rocketcsv

The only Rust-backed drop-in replacement for Python's csv module.

# Before
import csv

# After — everything else stays the same
import rocketcsv as csv

Same reader(), writer(), DictReader, DictWriter. Same parameters. Same behavior. 376 compatibility tests prove it.

Benchmarks

Three levels of speed, depending on how much code you want to change. Tested up to 877 MB. Full details: BENCHMARKS.md

Level 1 — Change one import line (zero other code changes)

import rocketcsv as csv          # swap this one line

with open("data.csv") as f:     # your existing code, untouched
    for row in csv.reader(f):
        process(row)

Scenario	Rows	stdlib	rocketcsv	Speedup
reader(), simple data	100K	0.068s	0.058s	1.2x
reader(), quoted fields	100K	0.127s	0.096s	1.3x
reader(), mixed types	100K	0.144s	0.115s	1.2x
DictReader(), quoted	100K	0.255s	0.175s	1.5x
writer(), mixed data	100K	0.208s	0.093s	2.2x

Level 2 — Pass a file path instead of a file object

rocketcsv extends reader() to accept a string path. When it sees a path, it reads the file entirely in Rust — no Python file handling overhead. This is not how stdlib csv.reader() works (stdlib requires a file object), but it's one small code change:

import rocketcsv as csv

# Before (stdlib-compatible):
# with open("data.csv") as f:
#     for row in csv.reader(f):

# After (rocketcsv extension — pass path directly):
for row in csv.reader("data.csv"):
    process(row)

File	Rows	stdlib open()+reader()	rocketcsv reader(path)	Speedup
175 MB	2,000,000	4.66s	3.32s	1.4x
395 MB	4,500,000	6.41s	2.65s	2.4x
877 MB	10,000,000	12.13s	7.20s	1.7x

Level 3 — Performance mode (lazy Rust-backed rows)

fast_reader_from_path() returns RocketRow objects instead of list[str]. Field data stays in Rust — a PyString is only created when you access row[i]. Columns you never touch cost zero.

import rocketcsv

for row in rocketcsv.fast_reader_from_path("data.csv"):
    if row[3] == "active":   # only this column is materialized
        print(row[0])         # and this one — the other 8 are free

File	Pattern	stdlib	fast_reader	Speedup
395 MB	Filter 1 column	4.82s	2.15s	2.2x
877 MB	Access 1 of 10 cols	10.77s	5.01s	2.2x
877 MB	Filter row[3] == "active"	10.63s	7.63s	1.4x

Writing — drop-in, same csv.writer() API

writerows() batches all formatting in Rust, then writes to your file in a single call.

import rocketcsv as csv

with open("out.csv", "w", newline="") as f:
    w = csv.writer(f)            # same API, nothing new to learn
    w.writerows(rows)            # 2.2x faster on mixed data

Scenario	Rows	stdlib	rocketcsv	Speedup
Mixed types (strings, numbers, dates)	100K	0.208s	0.093s	2.2x
Quoted fields (commas, newlines in data)	100K	0.127s	0.096s	1.3x
Narrow table (2 cols, high row count)	1M	0.234s	0.158s	1.5x

Compatibility

Metric	Result
Compat tests (Python 3.12)	376/376 pass
Compat tests (Python 3.11)	368/368 pass (8 skip: 3.12 features)
Real-world corpus (121 files)	100% pass

Installation

# From GitHub (requires Rust toolchain)
pip install git+https://github.com/leonardcatalinbira/rocketcsv.git

PyPI release coming soon. Python 3.11+. Linux, macOS, Windows.

Three ways to read

import rocketcsv as csv

# 1. Drop-in — swap one import, change nothing else (1.2-1.3x)
with open("data.csv") as f:
    for row in csv.reader(f):
        process(row)

# 2. Pass a path — auto-detects, reads in Rust (1.7-2.4x)
for row in csv.reader("data.csv"):
    process(row)

# 3. Performance mode — lazy rows, zero-cost unused columns (2.2x)
for row in rocketcsv.fast_reader_from_path("data.csv"):
    if row[2] == "IT":     # only this field is materialized
        name = row[0]       # and this one
        # row[1], row[3]..row[9] — never allocated

Full API

Everything in import csv works in import rocketcsv as csv:

Feature	Status
reader() / writer()	Drop-in compatible
DictReader / DictWriter	Drop-in compatible
Sniffer (sniff + has_header)	Drop-in compatible
Dialect support	Full (register, get, list, unregister)
All format parameters	Full (delimiter, quotechar, escapechar, doublequote, skipinitialspace, lineterminator, quoting, strict)
QUOTE_* constants	All 6 including 3.12+ (QUOTE_STRINGS, QUOTE_NOTNULL)
field_size_limit()	Enforced
reader_from_path()	rocketcsv-only, 1.4-2.4x faster
fast_reader() / fast_reader_from_path()	rocketcsv-only, lazy Rust-backed rows

How it works

CSV parsing happens in Rust via BurntSushi's csv crate. Python bindings via PyO3. Packaged with maturin.

• Per-column string interning — repeated values (status codes, countries, categories) are cached as Python objects and reused across rows. Auto-disables on high-cardinality columns
• Raw CPython ffi — PyUnicode_FromStringAndSize + PyList_SET_ITEM skip intermediate allocations and bounds checks
• Rust file I/O — reader_from_path() uses std::fs::read(), bypassing Python's file handling entirely
• Lazy RocketRow — fast_reader() parses in Rust but defers Python object creation to __getitem__. Columns you never access cost zero
• Batched writer — writerows() formats everything in a single Rust buffer, one .write() call

Why not Polars / PyArrow?

Those are great — for DataFrames. But they replace the API entirely. If your code uses csv.reader() or csv.DictReader(), switching to Polars means rewriting.

rocketcsv is for the millions of existing codebases that use import csv. One line change. Zero refactoring.

License

Dual-licensed:

• Open source: LGPLv3 — use freely in any project, open or proprietary. import rocketcsv in commercial software with zero obligation to open-source your code.
• Commercial: For modifying rocketcsv internals privately, or for warranty/support/indemnification. Contact for pricing.

Contributing

pip install maturin pytest
maturin develop
pytest tests/ -v   # 376 tests