Question 1

How would you implement a rolling window operation using stride tricks?

Accepted Answer

as_strided to expose sliding windows, then reduce; guard with correct shape/stride and avoid out-of-bounds. Evaluates practical hands-on knowledge of NumPy in real-world applications. Strong senior candidates explain their reasoning, pros/cons, and connect concepts to projects. Top candidates identify common pitfalls and explain how to mitigate them in production environments.

Question 2

Explain how to avoid double-dipping when stacking or target encoding.

Accepted Answer

Compute target encodings and meta-features strictly out-of-fold; never fit on full data before generating training features. Evaluates practical hands-on knowledge of Scikit-learn in real-world applications. Strong senior candidates explain their reasoning, pros/cons, and connect concepts to projects. Top candidates identify common pitfalls and explain how to mitigate them in production environments.

Question 3

When to use categorical dtype; how does it impact groupby, joins, and memory?

Accepted Answer

Use when low-cardinality strings/labels; reduces memory, speeds groupby/joins via integer codes; careful with category alignment. Evaluates practical hands-on knowledge of Pandas in real-world applications. Strong senior candidates explain their reasoning, pros/cons, and connect concepts to projects. Top candidates identify common pitfalls and explain how to mitigate them in production environments.

Question 4

Describe how to implement a custom gradient with `@tf.custom_gradient` and when it’s needed.

Accepted Answer

Decorate function with @tf; return forward value and custom backward grad. Evaluates practical hands-on knowledge of TensorFlow in real-world applications. Strong senior candidates explain their reasoning, pros/cons, and connect concepts to projects. Top candidates identify common pitfalls and explain how to mitigate them in production environments.

Question 5

How do you cap peak memory (activation/optimizer/state) on big models?

Accepted Answer

Use gradient checkpointing, AMP, smaller optimizer states (8-bit), and offload/freeze parts; watch activation sizes. Evaluates practical hands-on knowledge of PyTorch in real-world applications. Strong senior candidates explain their reasoning, pros/cons, and connect concepts to projects. Top candidates identify common pitfalls and explain how to mitigate them in production environments.

Question 6

How do you design reliable retries for flaky output sinks?

Accepted Answer

Exactly-once sinks or idempotent keys; retries with checkpointed progress. Evaluates practical hands-on knowledge of Spark in real-world applications. Strong senior candidates explain their reasoning, pros/cons, and connect concepts to projects. Top candidates identify common pitfalls and explain how to mitigate them in production environments.

Question 7

Design a CI/CD promoting models from Staging to Production with rollback.

Accepted Answer

GitOps/CI triggers tests → register → stage gate → promote; on failure roll back alias to prior version. Evaluates practical hands-on knowledge of Mlflow in real-world applications. Strong senior candidates explain their reasoning, pros/cons, and connect concepts to projects. Top candidates identify common pitfalls and explain how to mitigate them in production environments.