- Retention deep dive — Retention for your product is flat at 40% D30. What would you investigate and what would you try?
Execution
- Prioritisation frameworks — Walk me through how you prioritise your roadmap when you have more to build than capacity allows.
- Success metrics — How do you define success metrics for a new product feature before you build it?
- OKRs — How do you write good OKRs and how do you use them day-to-day?
- Sprint planning — How do you run an effective sprint planning session with your engineering team?
- Product roadmap — How do you build and communicate a product roadmap that different stakeholders can align on?
- Stakeholder management — Tell me about a time you had conflicting priorities from different stakeholders. How did you handle it?
- Go-to-market — Walk me through how you would launch a new feature to users.
- Technical debt — How do you decide when to ship fast vs invest in reducing technical debt?
Estimation
- Market sizing — Estimate the market size for a new Google feature that helps users manage their home energy usage.
- Fermi estimation — How many Uber rides happen in London on a Friday night?
- Back-of-envelope math — A new feature increases session length by 2 minutes. How much additional ad revenue does that generate for Meta per year?
- User base sizing — Estimate the number of active Gmail users in the UK.
- Revenue modelling — How would you model the revenue impact of a new subscription tier?
- Growth rate modelling — A product has 1M users growing at 15% MoM. When will it reach 10M users, and what assumptions would change that?
Strategy
- Competitive response — A well-funded competitor just launched a free version of your core product. How do you respond?
- Market entry — Should Google launch a competitor to Shopify? Make the case for or against.
- Platform and ecosystem — How do you decide whether to build a feature in-house or open it up as a platform for third-party developers?
- Monetisation strategy — You are the PM for a free product with 50M MAU. Design a monetisation strategy that does not kill growth.
Behavioural
- The screening round — How do you prepare for and perform in a FAANG PM screening call?
- Leadership examples — Tell me about a time you led a cross-functional team through a difficult project.
- Cross-functional collaboration — How do you keep engineering, design, marketing, and legal aligned on a complex launch?
- Influence without authority — How do you get engineers or designers to prioritise your work when they don't report to you?
- Data-driven decisions — Tell me about a decision where the data said one thing but your instinct said another. What did you do?
- Failure stories — Tell me about a product decision you made that turned out to be wrong. What did you learn?
- Conflict resolution — Tell me about a time you had a serious disagreement with an engineer or designer. How did you resolve it?
- Navigating organisational politics — Tell me about a time you had to navigate a difficult organisational dynamic — a reorg, a scope conflict, or a situation where doing good work wasn't enough.
- Getting promoted past senior — What does it take to get promoted past senior PM or staff level, and how is that different from getting your first promotion?
- Amazon Leadership Principles — Amazon's Bar Raiser will ask Leadership Principles questions in every round, not just the behavioural round. How do you prepare?
- Ambiguity and ownership — Tell me about a time you had to make a significant decision with incomplete information. How did you decide, and what happened?
- Why PM — Why do you want to be a product manager? What makes you good at it?
AI/ML for PMs
- How LLMs work — Explain how a large language model works to a non-technical stakeholder. What are its key limitations?
- LLM limitations — What are the main failure modes of LLMs and how do they affect product design?
- Prompt engineering — How do you design prompts to make an LLM-powered feature reliable for end users?
- RAG — What is Retrieval-Augmented Generation and when would you choose it over fine-tuning?
- Model evaluation — How do you evaluate whether an AI feature is performing well enough to ship?
- AI product design — Design an AI-powered feature for [product]. How do you handle cases where the AI is wrong?
- AI agent proficiency — What does it mean for a PM to be proficient with AI agents, and how does this change how product teams work?
- Activation in AI products — What makes activation uniquely difficult for AI products, and what approaches have actually worked?
- AIPM behavioral interview framework — What are the four categories of behavioral questions in AI PM interviews, and what does a strong answer in each category look like?
- ML evaluation: offline to business impact — Walk me through how you evaluate whether an ML feature is performing well, from model quality through to business impact.
- Fine-tuning — When would you recommend fine-tuning a model vs using a foundation model with prompting?
- Responsible AI — How do you ensure an AI feature is safe, fair, and doesn't cause harm?
