csv-turbo

High-performance CSV reader/writer for Python — streaming, typed columns, large files.

Features

• Streaming by default — iterate rows without loading the entire file into memory
• Typed columns — declare int, float, bool, date, datetime schemas and get cast values automatically
• Schema validation — catch missing/extra columns and type errors early
• Dialect detection — auto-detect delimiter, quoting style, and line endings
• Pipeline API — composable lazy transformations: filter, map, select, rename, chunk, sort, unique
• Column statistics — profile any CSV with counts, nulls, min/max, mean, std dev, top-N values
• Built-in dialects — RFC 4180, TSV, semicolon, pipe, Excel presets
• Zero dependencies — pure Python ≥ 3.9, standard library only

Installation

pip install csv-turbo

Quick Start

Reading

from csv_turbo import CsvReader

for row in CsvReader("data.csv"):
    print(row["name"], row["score"])

Writing

from csv_turbo import CsvWriter

rows = [
    {"id": 1, "name": "Alice", "score": 9.5},
    {"id": 2, "name": "Bob",   "score": 8.0},
]

with CsvWriter("output.csv") as w:
    w.writerows(rows)

Typed Schema

from csv_turbo import CsvReader, Schema, ColumnDef, INT, FLOAT, STRING, DATE

schema = Schema([
    ColumnDef("id",    INT,    nullable=False),
    ColumnDef("name",  STRING),
    ColumnDef("score", FLOAT,  nullable=True, default=0.0),
    ColumnDef("dob",   DATE),
])

for row in CsvReader("data.csv", schema=schema):
    # row["id"] is int, row["score"] is float, row["dob"] is datetime.date
    print(row["id"], row["score"])

Streaming Large Files (Chunked)

from csv_turbo import CsvReader

reader = CsvReader("huge_file.csv")
for batch in reader.chunks(500):
    db.bulk_insert(batch)           # list of 500 row dicts

Pipeline API

from csv_turbo import CsvReader
from csv_turbo.streaming import Pipeline

result = (
    Pipeline(CsvReader("sales.csv", schema=schema))
    .filter(lambda r: r["amount"] > 100)
    .map(lambda r: {**r, "vat": round(r["amount"] * 0.2, 2)})
    .select("id", "name", "amount", "vat")
    .sort(key=lambda r: r["amount"], reverse=True)
    .take(50)
    .to_list()
)

Write pipeline output directly to CSV:

Pipeline(CsvReader("raw.csv"))     \
    .where(status="active")        \
    .drop("internal_notes")        \
    .write_csv("clean.csv")

Dialect Detection

from csv_turbo.dialect import detect_dialect_from_file

dialect = detect_dialect_from_file("european_export.csv")
print(dialect.delimiter)   # likely ";"

for row in CsvReader("european_export.csv", delimiter=dialect.delimiter):
    ...

Use a preset dialect:

from csv_turbo import TSV, CsvReader

for row in CsvReader("data.tsv", delimiter=TSV.delimiter):
    ...

Column Statistics / Profiling

from csv_turbo import CsvReader, profile

rows = CsvReader("data.csv").read_all()
data_profile = profile(rows, top_n=5)
print(data_profile)

# Access per-column stats
stats = data_profile.column("score")
print(stats.mean, stats.std_dev, stats.null_count)

Infer Schema Automatically

from csv_turbo import CsvReader

schema = CsvReader.infer_schema("data.csv", sample_size=500)
print(schema)   # Schema([id:int, name:string, score:float, dob:date])

for row in CsvReader("data.csv", schema=schema):
    print(row)

Read from String (Testing / Inline Data)

from csv_turbo import CsvReader

csv_text = "id,name,score\n1,Alice,9.5\n2,Bob,8.0\n"
for row in CsvReader.from_string(csv_text):
    print(row)

Write to String

from csv_turbo import CsvWriter

csv_str = CsvWriter.to_string([
    {"x": 1, "y": 2},
    {"x": 3, "y": 4},
])
print(csv_str)

API Reference

CsvReader

CsvReader(
    source,            # str | Path | file-like
    *,
    schema=None,       # Schema for typed casting
    delimiter=",",
    quotechar='"',
    encoding="utf-8",
    has_header=True,
    skip_blank_lines=True,
    strip_whitespace=True,
    null_values={""},  # Values treated as None
    row_filter=None,   # Callable[[dict], bool]
    row_transform=None,# Callable[[dict], dict]
    strict=False,      # Raise on extra columns
    max_errors=0,      # Tolerated cast errors
)

Method / Property	Description
__iter__()	Iterate rows as dicts
chunks(size)	Iterate in batches
read_all()	Load all rows into a list
count_rows()	Count without materialising
headers	Column names after first read
error_count	Cast errors encountered
CsvReader.from_string(text)	Create from a CSV string
CsvReader.infer_schema(path)	Auto-detect schema

CsvWriter

CsvWriter(
    destination,           # str | Path | file-like
    *,
    fieldnames=None,       # Column order
    delimiter=",",
    quotechar='"',
    encoding="utf-8",
    write_header=True,
    formatter=None,        # Callable[[Any], str]
    column_formatters={},  # Per-column formatters
    row_transform=None,
    append=False,
    buffer_size=8192,
)

Method	Description
writerow(row)	Write one dict row
writerows(rows)	Write an iterable of rows
rows_written	Count of rows written
CsvWriter.to_string(rows)	Render to string

Schema & ColumnDef

Schema([
    ColumnDef(name, type, nullable=True, default=None, alias=None),
    ...
])

Built-in type singletons: STRING, INT, FLOAT, BOOL, DATE, DATETIME

Parameterised constructors: IntType(min_value=0), FloatType(precision=2), DateType(fmt="%d/%m/%Y"), StringType(max_length=255)

Pipeline

Method	Description
.filter(fn)	Keep rows where fn(row) is True
.map(fn)	Transform each row
.select(*cols)	Keep only named columns
.rename(mapping)	Rename columns
.drop(*cols)	Remove columns
.add_field(name, fn)	Compute a new column
.where(**kwargs)	Equality filter shorthand
.skip(n)	Skip first n rows
.take(n)	Keep first n rows
.chunk(size)	Group into batches
.sort(key)	Sort (materialises)
.unique(key)	Deduplicate rows
.peek(fn)	Side-effect per row
.to_list()	Materialise to list
.to_dict(key)	Index by column
.count()	Count rows
.first()	First row or None
.aggregate(col, ...)	Sum/mean/min/max
.write_csv(path)	Write to file

profile(rows)

data_profile = profile(rows, top_n=10)
# DataProfile.row_count, .columns, .column(name)
# ColumnStats.count, .null_count, .fill_rate, .unique_count
# ColumnStats.mean, .std_dev, .min_value, .max_value, .median (numeric)
# ColumnStats.min_length, .max_length, .avg_length (string)
# ColumnStats.top_values — list of (value, count)

Error Handling

All exceptions inherit from CsvTurboError:

Exception	When
ParseError	Malformed CSV structure
TypeCastError	Value cannot be cast to declared type
SchemaValidationError	Missing/extra columns
WriteError	Write failure
ConfigurationError	Invalid reader/writer options

from csv_turbo import CsvReader, TypeCastError, SchemaValidationError

try:
    for row in CsvReader("data.csv", schema=schema, max_errors=5):
        process(row)
except TypeCastError as e:
    print(f"Bad value '{e.value}' in column '{e.column}' on line {e.line}")
except SchemaValidationError as e:
    print(f"Missing columns: {e.missing_columns}")

License

MIT © Vladyslav Zaiets