Data Cleaning

Data cleaning — make dirty data trustworthy, and make the cleaning auditable

A clean table is typed + deduped + normalized + validated + reproducible. The deliverable here is never "I opened a notebook and fixed some rows by hand." It is a re-runnable function clean(raw) -> df plus a schema gate that fails loud when next month's file violates the contract. Reproducible means the same input always yields the same output: versions pinned, sorts deterministic, nothing random without a seed. If you can't re-run it tomorrow and get the identical result, you haven't cleaned the data — you've edited a snapshot.

Cleaning starts once you hold tabular rows and ends at a validated table/DataFrame/Parquet. Before that boundary the job is acquisition (data-scraper, structured-extraction); after it, consumption (spreadsheet-ops, analytics, business-intelligence, forecasting). Multi-GB analytical SQL is an engine choice, not a cleaning one — duckdb.

Current stack (verified 2026-06-02): pandas 3.0.x (3.0.0 shipped 2026-01-21) and pandera 0.31.1 (supports pandas ≥ 3) for in-pipeline schema validation; Polars and DuckDB when pandas runs out of RAM. Pin them: pandas==3.0.3, pandera==0.31.1.

The pipeline shape

One canonical order. Each step is positioned for a reason, not by habit.

import pandas as pd

def clean(raw_path: str) -> pd.DataFrame:
    df = read_typed(raw_path)     # 1. read with explicit dtypes — never let pandas guess
    df = normalize(df)            # 2. strings/categories/numbers/dates — collapse invisible variance
    df = dedupe(df)               # 3. AFTER normalize+type, so "1"/1 and "US "/"US" actually collapse
    df = handle_missing(df)       # 4. decide per column: drop / impute+flag / leave NA / quarantine
    df = Schema.validate(df, lazy=True)  # 5. the GATE — fail loud, surface every violation at once
    return df

• Type before dedupe — otherwise "1" (string) and 1 (int) survive as two distinct keys.
• Normalize before dedupe — "US " and "US" are the same customer; dedupe can't see that until whitespace/case are collapsed.
• Validate last — it is the gate, not a cleaning step. It asserts the contract holds after all fixes.
• Write to a NEW artifact — the raw file is read-only; you never overwrite your only source.

Read it right

The single most common reproducibility footgun: pandas' legacy numpy path silently casts an integer column containing one NaN to float64, so your id becomes 1001.0. Control the dtype on read.

# BAD — pandas guesses: ids become floats, "N/A" stays a string, "" is sometimes NaN sometimes ""
df = pd.read_csv("raw.csv")

# GOOD — explicit, deterministic, real nullable types
df = pd.read_csv(
    "raw.csv",
    dtype_backend="pyarrow",          # real nullable ints/strings; no silent float-cast
    na_values=["", "N/A", "NA", "null", "-1", "999"],  # YOUR sentinels become real NA
    keep_default_na=True,             # keep pandas' default NA tokens too
    encoding="utf-8",                 # state it; don't let locale decide
)

Two pandas 3.0 facts that read depends on. dtype_backend="pyarrow" only works if pyarrow is actually installed — PDEP-14 deliberately kept a NumPy-object fallback so PyArrow stays recommended, not required — so pip install pyarrow for the faster backed path, or pass dtype_backend="numpy_nullable" when it is absent. And the default str dtype (PyArrow-backed when pyarrow is present, NumPy-object-backed otherwise) uses NaN missing-value semantics like every other default dtype: test for null with pd.isna(), never by comparing against whichever null token happened to appear.

Profile before you fix

Let the numbers drive the plan, not a glance at df.head(). Run this first, every time.

def profile(df: pd.DataFrame) -> pd.DataFrame:
    return pd.DataFrame({
        "dtype":     df.dtypes.astype(str),
        "null_pct":  (df.isna().mean() * 100).round(1),
        "n_unique":  df.nunique(dropna=True),       # cardinality — catches category sprawl
        "sample":    df.apply(lambda s: s.dropna().unique()[:3].tolist()),
    })

print(profile(df))
print("rows:", len(df), "exact dupes:", df.duplicated().sum())

A column at 90% null is a drop candidate; one with 400 distinct "countries" needs a mapping table; an "age" with min -1/max 999 has sentinels to map. The profile is your TODO list.

Normalize

Each fix below: Bad → Good, with a one-line why.

Strings — invisible variance (trailing space, mixed case, lookalike unicode) silently breaks joins and dedupe.

# BAD: "US ", "us", "ｕｓ" all look different to a join
# GOOD:
s = df["country"].str.strip().str.casefold().str.normalize("NFKC")

Categories — use a mapping table, never a tower of regex. A dict is auditable and an unmapped value gets quarantined instead of silently passing through.

COUNTRY = {"usa": "US", "u.s.": "US", "united states": "US", "u.s.a.": "US", "es": "ES", "españa": "ES"}
key = df["country"].str.strip().str.casefold()
df["country"] = key.map(COUNTRY)            # unmapped -> NA, which the gate below will catch (no silent pass)

Numbers — turn sentinels into NA, then choose a range policy explicitly: clip (cap to bound) when out-of-range is plausibly a recording cap, reject (→ NA / quarantine) when it is impossible.

df["age"] = df["age"].mask(df["age"].isin([-1, 999]))   # sentinels -> NA
df["age"] = df["age"].clip(lower=0, upper=120)          # clip policy; or .mask(~df["age"].between(0,120)) to reject

Dates — state the format, coerce, then count the casualties. Never trust dayfirst inference; 03/04/2026 is ambiguous and pandas will pick silently.

parsed = pd.to_datetime(df["signup"], format="%Y-%m-%d", utc=True, errors="coerce")
bad = parsed.isna() & df["signup"].notna()
assert bad.sum() == 0, f"{bad.sum()} dates failed the expected format — inspect before proceeding"
df["signup"] = parsed

Copy-paste versions of all of these — category mapping with unmapped→quarantine, a robust date parser, unicode/encoding repair, a sentinel→NA table, numeric clip-vs-reject, plus Polars equivalents — are in references/normalization-recipes.md.

Dedupe

drop_duplicates(keep="first") is meaningless without a defined key and a stable sort — "first" of what order? Define both.

key = ["customer_id"]                                   # the BUSINESS key, stated explicitly
df = (df.sort_values(["customer_id", "updated_at"], ascending=[True, False], kind="stable")
        .drop_duplicates(subset=key, keep="first"))     # keep most-recent per customer, deterministically

Near-duplicates ("Acme Inc" vs "Acme, Inc.") are a normalization problem — collapse them in the normalize step first; only then does exact dedupe catch them. Fuzzy matching is a separate, riskier decision — make it visible, never automatic.

Missing values — decide per column

No silent fillna(0): a zero is a value, and treating "unknown" as zero poisons every mean, sum, and model downstream. Pick deliberately.

df["income_was_missing"] = df["income"].isna()
df["income"] = df["income"].fillna(df["income"].median())   # impute + flag, never bare fillna(0)