- Latency vs cost trade-offs — Your AI feature uses GPT-4 but latency is 4 seconds and cost is $0.05/query. What do you do?
- AI safety trade-offs — How do you balance making an AI assistant helpful vs preventing misuse?
- The PM role in the AI era — How do you see the PM role evolving as AI coding agents become widespread? What compounds, and what becomes obsolete?
- AI-native team operations — How are leading AI companies restructuring their product teams and processes to operate at AI speed?
- LLM inference cost optimisation — Your LLM costs $0.40 per query. At 100,000 queries/day that is $40,000/day. Logs show 60,000 of those queries are slight variations of the same 200 questions. How do you cut inference cost by 60% without users ever feeling like they got a cached or stale response?
- Navigating AI ethics conflicts — You discover mid-development that your ML feature has a bias or fairness problem that could violate regulations. How do you handle it as PM?
- Agent error compounding — Your support agent runs a five-step workflow: read the ticket, search the knowledge base, draft a reply, check it against policy, send. Each step is 95% accurate. What is the end-to-end success rate, and what do you change?
- Agent permissions and blast radius — Your agent can browse the web, call internal APIs, and write files. Scope its permissions. What do you refuse to let it do, and how do you enforce that rather than just prompting for it?
- Irreversible agent actions — Your agent took an action the user cannot undo — it emailed the wrong customer list. Design the product response. Not the apology: the system that means it does not happen the same way twice.
- Human-in-the-loop checkpoint placement — Where do you put human-in-the-loop checkpoints in an agent workflow, and what does putting them in the wrong place cost you?
- Building an eval set from production — You inherit an AI feature that is already live and has no evals. Build the eval set. Where do the cases come from, and how do you stop it going stale?
- EU AI Act deadline shift — Your roadmap was built around the EU AI Act high-risk obligations landing on 2 August 2026. The Digital Omnibus package then pushed stand-alone Annex III high-risk systems to 2 December 2027. What changes, and what does not?
- Classifying a feature as high-risk — Walk me through classifying whether your feature is high-risk under the EU AI Act. Use a product that screens, scores or ranks people as your example.
- Model deprecation and vendor dependency — Your provider announces that the model your product is built on will be retired in 60 days. Walk me through those 60 days.
- AI versus deterministic shipping rule — Write the rule your team uses to decide whether a feature ships with an LLM or with deterministic logic. It has to be specific enough that an engineer can apply it on a Tuesday without asking you.
Career & Mastery
- Consistency — How do you maintain consistent progress on long-term skill development, especially when motivation is low?
- Discipline — How do you sustain high-quality effort on hard work when motivation has run out?
- Motivation — How do you sustain motivation on a long-term project that isn't showing visible results yet?
- Skill trees — How do you approach learning a domain that feels completely overwhelming or out of reach?
- Deliberate practice — What does it mean to practice deliberately, and how is it different from just putting in hours?
- Learning mechanics — What's the most effective way to ensure something you study actually sticks long-term?
- Expertise — How do experts think differently from beginners, and what does that mean for how you build domain expertise?
- The long game — How do you stay oriented on long-term skill development when short-term results aren't visible?
- Earning mentors — How do you build relationships with senior people who can accelerate your career?
- Mission selection — How do you identify the right problem to work on, and how do you know if your direction is viable?
- Compounding principles — What separates someone who compounds over a decade from someone who stays in the same place?
Data Science
Statistics
- Probability distributions — Explain the difference between normal, Poisson, and binomial distributions. When would you use each?
- Hypothesis testing — Walk me through the steps of a hypothesis test. What is a p-value actually telling you?
- p-values — What are common misinterpretations of p-values and how would you explain them to a non-technical PM?
- Confidence intervals — What is a 95% confidence interval? How do you interpret it and what mistakes do people make?
- Bayes theorem — Explain Bayes theorem and give a product analytics example of where it applies.
- Central Limit Theorem — What is the Central Limit Theorem and why is it so important for A/B testing?
- Correlation vs causation — A PM says "users who use feature X retain 20% better — we should push everyone to feature X." What's wrong with this analysis?
