Lector aims to be a fast reader for potentially messy CSV files with configurable column type inference. It combines automatic detection of file encodings, CSV dialects (separator, escaping etc.) and preambles (initial lines containing metadata or junk unrelated to the actual tabular data). Its goal is to just-read-the-effing-CSV file without manual configuration in most cases. Each of the detection components is configurable and can be swapped out easily with custom implementations.
Also, since both pandas and Apache Arrow will destructively cast columns to the wrong type in some cases (e.g. large ID-like integer strings to floats), it provides an alternative and customisable inference and casting mechanism.
Under the hood it uses pyarrow's CSV parser for reading, and its compute functions for optional type inference.
Lector is used at Graphext behind the scenes whenever a user uploads a new dataset, and so implicitly has been validated across 1000s of different CSV files from all kinds of sources. Note, however, that this is Graphext's first foray into open-sourcing our code and still work-in-progress. So at least initially we won't provide any guarantees as to support of this library.
For quick usage examples see the Usage section below or the notebook in this repo.
For detailed documentation visit https://lector.readthedocs.io/.
While this library is not available yet on pypi, you can easily install it from Github with
pip install git+https://github.com/graphext/lector
Let's assume we receive a CSV file containing some initial metadata, using the semicolon as separator, having some missing fields, and being encoded in Latin-1 (you'd be surprised how common such files are in the real world).
Create example CSV file
csv = """
Some preamble content here
This is still "part of the metadata preamble"
id;genre;metric;count;content;website;tags;vecs;date
1234982348728374;a;0.1;1;; http://www.graphext.com;"[a,b,c]";"[1.3, 1.4, 1.67]";11/10/2022
;b;0.12;;"Natural language text is different from categorical data."; https://www.twitter.com;[d];"[0, 1.9423]";01/10/2022
9007199254740993;a;3.14;3;"The Project · Gutenberg » EBook « of Die Fürstin.";http://www.google.com;"['e', 'f']";["84.234, 12509.99"];13/10/2021
""".encode("ISO-8859-1")
with open("example.csv", "wb") as fp:
fp.write(csv)
To read this with lector into a pandas DataFrame, simply use
df = lector.read_csv("example.csv", to_pandas=True)
Printing the DataFrame and its column types produces the following output:
id genre metric count \
0 1234982348728374 a 0.10 1
1 <NA> b 0.12 <NA>
2 9007199254740993 a 3.14 3
content website \
0 <NA> http://www.graphext.com
1 Natural language text is different from catego... https://www.twitter.com
2 The Project · Gutenberg » EBook « of Die Fürstin. http://www.google.com
tags vecs date
0 [a, b, c] [1.3, 1.4, 1.67] 2022-10-11
1 [d] [0.0, 1.9423] 2022-10-01
2 [e, f] [84.234, 12509.99] 2021-10-13
id Int64
genre category
metric float64
count UInt8
content string
website category
tags object
vecs object
date datetime64[ns]
dtype: object
This is pretty sweet, because
- we didn't have to tell lector how to read this file (text encoding, lines to skip, separator etc.)
- we didn't have to tell lector the data types of the columns, but it inferred the correct and most efficient ones automatically, e.g.:
- a nullable
Int64
extension type was necessary to correctly represent values in theid
column - the
category
column was automatically converted to the efficientdictionary
(categorical) type - the
count
column uses the smallest integer type necessary - the
text
column, containing natural language text, has not been converted to a categortical type, but kept as string values (it is unlikely to benefit from dictionary-encoding) - the
date
column was converted to datetime's correctly, even though the original strings are not in an ISO format - the
tags
andvecs
columns have been imported withobject
dtype (since pandas doesn't officially support iterables as elements in a column), but its values are in fact numpy array of the correct dtype!
