What’s changed in data science interviews in 2025?
If you haven’t interviewed since 2022, you might be surprised by today’s data science interview questions. In this article, we’ll dissect 8 major changes in the last couple of years.
Have you tried getting a data science job recently? Many beginners have experienced firsthand how much the market has changed.
Even the more experienced among you — those who were lucky enough to get a job around 2022 and 2023 — who are now looking for better opportunities (or any opportunity after they’ve been laid off) right now, will find that the interviews feel…well, different.
So, what has changed? Here’s an overview of the changes that we’ll cover in this article.
1. Interviews now mirror the real job, not just the role
2022/2023: Interviews were mostly standardised: a basic SQL or Python screen, a take-home task, then a panel interview about exploratory analysis, model accuracy, and some generic business cases. The format didn’t change much between candidates, and it rarely reflected the actual pace or ambiguity of the job.
2025: Every round is tailored to what you’ll do on the job and how you’ll do it. Data science interview questions now test your depth in the specific area (e.g., causal inference for product analysts, orchestration for data engineers), but they also simulate real workplace conditions. You might be asked to debug a broken pipeline live, scope an ambiguous product metric, or pair with an interviewer on a system design task. The goal is not just to test technical correctness, but to see how you think, prioritise, and communicate in messy, collaborative situations.
What it means for you: Read the job description like a spec. Prep for both depth and realism. Practice ambiguous prompts, real-time collaboration, and structuring your thoughts aloud. The format now tests whether you’d thrive in the actual job, not just whether you can pass a technical screen.
2. Real-time systems are becoming more important
2022/2023: Data workflows were mostly batch. While real-time processing existed, it was often seen as niche and too out there for most companies, mostly reserved for Big Tech.
2025: While batch processing is still at the core of most systems, real-time processing is becoming more accessible and mainstream. Even the early-stage startups are now often building real-time systems by default. Some tools, such as Bytewax, Redpanda, Materialize, Snowflake Dynamic Tables, and Delta Live Tables (Databricks), lowered the barrier and made it easier to adopt real-time systems without having to build everything from scratch. Also, the rise of LLMs contributed indirectly by creating demand for systems that can feed embedded LLMs in real-time.
What it means for you: You don’t necessarily need to build streaming systems. However, you need to understand them, as there’s a higher chance you’ll have to work with them. Review how batch vs. real-time systems work and where they fit into modern architectures.
3. Generative AI and productization are now core skills
2022/2023: A few years ago, generative AI and LLMs were still niche. Most data science interviews didn’t include language models, embeddings, or prompt design. As for deploying models, that was usually someone else’s job, typically a machine learning engineer. Data scientists were expected to build models, maybe pickle them for storage, and explain their outputs, but not think much about how they would be used or maintained in production.
2025: Today, generative AI is everywhere: from customer support bots to content engines to internal search tools. Interviews now test whether you can turn models into usable, reliable systems. You’ll be expected to build APIs around models, design retrieval-augmented generation (RAG) pipelines that combine LLMs with vector search, write and refine prompts to reduce hallucination and improve output quality. You’ll also compare different embedding strategies and manage latency, cost, and token usage in production environments.
Especially in smaller or LLM-heavy teams, you may even be asked to prototype a full flow using FastAPI, Streamlit, MLflow, SageMaker, or LangChain. It’s not enough for your model to be accurate — it has to fit into a real product that delivers value.
What it means for you: Think beyond model accuracy. You need to design with users, systems, and scalability in mind. Practice building simple APIs, structuring prompts for different use cases, and balancing trade-offs like latency versus cost or accuracy versus interpretability. Familiarise yourself with the tools I mentioned above.
4. The bar for SQL is way higher
2022/2023: If you could write some joins, basic window functions, and perform data filtering and aggregation, you were generally fine. With classic Leetcode-style questions like “Find the second highest salary” or “Count consecutive login streaks”, the interviewers were testing whether you can code correctly, rather than focusing on business problems.
2025: Today, interview questions tend to be ambiguous and business-oriented. For example, in a question like “Who are our most valuable users over time?”, you get to define what “valuable” means. You must define metrics, make trade-offs, explain why your logic makes sense, and write the SQL code that will correctly output the users according to your logic. Additionally, your code should be optimized and ready for production.
What it means for you: Practice real-world cases and learn metrics for a particular industry you’re interested in. When coding, strive for clean, optimized, and maintainable SQL code.
5. Python code is expected to be modular and production-ready
2022/2023: Writing a readable inline code, typically using pandas and/or NumPy, in an exploratory style was often enough. Reusability and structure weren’t priorities.
2025: Now, the interviewers want code that can be plugged into MLflow, FastAPI, or a CLI script. In other words, your Python code should be modular, testable, and structured in functions or classes. Why? They want data scientists who can build things that integrate, scale, and can be handed off.
What it means for you: Practice writing Python code in functions and classes, separate the logic clearly, and perform unit tests.
6. More focus on product thinking
2022/2023: Two to three years ago, the interviewers would be happy if you could demonstrate your business acumen by explaining metrics or describing A/B test results.
2025: Now, you’re expected to be more of a product owner with a data science background. That means you should identify key performance indicators (KPIs) and prioritize features and experiments that will impact those KPIs. Some interviews will even simulate a product meeting scenario where you’re the only one from a data team.
What it means for you: Learn to translate vague business goals into metrics and features/experiments you would prioritise. Read product case studies and practice talking through trade-offs in plain language.
7. Quality over quantity in project portfolios
2022/2023: Previously, employers saw portfolios with numerous projects as proof of your knowledge, as well as your drive and willingness to learn. They valued quantity over quality.
2025: Employers realised that everybody was doing the same capstone projects (house price prediction, Titanic survival, MNIST digit, anyone?) in their GitHub notebooks and showing the same static dashboards. Now, they’re asking for fewer, but higher-quality projects, preferably relevant to the role you’re interviewing for. In this case, higher quality means end-to-end projects with hosted apps, documented pipelines, and real-data problem-solving.
What it means for you: Focus on fewer but deeper projects. Use real data, include a clear README, link to a live demo if possible, and explain your assumptions, trade-offs, and reasoning behind the design choice.
8. You can’t ignore MLOps anymore
2022/2023: You don’t know what MLOps is? It doesn’t matter. It’s 2022, after all. You are not an ML engineer, but a data scientist working in silos without worrying about deployment or monitoring.
2025: Today, you must know that Machine Learning Operations (MLOps) are tools and processes used to manage ML models from development to production and monitoring. Even if you’re not the one deploying models, you’re expected to understand how it happens. That means tracking version changes, scheduling retrainings, monitoring for data drift, and debugging model failures. Tools like MLflow, Weights & Biases, and feature stores (Feast, Tecton) are now standard parts of the interview vocabulary.
What it means for you: Understand MLOps stages and principles. Be able to explain how you’d detect your model silently failed (i.e., the model still runs, but no one notices the results are wrong or outdated), and how you’d fix it.
Conclusion
The actual job’s core tasks may be similar, but expectations around collaboration, specialization, and real-world performance have risen dramatically.
As for the technical changes, the most significant ones arose due to the emergence of LLMs and AI, as well as live data streaming becoming more mainstream.
In 2025, it’s not enough to know what to do — you need to know how, when, and why. And interviews reflect that.