- A/B test power — How do you calculate the required sample size for an A/B test?
- Covariance vs correlation — Explain the difference between covariance and correlation. When does covariance mislead you, and when is Pearson correlation insufficient?
- Stratified sampling — When would you use stratified sampling instead of simple random sampling, and how does it affect variance and statistical power?
- Bootstrapping — What is bootstrapping and when would you prefer it over a parametric statistical test? Walk through computing a 95% confidence interval via bootstrap.
- Outlier detection — Compare at least three techniques for detecting outliers in a dataset. How do you decide which to apply, and how do you handle detected outliers?
Machine Learning
- Bias-variance trade-off — Explain the bias-variance trade-off and how it guides your model selection decisions.
- Overfitting and regularisation — How do you detect and address overfitting in a production ML model?
- Decision trees — How does a decision tree work and what are its main weaknesses?
- Ensemble methods — Compare bagging, boosting, and stacking. When would you choose gradient boosting over a random forest?
- Gradient descent — Explain gradient descent and its variants. What is the learning rate and how do you tune it?
- Feature engineering — What feature engineering techniques do you apply when building a click-through rate prediction model?
- Model evaluation metrics — You're building a fraud detection model. Why is accuracy a bad metric and what would you use instead?
- Cross-validation — When would you use k-fold cross-validation vs a simple train/test split?
- Hyperparameter tuning — What are the main approaches to hyperparameter tuning and how do you choose between them?
- Feature importance — How do you explain which features are driving your model's predictions to a non-technical stakeholder?
- Missing data handling — A dataset you are training on has 15% missing values in a key feature. Walk through your decision process for how to handle it.
SQL & Data Engineering
- Window functions — Write a SQL query to find the top 3 users by revenue for each country in the last 30 days.
- CTEs — When do you use a CTE vs a subquery vs a temporary table? Write a query using a CTE.
- Query optimisation — Your SQL query on a 10B row table is taking 2 minutes. How do you debug and optimise it?
- Data modelling — When would you use a star schema vs a normalised data model in a data warehouse?
- ETL pipelines — How do you design a reliable ETL pipeline that handles late-arriving data and failures?
- Joins — Explain the difference between INNER, LEFT, RIGHT, and FULL OUTER joins with an example.
- Scaling data systems — Walk me through how you would identify bottlenecks in a data system and what approaches you would use to scale it to 10× the current load.
Product Analytics
- Funnel analysis — Walk me through how you would analyse a checkout funnel where conversion dropped 15% last week.
- Cohort analysis — You want to understand whether a product change improved long-term retention. How do you use cohort analysis?
- Retention analysis — How do you measure and improve retention for a consumer app?
- KPI trees — Build a KPI tree for an e-commerce platform's revenue metric.
- North Star Metric — How do you choose a North Star Metric for a two-sided marketplace like Airbnb?
- Dashboarding — What makes a good analytics dashboard and what are common mistakes?
- Metric query patterns — Write a SQL query to compute 7-day retention for a mobile app, given a table of user_id and event_date.
Experimentation
- A/B test design — Walk me through how you would design an A/B test for a new recommendation algorithm.
- Type I and Type II errors — Explain Type I and Type II errors in the context of A/B testing. Which is worse?
- Sample size calculation — How do you calculate the minimum sample size needed for an A/B test?
- CUPED — What is CUPED and why does it reduce the sample size needed for A/B tests?
- Novelty effects — Your experiment shows a 5% lift in engagement in week 1 but it fades by week 4. What's happening and what do you do?
- Sequential testing — What is sequential testing and when is it preferable to fixed-horizon A/B tests?
- Multiple testing — You're running 20 A/B tests simultaneously. How do you avoid false discoveries?
- Power analysis — Your team wants to detect a 1% improvement in conversion. What are the implications for your experiment?
- Network effects and interference — You are running an A/B test on a new driver incentive at Uber. Why does standard user-level randomisation fail, and what do you do instead?
- Causal inference methods — A feature cannot be A/B tested because it was rolled out to all users at once. How do you measure its causal impact?
