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.

Situation	Action	Why
Column is mostly null (e.g. >70%) and not load-bearing	Drop the column	Imputing it invents signal that isn't there
A few rows missing a required key (id, date)	Drop the row (and log/quarantine)	Can't dedupe or join without the key
Numeric gap you must fill for a model	Impute and add a _was_missing flag	The model can learn "was missing"; you keep the audit trail
Genuinely optional field	Leave NA	NA is information; don't fabricate a value
Value is present but invalid (unmapped category, bad date)	Quarantine the row	Don't drop silently and don't let it pass the gate

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

Validate — the gate

This is where cleaning becomes trustworthy. Declare the contract as a pandera DataFrameModel, validate output (and input expectations where they exist), and split valid rows from failures instead of crashing — the failures become your quarantine.

import pandera.pandas as pa
from pandera.typing import Series

class CustomerSchema(pa.DataFrameModel):
    customer_id: Series[int]   = pa.Field(unique=True, ge=1)
    country:     Series[str]   = pa.Field(isin=["US", "ES", "FR"])     # only mapped categories survive
    age:         Series[float] = pa.Field(ge=0, le=120, nullable=True)
    signup:      Series[pa.DateTime] = pa.Field(nullable=False)

    class Config:
        strict = True       # reject unexpected columns
        coerce = True       # coerce to declared dtype, fail loud if impossible

# lazy=True collects EVERY violation at once instead of dying on the first
try:
    valid = CustomerSchema.validate(df, lazy=True)
except pa.errors.SchemaErrors as e:
    failures = e.failure_cases          # dataframe of exactly which rows/checks failed
    failures.to_parquet("quarantine.parquet")   # keep, don't drop — someone investigates these
    valid = df.drop(index=e.failure_cases["index"].dropna().unique())  # proceed with the clean subset

coerce=True fixes types the contract expects; nullable states which columns may hold NA; field Checks (ge, le, isin, unique) are the allowed-value rules. strict catches columns that shouldn't be there. Together they are the data contract in code. Log the row-count diff on every run — in, out, coerced, quarantined — so what the pipeline changed is an auditable record, not an assumption.

When to escalate beyond pandera: reach for GX Core 1.0 (Great Expectations' rebranded OSS — Data Context → Data Source → Expectation Suite → Validation Definition → Checkpoint) when you need a shared data-quality platform across many datasets and teams with a results store and docs. Use dbt model contracts (enforced at build) plus dbt tests (post-materialization) when the cleaning lives in a SQL warehouse, not Python. The full DataFrameModel (custom @pa.check, lazy SchemaErrors report, valid/quarantine split helper), the GX checkpoint sketch, the dbt model-contract + data_tests YAML, and the "which validator" chooser are in references/validation-patterns.md.

Scale — when pandas hurts

Heuristic: pandas is fine while the data fits comfortably in RAM (roughly ≤ 1–2 GB working set). Beyond that, or when a groupby/join dominates the runtime, switch the mechanics (not the principles):

• Polars for clean-at-scale: pl.scan_csv(...) (lazy, parallel, Rust), then .unique(), .drop_nulls(), .fill_null(...), .str.* — the same profile→normalize→dedupe→validate shape, faster. pandera validates Polars frames too, and the recipes reference has the Polars equivalent of every fix above.
• DuckDB when the bottleneck is analytical SQL over multi-GB files — point heavy joins/aggregations there: duckdb. It is an engine choice; correctness/normalization is still this skill's job.

Anti-patterns

Anti-pattern	Why it breaks
"fillna(0) to get rid of the nulls"	Zero is a value; it distorts every mean/sum/model. Impute deliberately and add a _was_missing flag.
"drop_duplicates() — done"	No subset, no sort → which row survives is nondeterministic. Define the key, sort_values(kind="stable"), set keep.
"pd.read_csv(path) and start cleaning"	pandas guesses: ids become floats, dates become strings. Pass dtype_backend + na_values.
"I fixed the rows in a notebook cell"	Not reproducible — next month's file gets nothing. Wrap it in clean(raw) -> df.
"Drop the rows that look wrong"	Silent data loss with no audit trail. Quarantine to a file; someone investigates.
"A few regexes will normalize the countries"	Unmaintainable and silent on new values. Use a mapping dict; unmapped → NA → caught by the gate.
"pd.to_datetime figures out the format"	Ambiguous dates parse silently wrong. State format=, errors="coerce", then assert the NaT count.
"Validation passed, so we're good"	A gate that never fails is a no-op. Feed it a known-bad row and confirm it rejects.
"It's slow, rewrite everything in Polars"	Switch the engine, not the discipline — profile→normalize→dedupe→validate still applies.

Verify

scripts/verify.sh runs from anywhere, no network. It does static structure checks on this skill (frontmatter keys, references present) always, and — when pandas + pandera are installed — extracts the documented pattern, feeds it one clearly-good row and one clearly-bad row, and asserts the good row PASSES validation while the bad row is FLAGGED/quarantined, proving the gate is not a no-op. Without pandas/pandera it prints SKIP for the runtime check and still passes the static checks.

Project grounding (02-DOCS + CLAUDE.md)

In a project with a 02-DOCS/ layer (the harness wiki), record this dataset's cleaning decisions — the schema/contract, the category mapping tables, the dedupe key, the quarantine location, version pins — in 02-DOCS/wiki/data/<dataset>.md, link it from the root CLAUDE.md ## Knowledge map, and read it first on every re-run so the contract stays consistent. No 02-DOCS/? Skip silently. Conventions are recorded, never gated.