Floating-Point & Numeric Bugs

Q: Why does Python show 0.30000000000000004?

That is the closest binary64 representation. print shows shortest round-trip form; the value is not exactly 0.3.

Q: What rounding mode should billing use?

Depends on regulation; ROUND_HALF_UP is common for retail. Encode the rule in one function and test edge cases (0.5 ulp).

Q: How do pandas nullable Int64 columns help?

Store cents as Int64 in pandas 2.2+ to avoid float altogether until display formatting.

Q: Does Polars handle Decimal?

Polars 1.x supports Decimal dtypes for ETL; convert to string/int for JSON APIs explicitly.

Binary floats cannot represent many decimal fractions exactly. Financial totals, thresholds, and equality checks fail when code treats float as exact decimal arithmetic.

Recipe

Quick-reference recipe card - copy-paste ready.

from decimal import Decimal, ROUND_HALF_UP
 
price = Decimal("19.99")
qty = Decimal("3")
total = (price * qty).quantize(Decimal("0.01"), rounding=ROUND_HALF_UP)
 
# Never: if a == b for floats
import math
if math.isclose(a, b, rel_tol=1e-9, abs_tol=0.0):
    ...

When to reach for this:

0.1 + 0.2 != 0.3 breaks tests or billing
Pandas float64 aggregates disagree with spreadsheet totals
Sorting or binning produces off-by-one at boundaries
JSON APIs return noisy float tails (19.990000000002)

Working Example

from decimal import Decimal, ROUND_HALF_UP, getcontext
import math
 
# --- classic float surprise ---
assert 0.1 + 0.2 != 0.3  # True in IEEE 754 float
 
# --- money with Decimal ---
def line_total(unit_price: str, qty: int) -> Decimal:
    price = Decimal(unit_price)
    return (price * qty).quantize(Decimal("0.01"), rounding=ROUND_HALF_UP)
 
assert line_total("19.99", 3) == Decimal("59.97")
 
# --- safe float comparison ---
def nearly_equal(a: float, b: float) -> bool:
    return math.isclose(a, b, rel_tol=1e-9, abs_tol=1e-12)
 
# --- pandas 2.2+ aware aggregation ---
import pandas as pd
 
df = pd.DataFrame({"amount": ["10.10", "20.20"]})
# BUG path: astype(float) early can still accumulate binary error at scale
df["amount_dec"] = df["amount"].map(lambda x: Decimal(x))
total = sum(df["amount_dec"], Decimal("0"))
assert total == Decimal("30.30")
 
# --- context for banking rules ---
getcontext().prec = 28
tax_rate = Decimal("0.0825")
subtotal = Decimal("100.00")
tax = (subtotal * tax_rate).quantize(Decimal("0.01"), rounding=ROUND_HALF_UP)
assert tax == Decimal("8.25")

What this demonstrates:

Float equality is unsafe; math.isclose compares with tolerance
Decimal constructed from strings preserves decimal intent
quantize applies explicit rounding rules for money
pandas pipelines should delay float conversion or stay in Decimal until presentation

Deep Dive

How It Works

float is IEEE 754 binary64; many decimal rationals have no exact binary representation
Decimal stores base-10 coefficients with adjustable precision and rounding modes
Fraction handles rational numbers exactly when values are rational
Integer cents (int) avoids float entirely for money when scale is fixed

Type Picking

Use case	Type	Why
Money, tax, invoices	`Decimal` or int cents	deterministic decimal rules
Science/engineering	`float` + `isclose`	performance, numpy ecosystem
Ratios like 1/3	`Fraction`	exact rational
ML features	`float32`/`float64`	tensor APIs expect floats

Python Notes

# BAD Decimal from float - inherits binary error
Decimal(0.1)  # Decimal('0.1000000000000000055511151231257827021181583404541015625')
 
# GOOD
Decimal("0.1")

Gotchas

Decimal(float) constructor - imports float imprecision into Decimal. Fix: Decimal(str(x)) or parse from string input.
Mixing Decimal and float in ops - raises TypeError or coerces unexpectedly. Fix: pick one type per boundary.
JSON serialization - Decimal not JSON-native. Fix: quantize to string or int cents in API layer.
numpy/pandas float64 sum order - non-associative sums differ by order. Fix: Kahan summation, Decimal pass, or integer cents.
Floor/ceil on negatives - int(-1.2) truncates toward zero, not floor. Fix: math.floor / math.ceil.
Display rounding vs stored value - UI shows 2 decimals, backend compares full float. Fix: quantize at comparison boundary.

