Data types (dtypes) & memory

Every NumPy array has a single dtype that decides how each element is stored in memory. Picking the right one can shrink a dataset several times over, and picking the wrong one can silently give wrong answers.

Common dtypes

dtype	Bytes per value	Range / precision	Typical use
int8	1	-128 to 127	Small codes, ratings 1–5
int32	4	about ±2.1 billion	Counts, IDs
int64	8	about ±9.2 x 10^18	Default integer on most platforms
float32	4	~7 significant digits	ML features, images
float64	8	~15–16 significant digits	Default float, finance, science
bool	1	True / False	Masks

Inspecting and converting

import numpy as np

qty = np.array([3, 5, 2, 8])
print(qty.dtype, qty.itemsize, qty.nbytes)   # int64 8 32

qty_small = qty.astype(np.int16)
print(qty_small.nbytes)                      # 8

prices = np.array([199.0, 249.5, 99.9], dtype=np.float32)
print(np.iinfo(np.int16).max)                # 32767
print(np.finfo(np.float32).eps)              # machine precision of float32

astype always returns a new array, so assign the result back.

Memory at scale

One crore (10 million) sensor readings:

n = 10_000_000
print(np.zeros(n, dtype=np.float64).nbytes / 1e6)   # 80.0 MB
print(np.zeros(n, dtype=np.float32).nbytes / 1e6)   # 40.0 MB
print(np.zeros(n, dtype=np.int8).nbytes / 1e6)      # 10.0 MB

For store IDs between 0 and 500, int16 is plenty. For temperatures to one decimal place, float32 is enough.

Trap 1: integer overflow wraps silently

Array arithmetic does not check for overflow. It simply wraps around:

stock = np.array([100, 120], dtype=np.int8)
print(stock + stock)        # [-56 -16]  -- wrong, no error raised!

Always check np.iinfo(dtype) limits before downcasting, especially for sums and products, which grow fast.

Trap 2: floating-point precision

Binary floats cannot represent most decimals exactly:

print(0.1 + 0.2 == 0.3)               # False
print(np.isclose(0.1 + 0.2, 0.3))     # True

Compare floats with np.isclose or np.allclose, never ==. For money, many teams store paise as integers (₹249.50 becomes 24950) to avoid rounding drift in totals.

Trap 3: upcasting in mixed operations

Combining dtypes promotes to a type that can hold both, so an int8 array plus a float64 array gives float64. Your carefully small array can quietly become large again.

Strings and objects

np.array(["Pune", "Delhi"]) creates a fixed-width Unicode dtype such as <U5, sized to the longest string. Mixed Python objects give dtype=object, which loses NumPy's speed because every element is a Python object. For text-heavy data, pandas is usually the better tool.

Key takeaways

• Check dtype, itemsize and nbytes when memory matters.
• Downcast only after checking value ranges with np.iinfo / np.finfo.
• Integer overflow in arrays is silent; floats are approximate.
• Avoid object arrays for numerical work.