Банк вопросов из реальных собеседований

There are historical transactions, platform logs, more than a million suppliers and about 100 customer companies. What data to use and how does scale affect the architecture?

What facts from company PDF reports are useful for production forecasts, and how can you distinguish them from noisy text?

After launching the MVP, what events and features need to be collected to train the user-video ranking model?

What are the typical problems of recommendation systems and how can they be measured or reduced?

What features of the catalog need recommendations in the basket and what quality problems arise in the commodity taxonomy?

Why reduce the accuracy of language model weights and activations, what quantization options exist, and what losses should be measured?

What does the language model’s KV cache store, why does it speed up generation, and what limitations does it create when serving multiple queries?

How to choose between improving the instruction, searching for knowledge, and training the model, including the LoRA adapters?

When would you choose a columnar database over Redis, MongoDB or a row-oriented relational database for ML/data pipelines?

When should you use a classic batch ETL and when is streaming for recommendations, analytics or ML features?

Search has embeddings and a full-text index. When to use both approaches and how to combine them?

The product has a search for documents/model files. When to use full-text, when to use vector search, and why might you need hybrid retrieval?

How to design caching and latency budget for recommendation API?

How to use location and image quality in a price or recommendation model without mixing product condition with photo quality?

With 10,000 calls per day and many quick rejections, how would you structure the pipeline and choose metrics before running an expensive extractor?

How to evaluate the quality of search or RAG-system offline and online?

Какие offline, online и guardrail-метрики выбрать для A/B-теста динамической стоимости доставки?

We design ML for search on the marketplace. What business, online and offline metrics to choose?

What metrics should you use in a marketplace where clicks, contacts, deals, and distribution of impressions between sellers reflect different goals?

What offline, online, and security metrics fit the recommendation box in the shopping cart if clicking doesn’t already mean buying?

A production service already has data, but you need to change the database schema. Describe a safe migration process.

Design a basic multi-head attention block in PyTorch. What tensors do Q, K and V have, when do you split heads, and what common shape mistakes should you avoid?

Explain the difference between a Kubernetes pod, service, deployment and node.

Explain the difference between a Kubernetes pod, service, deployment and node.