Alternatives

Alternative	Use When	Don't Use When
Integer minor units (cents)	fixed currency scale	FX with many decimal places
`Decimal`	human decimal rules	high-throughput ML inner loops
`float` + epsilon	graphics/simulation	regulatory financial totals
Polars 1.x Decimal dtype	typed ETL decimal columns	simple scripts

FAQs

Why does Python show 0.30000000000000004?

That is the closest binary64 representation. print shows shortest round-trip form; the value is not exactly 0.3.

Should APIs use float or string for money?

Prefer string decimal or integer minor units in JSON. If float is unavoidable, document tolerance and never use raw equality.

What rounding mode should billing use?

Depends on regulation; ROUND_HALF_UP is common for retail. Encode the rule in one function and test edge cases (0.5 ulp).

Is math.isclose enough for money?

No for equality of money. Use Decimal quantize or integer cents. isclose suits scientific floats.

How do pandas nullable Int64 columns help?

Store cents as Int64 in pandas 2.2+ to avoid float altogether until display formatting.

Does Polars handle Decimal?

Polars 1.x supports Decimal dtypes for ETL; convert to string/int for JSON APIs explicitly.

What about numpy.isclose vs math.isclose?

Both exist; numpy version handles arrays. Pick one module per codebase layer for consistency.

Can I use Fraction for money?

Possible for rational prices but awkward for decimal tax tables. Decimal matches human base-10 rules better.

Why do CSV imports drift totals?

Locale commas and implicit float parsing. Read as string, parse to Decimal, then quantize.

How do I test rounding?

Table-driven tests on half-cent boundaries, negative amounts, and three-way tax splits that must sum to invoice total.

Does PyTorch 2.6+ use float32 by default?

Many tensors default to float32. Do not use torch floats for ledger math; keep finance in Decimal/int paths.

When is binary float OK for currency display only?

When values are already quantized to cents and you only format for UI, but keep ledger in int cents end-to-end.

Data Pipeline Defects - dtype coercion side effects
Encoding & Unicode Bugs - parsing locale number strings
Debugging Refactor Snippets - float to Decimal refactors
Domain Modeling - Money value objects
Debugging Tools - inspect values in pdb

Stack versions: This page was written for Python 3.14.0 (stable 3.14, maintenance 3.13), FastAPI 0.115+, Django 5.2, Flask 3.1, Pydantic 2, PyTorch 2.6+, pandas 2.2+, Polars 1.x, ruff 0.9+, and uv 0.6+.

Recipe

Quick-reference recipe card - copy-paste ready.

from decimal import Decimal, ROUND_HALF_UP
 
price = Decimal("19.99")
qty = Decimal("3")
total = (price * qty).quantize(Decimal("0.01"), rounding=ROUND_HALF_UP)
 
# Never: if a == b for floats
import math
if math.isclose(a, b, rel_tol=1e-9, abs_tol=0.0):
    ...

When to reach for this:

0.1 + 0.2 != 0.3 breaks tests or billing
Pandas float64 aggregates disagree with spreadsheet totals
Sorting or binning produces off-by-one at boundaries
JSON APIs return noisy float tails (19.990000000002)

Working Example

from decimal import Decimal, ROUND_HALF_UP, getcontext
import math
 
# --- classic float surprise ---
assert 0.1 + 0.2 != 0.3  # True in IEEE 754 float
 
# --- money with Decimal ---
def line_total(unit_price: str, qty: int) -> Decimal:
    price = Decimal(unit_price)
    return (price * qty).quantize(Decimal("0.01"), rounding=ROUND_HALF_UP)
 
assert line_total("19.99", 3) == Decimal("59.97")
 
# --- safe float comparison ---
def nearly_equal(a: float, b: float) -> bool:
    return math.isclose(a, b, rel_tol=1e-9, abs_tol=1e-12)
 
# --- pandas 2.2+ aware aggregation ---
import pandas as pd
 
df = pd.DataFrame({"amount": ["10.10", "20.20"]})
# BUG path: astype(float) early can still accumulate binary error at scale
df["amount_dec"] = df["amount"].map(lambda x: Decimal(x))
total = sum(df["amount_dec"], Decimal("0"))
assert total == Decimal("30.30")
 
# --- context for banking rules ---
getcontext().prec = 28
tax_rate = Decimal("0.0825")
subtotal = Decimal("100.00")
tax = (subtotal * tax_rate).quantize(Decimal("0.01"), rounding=ROUND_HALF_UP)
assert tax == Decimal("8.25")