- Ad-load probabilistic experiment design — You want to test whether increasing ad load from 3 to 4 ads per session harms retention. The effect may be small and heterogeneous across user segments. How do you design the experiment and analyse the results?
LLMs & GenAI
- Transformer architecture — Explain the transformer architecture. What problem did it solve that RNNs couldn't?
- Tokenisation — How does tokenisation work and why does it matter for LLM applications?
- RAG — Design a RAG system for a customer support chatbot. What are the key components and failure modes?
- Fine-tuning — When would you fine-tune a model vs use RAG vs just prompt-engineer a foundation model?
- RLHF — Explain RLHF at a high level. What problem does it solve and what are its limitations?
- Prompt engineering — What techniques do you use to improve LLM output quality through prompting alone?
- Vector databases — How do vector databases work and what are the trade-offs between exact and approximate nearest neighbour search?
- Embeddings — What are embeddings and how are they used in LLM-based applications?
- LLM evaluation — How do you evaluate whether an LLM-powered feature is performing well in production?
- AI agents — What is an AI agent and what are the key design decisions when building one?
- Sampling parameters: temperature, top-k, top-p — Explain temperature, top-k, and top-p (nucleus) sampling. How do you tune them for a customer-facing vs a code-generation use case?
- Context window overflow solutions — A user's conversation history plus retrieved documents exceeds the model's context window. What strategies do you use to handle this without losing critical information?
- RAG chunking strategies — What chunking strategy would you use for a RAG system over a large technical documentation corpus? Walk through the trade-offs of fixed-size, semantic, and hierarchical chunking.
- RAG failure modes — A RAG system is giving wrong answers even though the relevant document is in the corpus. Walk through the failure modes and how you diagnose each one.
- Hybrid search: BM25 + dense retrieval — What is hybrid search in a RAG system and why does combining BM25 with dense vector search outperform either approach alone?
- Re-ranker in RAG — What role does a re-ranker play in a RAG pipeline, and when is the added latency worth it?
- HNSW algorithm — Explain how HNSW (Hierarchical Navigable Small World) enables fast approximate nearest-neighbour search in a vector database. What are the key parameters?
- Scaling to 150M embeddings — You need to serve approximate nearest-neighbour search over 150 million embedding vectors with p99 latency < 50ms. Walk through the architecture choices.
- Vector search recall debugging — Your RAG system's retrieval recall has dropped from 85% to 60% after a model migration. How do you diagnose and fix it?
- Structured output and JSON prompting — How do you reliably get an LLM to return structured JSON output for use in a downstream system? What fails and how do you prevent it?
- Prompt injection mitigation — What is prompt injection and how do you defend against it in a production LLM application that processes untrusted user input?
- Reducing hallucinations via prompting — Without changing the model, what prompt engineering and system design techniques reduce hallucination rates in a production LLM application?
- RAGAS evaluation framework — What is RAGAS and what metrics does it provide for evaluating a RAG pipeline end-to-end?
- LLM-as-judge — Explain the LLM-as-judge evaluation paradigm. What are its strengths over human evaluation, and what biases make it unreliable?
- Hallucination detection — Beyond temperature tuning and grounding prompts, how do you detect and measure hallucinations in a production LLM system at scale?
- LLM eval flakiness — Your LLM evaluation suite gives different pass/fail results on the same code change across runs. How do you diagnose and fix eval flakiness?
- Multi-tenant RAG isolation — You're building a RAG product used by 500 enterprise customers, each with private documents. How do you ensure strict data isolation between tenants in the vector store?
- MLflow for reproducibility — How do you use MLflow to ensure an LLM or ML experiment is fully reproducible six months later?
- Trace logging for LLM explainability — How do you implement observability and trace logging for an LLM application so you can debug why a specific user got a wrong answer?
Coding
- Pandas data cleaning — Write a function clean_dataframe(df) that handles: (1) drops exact duplicate rows, (2) fills numeric nulls with column median, (3) fills string nulls with "unknown", (4) strips leading/trailing whitespace from all string columns. Return the cleaned DataFrame.
- Train/test split from scratch — Implement train_test_split(X, y, test_size=0.2, random_seed=42) without using sklearn. Shuffle the data, split it, and return (X_train, X_test, y_train, y_test) as numpy arrays.
