Money Decimal

Money with Decimal (fixed precision, never float)

When to use

Any field, calculation, or serializer that carries money — line totals, taxes, fees, discounts, withholdings. On money, a rounding bug is a correctness bug that shows up as customer-facing cents that do not add up, so treat precision as an invariant, not a formatting detail.

Pattern

Two rules, held everywhere money is stored or computed:

1. Money is Decimal at a fixed scale, never float. Binary floats cannot

represent decimal fractions exactly (0.1 + 0.2 != 0.3), so float money drifts by a cent under multiplication and summation. Store as DecimalField at the currency's minor-unit precision.

1. quantize with an explicit rounding mode at every step of a multi-step chain,

not only on the final total. Rounding a tax, then rounding the fee on top of the already-rounded tax, gives a different answer than rounding once at the end — and the stepwise answer is the one on the printed document.

Concurrency (two requests writing the same balance) is a separate concern — lock the row with db-concurrency. This skill is only about precision and rounding.

Adapt to your repo

Define the currency scale as a parameter, not a magic number: most currencies use 2 decimal places, some use 0 or 3. Set decimal_places to your minor-unit precision and max_digits to cover the largest total you expect. Declare the valid rate set and the precision in exactly one module and import it into the model, the serializer, and (mirrored) the frontend, so a rate added in one place cannot silently be missing in another. Rename the example fields to your domain.

# money.py — the single source of truth, imported everywhere
from decimal import Decimal, ROUND_HALF_UP

MONEY_SCALE = Decimal("0.01")            # 2 dp; use "1" for 0-dp, "0.001" for 3-dp
VALID_TAX_RATES = frozenset(map(Decimal, ["0", "0.10", "0.20"]))  # tune per jurisdiction

def money(value) -> Decimal:
    """Round any intermediate to the currency scale with an explicit mode."""
    if isinstance(value, float):
        # a float has already lost precision; quantize would only hide it.
        # Don't "fix" this with Decimal(str(value)) — that launders the error.
        raise TypeError("money() rejects float — pass Decimal, int, or str")
    return Decimal(value).quantize(MONEY_SCALE, rounding=ROUND_HALF_UP)

# stepwise chain: quantize the tax, THEN the fee — not once at the end
net  = money(unit_price * quantity)
tax  = money(net * tax_rate)                 # rate ∈ VALID_TAX_RATES
fee  = money((net + tax) * fee_rate)
total = net + tax + fee                       # already-quantized parts sum exactly

Model and serializer both pull from the same module:

# models.py
amount = models.DecimalField(max_digits=14, decimal_places=2)  # decimal_places = scale

Gotchas

• Never build a Decimal from a float literal — Decimal(0.1) carries the float's

error. Pass a string or int: Decimal("0.1"). Enforce it at the boundary — money() raises TypeError on a float rather than quantizing the error away. quantize hides the drift for most values, so a laundered float is a latent cent-drift, not a loud failure: money(2.675) would give 2.67 where money(Decimal("2.675")) gives 2.68.

• Pick the rounding mode deliberately and reuse it; ROUND_HALF_UP and

ROUND_HALF_EVEN disagree on exact-half cases, and the choice is a policy, not a default. Whatever you pick, apply it consistently across every step.

• Validate incoming rates against the shared VALID_TAX_RATES set — an unlisted rate

is a data error, not a silent computation.

• Summing a column of un-quantized intermediates then rounding once can differ from

the document's line-by-line total; round at the boundary the document rounds at.

• The DB column's decimal_places and the app's MONEY_SCALE must match, or a

round-trip silently re-rounds.

Golden-fixture regression test

Pin one hand-verified worked example so any drift in the calculation fails loudly. Compute a realistic multi-line order by hand once, assert the exact total, and also test the boundaries — zero, each rate-bracket edge, and the min/max representable amount (see write-tests).

def test_order_total_golden():
    # hand-verified: net 14.97 (3 x 4.99) + tax 2.99 (20% of 14.97, half-up) + flat fee 1.50 = 19.46
    assert compute_total(unit_price=Decimal("4.99"), quantity=3,
                         tax_rate=Decimal("0.20"), fee=Decimal("1.50")) == Decimal("19.46")

def test_zero_and_bracket_edges():
    assert compute_total(Decimal("0"), 0, Decimal("0"), Decimal("0")) == Decimal("0.00")

See also

• db-concurrency
• write-tests
• drf-api

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: Deadlymind
• Source: Deadlymind/nanolama
• License: MIT

Install and usage instructions live in the source repository linked above.