Hire elite Python developers.
AI & ML experts, ready in 15 days.

Expected answer

A list stores all elements in memory at once, while a generator yields elements one at a time using lazy evaluation. Use lists when you need random access, slicing, or will iterate multiple times. Use generators for large datasets where loading everything into memory is impractical — e.g., processing millions of database rows or streaming API responses. Generators use the yield keyword and consume O(1) memory regardless of dataset size.

Expected answer

I would use async/await throughout to handle I/O-bound operations concurrently. Key optimizations: async database driver (asyncpg for PostgreSQL), connection pooling, Redis caching for hot data, Pydantic v2 for fast serialization, background tasks for non-critical operations, and horizontal scaling behind a load balancer. For the database layer, I would use read replicas and implement query optimization with proper indexing.

Expected answer

The Global Interpreter Lock (GIL) prevents multiple native threads from executing Python bytecodes simultaneously in CPython. This means CPU-bound tasks don't benefit from threading. Workarounds: use multiprocessing for CPU-bound work, asyncio for I/O-bound work, or C extensions that release the GIL. In Python 3.13+, there's experimental support for disabling the GIL entirely (PEP 703).

Expected answer

Use a multi-step approach: 1) Add new columns as nullable (no downtime), 2) Deploy code that writes to both old and new columns, 3) Backfill data with a management command, 4) Deploy code that reads from new columns, 5) Remove old columns. Use django-migration-linter to catch unsafe migrations. Never rename columns directly — always add-new, migrate-data, remove-old.

Expected answer

Architecture: Feature store (Feast) for consistent features → Training pipeline (Airflow + MLflow) for model versioning → Model serving (FastAPI + Redis for caching predictions) → A/B testing framework → Monitoring (data drift detection with Evidently). Key considerations: separate training and inference, version everything (data, features, models), monitor prediction quality in production, and implement graceful degradation (fallback to popularity-based recommendations).