lazycsv

lazycsv - a memory-efficient csv parser

Developers: Michael Green, Chris Perkins

lazycsv is a C implementation of a csv parser for python. The aim of this parser is to provide for fast extraction of sequences of data from a CSV file in a memory-efficient manner, with zero dependencies.

LazyCSV utilizes memory mapped files and iterators to parse the file without persisting any significant amounts of data to physical memory. The design allows a user to generate PyObject's from a csv file lazily.

The parser works as follows:

First, The user file is memory-mapped internally to the LazyCSV object. That file is used to generate three indexes. The first is an index of values which correspond to the position in the user file where a given CSV field starts. This value is always a uint16_t which we found to be the optimal bit size for disk usage and execution performance (This type can however be changed by setting the LAZYCSV_INDEX_DTYPE environment variable to any unsigned integer type). For index values outside the range of an unsigned short, An "anchor point" is created, which is a pair of size_t values that mark both the value which is subtracted from the index value such that the index value fits within 16 bits, and the first column of the CSV where the anchor value applies. This anchor point is periodically written to the second index file when required for a given comma index. Finally, the third index writes the index of the first anchor point for each row of the file.

When a user requests a sequence of data (i.e. a row or a column), an iterator is created and returned. This iterator uses the value of the requested sequence and its internal position state to index into the index files the values representing the index of the requested field, and its length. Those two values are then used to create a single PyBytes object. These PyBytes objects are then yielded to the user per-iteration.

This process is lazy, only yielding data from the user file as the iterator is consumed. It does not cache results as they are generated - it is the responsibility of the user to store in physical memory the data which must be persisted. The only persisted overhead in physical memory is the LazyCSV object itself, any created iterators, a small cache of common length-0 and length-1 PyObject*'s for fast returns, and optionally the headers of the CSV file.

>>> from lazycsv import lazycsv
>>> lazy = lazycsv.LazyCSV("tests/fixtures/file.csv")
>>> lazy
<lazycsv.LazyCSV object at 0x7f5b212ea3d0>
>>> (col := lazy.sequence(col=0))
<lazycsv_iterator object at 0x7f5b212ea420>
>>> next(col)
b'0'
>>> next(col)
b'1'
>>> next(col)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
StopIteration

Since data is yielded through the iterator protocol, lazycsv pairs well with many of the builtin functional components of Python, and third-party libraries with support for iterators. This has the added benefit of keeping iterations in the C level, maximizing performance.

>>> row = lazy.sequence(row=1)
>>> list(map(lambda x: x.decode('utf8'), row))
['1', 'a1', 'b1']
>>>
>>> import numpy as np
>>> np.fromiter(map(int, lazy.sequence(col=0)), dtype=np.int64)
array([0, 1])

The lazy object also supports indexing operations for expressive iterables. The axis for iteration can be passed as a slice object, and the index of the iterable can be passed as a integer. Individual coordinate values can also be passed as a pair of integers, this call will eagerly return the value at that index.

>>> list(lazy[::-1, 1])
[b'a1', b'a0']
>>> lazy[-1, -1]
b"b1"

Iterators can be materialized at any point by calling the to_list() or to_numpy() methods on the iterator object (to enable optional numpy support, see the Numpy section of this document). These methods exhaust the iterator, placing the remaining PyBytes values into a PyObject.

>>> col = lazy[:, 0]
>>> next(col)
b'0'
>>> col.to_list()
[b'1']
>>>

Headers are by default parsed from the csv file and packaged into a tuple under a .headers attribute. This can be skipped by passing skip_headers=True to the object constructor. Skipping the header parsing step results in the header value being included in the iterator.

Note: lazycsv makes no effort to deduplicate headers and it is the responsibility of the user to make sure that columns are properly named.

>>> lazy.headers
(b'', b'ALPHA', b'BETA')
>>> (col := lazy.sequence(col=1))
<lazycsv_iterator object at 0x7f599fd86b50>
>>> list(col)
[b'a0', b'a1']
>>> lazy = lazycsv.LazyCSV(FPATH, skip_headers=True)
>>> (col := lazy[:, 1])
<lazycsv_iterator object at 0x7f59d1b21890>
>>> list(col)
[b'ALPHA', b'a0', b'a1']

Fields which are double-quoted by default are yielded without quotes. This behavior can be disabled by passing unquoted=False to the object constructor.

