Банк вопросов из реальных собеседований

How do async and await work in Python and how does competitive I/O performance differ from parallel computing?

What does await do in asyncio and why is it important for a non-blocking service?

How to choose the delay between repeated requests and why add random variation to it?

There is a legacy C component and a desire to give users a convenient Python/API layer. How to decide: make a wrapper around C or completely rewrite the implementation in Python?

After BM25 baseline you need to strengthen the search for articles. How to design a candidate generator, hybrid retrieval and reranker?

How can a forecasting system support multiple prediction horizons, and what does it mean that SHAP is model-agnostic?

What should a stream job look like for a CTR dashboard: what does it read, what does it count, and where does it write the result for the advertiser’s chart?

If the CTR panel receives millions of events per second, how do you estimate the number of Kafka partitions and Spark handlers and where to look for the bottleneck?

If the CTR panel receives millions of events per second, how do you estimate the number of Kafka partitions and Spark handlers and where to look for the bottleneck?

The casino product data is stored in storage and comes through a message queue. You need to regularly update the signs and prompt users. How to design a production pipeline?

The casino product data is stored in storage and comes through a message queue. You need to regularly update the signs and prompt users. How to design a production pipeline?

The casino product data is stored in storage and comes through a message queue. You need to regularly update the signs and prompt users. How to design a production pipeline?

In order history, you can use winner label, bid/no-bid, manual removal of supplier and other events. What labels and metrics should I choose for the candidate generator and ranker?

How does scalable scalar attention work, what is its computational complexity, and what changes GQA, MQA, and the limited attention window?

Several annotators transcribe the same low-resource-language audio and produce inconsistent text, numbers and addresses. How do you clean or aggregate the labels?

There are 1000 emails, 100 are truly spam. A model predicts 120 emails as spam, and 90 of those predicted spam emails are truly spam. Compute precision and recall.

How would you choose between SQL and NoSQL storage, and what would you add so the data is not lost?

When are SQL window functions useful, how are they different from GROUP BY, and what ClickHouse MergeTree details matter when writing analytical queries?

How do WHERE and HAVING differ, when should UNION ALL be used, and how does a window function aggregate without collapsing rows?

In Postgres, a complex query with joins and filters is slow. How should I investigate the problem and what remediation options should I consider?

You collected months of human-reviewer decisions for task outputs. How could you use this data to improve the automatic checker?

The model received some value of the average square error in the future period. How do you know if this is a good enough result?

How are matrix equation, least squares, gradient descent and L1/L2 regularization related?

You have predicted boxes and ground-truth boxes for one image. Why is naive average IoU misleading, and how would you match boxes for a useful detection metric?