Interview questions

Lists are mutable (can change) and use []; tuples are immutable and use (). Tuples are hashable (if their items are) so they can be dict keys, and are slightly faster and lighter.

*args collects extra positional arguments into a tuple; **kwargs collects extra keyword arguments into a dict. They make functions flexible and are common in decorators and wrappers.

A decorator is a function that takes a function and returns a new function, adding behaviour without modifying the original — e.g. logging, caching (functools.lru_cache), auth checks (@login_required).

CPython's Global Interpreter Lock allows only one thread to execute Python bytecode at a time. CPU-bound work should use multiprocessing (or native extensions); I/O-bound work benefits from threads or asyncio.

A shallow copy (copy.copy, list[:]) copies the outer container but shares nested objects. A deep copy (copy.deepcopy) recursively copies everything.

WHERE filters rows before grouping; HAVING filters groups after aggregation. WHERE can't use aggregate functions; HAVING can.

An index (usually a B-tree) speeds up lookups and joins. It costs disk space and slows INSERT/UPDATE/DELETE because the index must be maintained. Avoid over-indexing low-selectivity columns.

Organising tables to reduce redundancy: 1NF atomic values, 2NF no partial dependency on a composite key, 3NF no transitive dependencies. Analytical systems often denormalise for speed.

loc selects by labels (index names, column names, boolean arrays); iloc selects by integer positions.

Inspect with isna().sum(), then drop (dropna), fill with a constant/mean/median/mode (fillna), forward-fill for time series, or impute with a model — choosing based on why data is missing.

Bias: error from overly simple assumptions (underfitting). Variance: sensitivity to training data (overfitting). Increasing model complexity lowers bias but raises variance; we aim for the sweet spot via validation, regularisation and more data.

Splitting data into k folds, training on k−1 and validating on the remaining fold, rotating k times. It gives a more reliable performance estimate than a single split.

Precision = TP / (TP + FP): of predicted positives, how many are correct. Recall = TP / (TP + FN): of actual positives, how many we found. F1 is their harmonic mean.

It trains many decision trees on bootstrapped samples with random feature subsets, then averages (regression) or votes (classification). This reduces variance compared to a single tree.

A regularisation technique that randomly zeros a fraction of neurons during training, forcing the network to learn redundant representations and reducing overfitting.

ReLU doesn't saturate for positive inputs, mitigating vanishing gradients, and is cheap to compute — leading to faster training in deep networks.

Ground answers with RAG, instruct the model to cite sources and say 'I don't know', lower temperature, validate outputs against schemas or databases, and evaluate with a test set of known Q&A.

Malicious input that tries to override the system's instructions (e.g. 'ignore previous instructions'). Mitigate by separating untrusted content, limiting tool permissions, validating outputs and never exposing secrets to the model.

Looping over objects and accessing a related field triggers one query per object. Fix with select_related (FK/OneToOne JOIN) or prefetch_related (separate query for M2M/reverse FK).

Auto-escaping templates (XSS), CSRF tokens, parameterised ORM queries (SQL injection), clickjacking headers, secure password hashing, and security settings like HSTS and secure cookies.