Production checklist for NumPy

Production pipelines using NumPy need dtype discipline, shape validation, reproducible RNG, and explicit copies at boundaries—especially before Pandas handoff or model serving.

Before shipping numeric code

• Assert expected shape and dtype at API boundaries
• Document axis conventions (samples × features)
• Seed RNG (default_rng) in tests and training
• Use np.save / versioned artifacts for array checkpoints
• Handle NaN/inf explicitly—don't silently propagate

Performance in production

• Avoid object arrays in hot paths
• Minimize copies between Pandas ↔ NumPy ↔ model runtime
• Profile with realistic batch sizes
• Consider float32 where precision sufficient

Testing

import numpy as np
expected = np.array([1, 2, 3])
actual = np.array([1, 2, 3])
assert np.array_equal(expected, actual)
print('arrays match')

Important interview questions and answers

1. Q: array_equal vs (a==b).all()?
   A: array_equal handles NaN with equal_nan option; raw == fails on NaN.
2. Q: Train-serve dtype mismatch?
   A: float64 train vs float32 serve can shift predictions—standardize.

Self-check

1. List five production NumPy checklist items.
2. Why assert shapes at boundaries?
3. How to compare arrays in tests?

Tip: Assert shapes at every pipeline stage boundary.