Банк вопросов из реальных собеседований

What is the difference between RDD, DataFrame and Dataset in Spark? Why is DataFrame usually faster, and how do repartition, coalesce, cache, and persist affect Spark tasks?

What should a stream job look like that calculates CTR by campaign_id and time windows?

Why does a custom nn.Module need super().__init__()? Separately, why is tags=[] as a default argument in Python dangerous?

What blocks does a language model agent consist of, where is its state stored, and how does it safely invoke external tools?

Does Python int overflow? How can you roughly estimate how much memory n! needs without computing the factorial?

For users who already have partial subscription history, what transaction features would you build for LTV prediction, and how would you avoid confusing churned users with recently returned users?

Which Candidate Generators Can Be Used in the Recommended System? How does ALS implicit feedback fit into this stack?

What boundary cases should be considered if the LRU cache stores arbitrary user values?

What is a decorator, why is it needed, and why is the code inside the generator executed when iterated rather than when created?

How to write a decorator to log a call, why is functools.wraps needed, and how to lazily import a module inside a wrapper?

What are the tasks in the industrial ML MLflow, reproducible piplins, PySpark and feature storage?

What are the tasks in the industrial ML MLflow, reproducible piplins, PySpark and feature storage?

What are the tasks in the industrial ML MLflow, reproducible piplins, PySpark and feature storage?

What happens inside the Python for-loop, how does an iterator differ from a generator, and why is StopIteration needed?

How are arguments passed to functions in Python? What happens if a function mutates a list argument versus reassigning an immutable value?

How to compare forecast models if LLM-extractor can know future facts from pretraining?

The business wants to know whether the user will return and whether to give him a discount. How to formulate ML-task, target and signs?

In a casino product, the sales team needs to understand as early as possible whether the new player will become a VIP in terms of deposits and turnover. How to formulate a ML-task, target, forecast horizon and business action?

The interview shows a class that reads the file, stores the DataFrame and does the processing. What problems do you look for in this code?

Explain how the Python dict works and how the regular list differs from the NumPy array.

What neural network approaches can be used in RecSys and where do they stand in the pipeline?

The player has just entered the casino product. What signs can be collected in the first days to distinguish a potential VIP from an ordinary player?

You have a linear-regression baseline for LTV. When is gradient boosting likely to help, and how would you decide whether to move to it?

How would you design a backend pipeline that ingests market and alternative data, produces ML signals, applies risk controls, and hands decisions to execution?