NumPy: arrays at speed

Why this matters

Pandas, scikit-learn, PyTorch tensors, and every quant/ML stack sit on numpy's ndarray and its C-level broadcasting. The staff-level skill is spotting Python-level for loops over arrays and rewriting them as vectorized ufuncs — often a 100-1000× speedup and a memory-contiguous win for downstream BLAS calls.

Common pitfalls

• Iterating with for x in arr — kills the point of numpy; use vectorized ops or np.vectorize only as a last resort.
• Mixing dtypes silently — an int64 array + Python float upcasts to float64 and doubles memory.
• Assuming slices are copies — arr[1:3] is a view; mutating it edits the original. Use .copy() when needed.