What this demonstrates:

Float equality is unsafe; math.isclose compares with tolerance
Decimal constructed from strings preserves decimal intent
quantize applies explicit rounding rules for money
pandas pipelines should delay float conversion or stay in Decimal until presentation

Deep Dive

How It Works

float is IEEE 754 binary64; many decimal rationals have no exact binary representation
Decimal stores base-10 coefficients with adjustable precision and rounding modes
Fraction handles rational numbers exactly when values are rational
Integer cents (int) avoids float entirely for money when scale is fixed

Type Picking

Use case	Type	Why
Money, tax, invoices	`Decimal` or int cents	deterministic decimal rules
Science/engineering	`float` + `isclose`	performance, numpy ecosystem
Ratios like 1/3	`Fraction`	exact rational
ML features	`float32`/`float64`	tensor APIs expect floats

Python Notes

# BAD Decimal from float - inherits binary error
Decimal(0.1)  # Decimal('0.1000000000000000055511151231257827021181583404541015625')
 
# GOOD
Decimal("0.1")

Gotchas

Decimal(float) constructor - imports float imprecision into Decimal. Fix: Decimal(str(x)) or parse from string input.
Mixing Decimal and float in ops - raises TypeError or coerces unexpectedly. Fix: pick one type per boundary.
JSON serialization - Decimal not JSON-native. Fix: quantize to string or int cents in API layer.
numpy/pandas float64 sum order - non-associative sums differ by order. Fix: Kahan summation, Decimal pass, or integer cents.
Floor/ceil on negatives - int(-1.2) truncates toward zero, not floor. Fix: math.floor / math.ceil.
Display rounding vs stored value - UI shows 2 decimals, backend compares full float. Fix: quantize at comparison boundary.

Alternatives

Alternative	Use When	Don't Use When
Integer minor units (cents)	fixed currency scale	FX with many decimal places
`Decimal`	human decimal rules	high-throughput ML inner loops
`float` + epsilon	graphics/simulation	regulatory financial totals
Polars 1.x Decimal dtype	typed ETL decimal columns	simple scripts

FAQs

Why does Python show 0.30000000000000004?

That is the closest binary64 representation. print shows shortest round-trip form; the value is not exactly 0.3.

Should APIs use float or string for money?

Prefer string decimal or integer minor units in JSON. If float is unavoidable, document tolerance and never use raw equality.

What rounding mode should billing use?

Depends on regulation; ROUND_HALF_UP is common for retail. Encode the rule in one function and test edge cases (0.5 ulp).

Is math.isclose enough for money?

No for equality of money. Use Decimal quantize or integer cents. isclose suits scientific floats.

How do pandas nullable Int64 columns help?

Store cents as Int64 in pandas 2.2+ to avoid float altogether until display formatting.

Does Polars handle Decimal?

Polars 1.x supports Decimal dtypes for ETL; convert to string/int for JSON APIs explicitly.

What about numpy.isclose vs math.isclose?

Both exist; numpy version handles arrays. Pick one module per codebase layer for consistency.

Can I use Fraction for money?

Possible for rational prices but awkward for decimal tax tables. Decimal matches human base-10 rules better.

Why do CSV imports drift totals?

Locale commas and implicit float parsing. Read as string, parse to Decimal, then quantize.

How do I test rounding?

Table-driven tests on half-cent boundaries, negative amounts, and three-way tax splits that must sum to invoice total.

Does PyTorch 2.6+ use float32 by default?

Many tensors default to float32. Do not use torch floats for ledger math; keep finance in Decimal/int paths.

When is binary float OK for currency display only?

When values are already quantized to cents and you only format for UI, but keep ledger in int cents end-to-end.

Data Pipeline Defects - dtype coercion side effects
Encoding & Unicode Bugs - parsing locale number strings
Debugging Refactor Snippets - float to Decimal refactors
Domain Modeling - Money value objects
Debugging Tools - inspect values in pdb

Stack versions: This page was written for Python 3.14.0 (stable 3.14, maintenance 3.13), FastAPI 0.115+, Django 5.2, Flask 3.1, Pydantic 2, PyTorch 2.6+, pandas 2.2+, Polars 1.x, ruff 0.9+, and uv 0.6+.

Floating-Point & Numeric Bugs

Search Documentation

Floating-Point & Numeric Bugs