- Precision, recall & F1 from scratch — Implement precision(y_true, y_pred), recall(y_true, y_pred), and f1_score(y_true, y_pred) from scratch using only Python built-ins. Handle the zero-division edge case gracefully.
- Rolling window features — Given a DataFrame with columns ["user_id", "date", "value"] sorted by user and date, write code to add per-user rolling 7-day mean, rolling 7-day std, and a 1-day lag of "value". Do not leak future data into past rows.
- K-means from scratch — Implement kmeans(X, k, max_iters=100, random_seed=42) from scratch using only numpy. Return (labels, centroids). Handle the edge case where a centroid gets no assigned points.
Behavioural
- Disagreeing with technical direction — Tell me about a time you disagreed with your manager or senior engineer on a technical decision. How did you handle it and what was the outcome?
- Deciding with incomplete information — Describe a situation where you had to make an important technical or analytical decision with incomplete data. How did you decide and what framework did you use?
- Production data incident response — Walk me through a time a data pipeline or model failure caused a business impact. How did you detect it, respond, and prevent recurrence?
- Discovering a data quality issue post-deployment — You deployed a new pipeline and a week later discover that a metric used by executives has been silently wrong for 7 days. How do you handle it?
- Mission and impact motivation — Why do you want to work on AI / data infrastructure at this company specifically? What draws you to this problem space?
- Influencing without authority — Describe a time you convinced a team or stakeholder to adopt a technical approach you believed in, when you had no formal authority over them.
Data Engineering
Coding
- SQL window function query — You have a table orders(order_id, user_id, order_date, revenue). Write a single SQL query that returns: user_id, order_date, revenue, a 7-day rolling sum of revenue per user, and their revenue rank on each day across all users.
- Upsert with MERGE — Write a SQL MERGE statement that upserts from staging.users into production.users. Match on user_id. If matched and email or name changed: update those fields and set updated_at = NOW(). If not matched: insert the full row.
- PySpark transformation — Write a PySpark job that: (1) reads a JSON file from S3, (2) filters out rows where "status" is null, (3) adds a "processed_date" column with today's date, (4) writes the output as Parquet partitioned by "country". Use the DataFrame API.
- Incremental ETL in Python — Write a Python function run_incremental_load(source_conn, target_conn, watermark_table) that: reads the last watermark from a state table, queries the source for rows updated_at > watermark, upserts them into the target table, then updates the watermark. Handle failures safely.
- SQL data quality checks — Write SQL to run data quality checks on a table users(user_id, email, created_at, country, age). Return a single summary result set with one row per check: check_name, status (PASS/FAIL), and failing_count.
- Clustered vs non-clustered indexes — Explain the difference between clustered and non-clustered indexes. When would adding an index hurt write performance more than it helps read performance?
- N+1 query problem — What is the N+1 query problem? Give a concrete example and explain how to fix it at the application and database levels.
Pipeline Engineering
- Airflow & DAG design — How would you design an Airflow DAG for a daily ingestion pipeline that pulls from 5 sources, transforms the data, and loads it into a data warehouse — with proper error handling and no duplicate records?
- dbt fundamentals — Explain the role of dbt in a modern data stack. How does it differ from a traditional ETL tool, and what problems does it solve?
- Idempotency — What does it mean for a data pipeline to be idempotent, and why is it critical? Give a concrete example of an idempotent and a non-idempotent load pattern.
- Incremental loading — When would you choose incremental loading over a full refresh? What patterns exist and what are the trade-offs?
- Pipeline debugging — Your pipeline ran successfully but the downstream dashboard shows a 30% drop in row counts. Walk me through your debugging process.
- Change Data Capture — What is Change Data Capture (CDC) and when would you use it instead of batch ingestion? What are the implementation trade-offs?
- dbt incremental strategies — What is an incremental model in dbt? Walk through the available strategies and how you handle late-arriving data.
- CI/CD for dbt — How would you set up a CI/CD pipeline for a dbt project? What checks run on each pull request and how do you avoid running the entire model graph on every PR?
- Spark architecture — Explain Spark's execution model. What are the roles of the driver, executors, and cluster manager, and how does Spark achieve parallelism?