>>> lazy = lazycsv.LazyCSV(
...     "tests/fixtures/file_crlf2.csv"
... )
>>> lazy.headers
(b'', b'This,that', b'Fizz,Buzz')
>>> lazy = lazycsv.LazyCSV(
...     "tests/fixtures/file_crlf2.csv", unquote=False
... )
>>> lazy.headers
(b'', b'"This,that"', b'"Fizz,Buzz"')

LazyCSV also provides the option to specify a delimiter and a quote character. Pass the keywords delimiter= and quotechar= to the object contstructor to use custom values. By default, delimiter defaults to , and quotechar defaults to ".

>>> lazy = lazycsv.LazyCSV(
...     "tests/fixtures/file_delimiter_and_quotechar.csv",
...     quotechar="|",
...     delimiter="\t",
...     unquote=False,
... )
...
>>> open(lazy.name, "rb").read()
b'INDEX\tATTR\n0\t|A|\n1\t|B|\n'
>>> list(lazy[:, 1])
[b'|A|', b'|B|']

Numpy

Optional, opt-in numpy support is built into the module. Access to this extended feature set can be had by building the extension from source while setting a LAZYCSV_INCLUDE_NUMPY environment variable to 1. This adds a to_numpy() method to the iterator, which allows iterators to materialize in a 1-dimensional numpy array without creating intermediary PyObject*'s for each field of the CSV file.

Access to this feature requires numpy to be preinstalled as this feature makes numpy a compilation dependency.

$ LAZYCSV_INCLUDE_NUMPY=1 python -m pip install lazycsv

>>> import numpy as np
>>> from lazycsv import lazycsv
>>> lazy = lazycsv.LazyCSV("")
>>> lazy = lazycsv.LazyCSV("./tests/fixtures/file.csv")
>>> lazy.sequence(col=0).to_numpy().astype(np.int8)
array([0, 1], dtype=int8)

Users pinned to an older version of numpy (<1.7) may wish to instead compile using a LAZYCSV_INCLUDE_NUMPY_LEGACY=1 flag, which drops the API pin in the module while still compiling with numpy support.

Benchmarks (CPU)

CPU benchmarks are included below, benchmarked on a Ryzen 7 5800X inside a stock python3.9 docker container.

root@aa9d7c7ffb59:/code# python tests/benchmark_lazy.py
filesize: 0.134gb
cols=10000
rows=10000
sparsity=0.95

benchmarking lazycsv:
indexing lazy... time to index: 0.450414217018988
parsing cols... time to parse: 1.5233540059998631
total time: 1.9737682230188511

benchmarking datatable:
100% |██████████████████████████████████████████████████| Reading data [done]
creating datatables frame... time to object: 0.40828132900060154
parsing cols... time to parse: 3.810204313998838
total time: 4.21848564299944

benchmarking polars (read):
creating polars df... time to object: 2.357821761001105
parsing cols... time to parse: 1.3874979300017003
total time: 3.7453196910028055

root@aa9d7c7ffb59:/code# python tests/benchmark_lazy.py
filesize: 1.387gb
cols=10000
rows=100000
sparsity=0.95

benchmarking lazycsv:
indexing lazy... time to index: 4.298127760004718
parsing cols... time to parse: 18.591125406033825
total time: 22.889253166038543

benchmarking datatable:
100% |██████████████████████████████████████████████████| Reading data [done]
creating datatables frame... time to object: 2.4456441220027045
parsing cols... time to parse: 37.424315700998704
total time: 39.86995982300141

benchmarking polars (read):
creating polars df... time to object: 22.383294907001982
parsing cols... time to parse: 14.16580996599805
total time: 36.54910487300003

filesize: 14.333gb
cols=100000
rows=100000
sparsity=0.95

benchmarking lazycsv:
indexing lazy... time to index: 55.42112316700002
parsing cols... time to parse: 362.268973717
total time: 417.690096884

benchmarking datatable:
 58% |█████████████████████████████▍                    | Reading data Killed

benchmarking polars (read):
Killed