- Data skew in Spark — What is data skew in Spark and why does it cause jobs to stall? Walk through three techniques to mitigate it.
- Medallion Architecture — What is the Medallion Architecture (Bronze/Silver/Gold)? How does it structure a data lakehouse, and what are the trade-offs versus a single-layer approach?
- Surrogate vs natural keys — When would you use a surrogate key versus a natural key as the primary key in a data warehouse dimension table? What are the practical trade-offs?
- Airflow retry and failure handling — A dbt model running in Airflow fails halfway through because of a transient API timeout. How do you design the DAG and the dbt model so the retry is safe and idempotent?
Streaming Systems
- Kafka fundamentals — Explain Kafka's architecture. What are topics, partitions, and consumer groups, and how do they work together to enable scalable, fault-tolerant message processing?
- Exactly-once semantics — What are the three delivery guarantees in distributed messaging systems, and what does exactly-once processing actually require to implement?
- Streaming vs batch — You're building a fraud detection system. When would you use streaming over batch processing, and what are the engineering trade-offs?
- Late-arriving data — How do you handle late-arriving events in a streaming pipeline? Explain watermarks and the trade-off between completeness and latency.
- Kafka partitioning strategy — How do you decide how many partitions a Kafka topic should have, and how do you choose a partition key? What goes wrong with a bad choice?
- Backpressure in streaming — What is backpressure in a streaming system and how do Flink and Kafka Streams handle it? What happens if you ignore it?
- Flink vs Spark Streaming — Compare Apache Flink and Spark Structured Streaming. When would you choose Flink over Spark for a production streaming job?
- Lambda vs Kappa architecture — Explain the Lambda and Kappa architectures. What problem does each solve, and what are the operational trade-offs?
Cloud Data Warehouses
- Warehouse architecture comparison — Compare BigQuery, Snowflake, and Redshift on architecture, pricing model, and use case. When would you choose each?
- Partitioning & clustering — How does partitioning differ from clustering in a data warehouse like BigQuery, and how do you decide which (or both) to apply to a table?
- Cost optimisation — Your BigQuery bill doubled last month. Walk me through how you'd investigate the cause and reduce costs.
- Parquet over CSV — Why would you choose Parquet over CSV for storing large analytical datasets? Explain the architectural differences and when CSV is still acceptable.
- Delta Lake vs Iceberg vs Hudi — What are Delta Lake, Apache Iceberg, and Apache Hudi? What problem does each solve, and when would you choose one over the others?
- Avro vs Parquet — When would you use Avro versus Parquet? Describe the architectural difference and give a concrete example for each.
- Snowflake architecture — Explain Snowflake's three-layer architecture. How does separating storage and compute enable features that traditional MPP warehouses cannot offer?
- Z-ordering — What is Z-ordering (space-filling curve ordering) in Delta Lake and Databricks, and when does it improve query performance over simple column sorting?
- Materialized views — What are materialized views, when do they outperform regular views, and what are the consistency trade-offs to watch for?
- Snowflake Time Travel and cloning — Explain Snowflake's Time Travel and zero-copy cloning features. How would you use each in a production data engineering workflow?
- Snowflake micro-partitions — What are Snowflake micro-partitions and how does Snowflake use them for automatic query pruning? Why does this change how you think about clustering?
Data Quality & Observability
- Data quality dimensions — What are the key dimensions of data quality, and how would you build an observability framework to monitor them across your pipelines?
- Schema evolution — A source system adds a new column and silently removes an old one. How do you handle schema evolution without breaking downstream pipelines?
- Data contracts — What is a data contract and why is it becoming a standard practice in data engineering teams?
- Testing data pipelines — How do you test a data pipeline end-to-end? What types of tests exist and where do they sit in your CI/CD process?
- Data quality framework at scale — How would you design a data quality framework for a platform with 500+ tables and dozens of pipelines? What do you automate vs review manually?
- SLAs and SLOs for data products — How do you define and enforce SLAs and SLOs for data pipelines? Give an example of a freshness SLO and explain what happens when it is breached.
- Anomaly detection in pipelines — How do you automatically detect anomalies in data pipelines — such as unexpected row count drops or distribution shifts — without requiring manual threshold configuration for every table?
- Schema change triage — You get paged at 2am: a source system silently removed a column your pipeline depends on. Walk through how you triage and resolve this without data loss.
Data Governance
- PII handling — You're building a pipeline that processes user data including names, emails, and IP addresses. What steps do you take to handle PII correctly end-to-end?
- GDPR compliance — What are the main data engineering implications of GDPR? How do you build pipelines that stay compliant as your data volume grows?
- Data access controls — How do you implement fine-grained access controls in a data warehouse? Explain the difference between row-level and column-level security.
- Data lineage — What is data lineage, why does it matter for engineering teams, and how do you implement it in a modern data stack?
ML Infrastructure
- Feature stores — What is a feature store, what problem does it solve, and when would a data engineering team invest in building or adopting one?
- Model monitoring — Once a model is deployed to production, what data infrastructure do you need to monitor it effectively and detect when it needs retraining?
- Training data pipelines — What are the unique challenges of building pipelines for ML training data, compared to analytics pipelines?
- Embedding pipeline for fine-tuning — Design the data pipeline required to prepare a fine-tuning dataset for a domain-specific LLM. What are the key steps and quality gates?
- Eval data pipeline — How do you build and maintain a reliable evaluation dataset pipeline for a production ML or LLM system?
Analytics Engineering
- Analytics engineering layer — What is the analytics engineering layer? How does it sit between raw data engineering and business analysis, and what does an analytics engineer own?
- Metrics layer — What problem does a semantic/metrics layer solve, and when would you introduce one into your data platform?
- Self-service analytics — How would you design a data platform that enables non-technical stakeholders to explore data independently without needing an analyst for every question?
System Design
- Real-time log aggregation system — Design a real-time log aggregation system that ingests 1 million log events per second from 10,000 microservices, stores them durably, and makes them searchable within 5 seconds.
- IoT streaming platform — Design a data platform to ingest telemetry from 10 million IoT devices sending sensor readings every 30 seconds. The platform must support real-time alerting (< 10 seconds) and historical analytics.
- 10TB daily clickstream pipeline — Design an end-to-end pipeline to process 10TB of raw clickstream events per day, transform them into session-level and user-level aggregates, and serve them to a BI tool by 08:00 UTC each morning.
- MAU computation pipeline — How would you design a pipeline to compute and serve Monthly Active Users (MAU) in near-real-time for a product with 500 million daily events, without scanning the full month's data on every query?
- Top-K trending items pipeline — Design a real-time system to surface the top-K trending items (e.g. trending hashtags, search queries) over a sliding 1-hour window across 100,000 events per second.
- Point-of-sale data pipeline — Design a data pipeline for a retail chain with 5,000 stores, each sending POS transaction data every 15 minutes. The system must handle intermittent connectivity and produce daily sales reports accurate to the penny.
- Enterprise Data Lake design — A 10,000-person enterprise wants to consolidate 50 data sources into a single data lake. Design the architecture, governance model, and access control strategy.
AI
ML & GenAI Fundamentals
Google / Gemini
Meta / Meta AI
- Design AI assistant for Meta platforms — Design an AI assistant for WhatsApp, Instagram, or creators.
- AI to improve Reels creation or discovery — How would you use AI to improve Reels creation or discovery?
- Metrics for Meta AI — What metrics define success for Meta AI?
- Engagement up but trust down — Engagement is up but user trust is down. What do you do?
- Detect and reduce harmful AI-generated content — How would you detect and reduce harmful AI-generated content?
- Open-source powerful model tradeoffs — Should Meta open-source a powerful new model? Walk through the tradeoffs.
- AI agent for small businesses on Instagram — Design an AI agent for small businesses on Instagram.
- Debug retention drop after AI feature launch — How would you debug a drop in retention after launching an AI feature?
- Drove alignment across design, ML, data, policy — Tell me about a time you drove alignment across design, ML, data, and policy.
- Prioritize latency, cost, quality, safety — How would you prioritize among latency, cost, quality, and safety?
OpenAI
Anthropic