- a nullable
Neither pandas nor arrow will do this. In fact, they cannot even import this data correctly, without attempting to do any smart type inference. Compare e.g. with pandas attempt to read the same CSV file:
Pandas and Arrow fail
Firstly, to get something close to the above, you'll have to spend a good amount of time manually inspecting the CSV file and come up with the following verbose pandas call:dtypes = {
"id": "Int64",
"genre": "category",
"metric": "float",
"count": "UInt8",
"content": "string",
"website": "category",
"tags": "object",
"vecs": "object"
}
df = pd.read_csv(
fp,
encoding="ISO-8859-1",
skiprows=3,
sep=";",
dtype=dtypes,
parse_dates=["date"],
infer_datetime_format=True
)
While this parses the CSV file alright, the result is, urm, lacking. Let's see:
id genre metric count \
0 1234982348728374 a 0.10 1
1 <NA> b 0.12 <NA>
2 9007199254740992 a 3.14 3
content \
0 <NA>
1 Natural language text is different from catego...
2 The Project · Gutenberg » EBook « of Die Fürstin.
website tags vecs date
0 http://www.graphext.com [a,b,c] [1.3, 1.4, 1.67] 2022-11-10
1 https://www.twitter.com [d] [0, 1.9423] 2022-01-10
2 http://www.google.com ['e', 'f'] ["84.234, 12509.99"] 2021-10-13
id Int64
genre category
metric float64
count UInt8
content string
website category
tags object
vecs object
date datetime64[ns]
dtype: object
A couple of observations:
- Pandas will cast numeric columns with missing data to the float type always, before any of our custom types are applied. This is a big problem, as we can see in the
id
column, since not all integers can be represented exactly by a 64 bit floating type (the correct value in our file is9007199254740993
👀). It is also a sneaky problem, because this happens silently, and so you may not realize you've got wrong IDs, and may produce totally wrong analyses if you use them down the line for joins etc. The only way to import CSV files like this with pandas correctly is to inspect the actual data in a text editor, guess the best data type, import the data without any type inference, and then individually cast to the correct types. There is no way to configure pandas to import the data correctly. - Pandas has messed up the dates. While at least warning us about it, pandas doesn't try to infer a consistent date format across all rows. While the CSV file contains all dates in a single consistent format (
%d/%m/%Y
), pandas has used mixed formats and so imported some dates wrongly. - The
category
andtext
columns have been imported with theobject
dtype, which is not particularly useful, but not necessarily a problem either. - Since pandas doesn't support iterable dtypes, the tags and vecs columns haven't been parsed into any useful structures
Note that Arrow doesn't fare much better. It doesn't parse and infer its own list
data type, it doesn't know how to parse dates in any format other than ISO 8601, and commits the same integer-as-float conversion error.
To install a local copy for development, including all dependencies for test, documentation and code quality, use the following commands:
clone git+https://github.com/graphext/lector
cd lector
pip install -v -e ".[dev]"
pre-commit install
The pre-commit command will make sure that whenever you try to commit changes to this repo code quality and formatting tools will be executed. This ensures e.g. a common coding style, such that any changes to be commited are functional changes only, not changes due to different personal coding style preferences. This in turn makes it either to collaborate via pull requests etc.
To test installation you may execute the pytest suite to make sure everything's setup correctly, e.g.:
pytest -v .
The documentation is created using Sphinx and is available here: https://lector.readthedocs.io/.
You can build and view the static html locally like any other Sphinx project:
(cd docs && make clean html)
(cd docs/build/html && python -m http.server)
- Parallelize type inference? While type inference is already pretty fast, it can potentially be sped up by processing columns in parallel.
- Testing. The current pytest setup is terrible. I've given
hypothesis_csv
a try here, but I'm probably making bad use of it. Tests are convoluted and probably not even good a catching corner cases.
This project is licensed under the terms of the Apache License 2.0.
- Documentation: https://lector.readthedocs.io/
- Source: https://github.com/graphext/lector
- Graphext: https://www.graphext.com
- Graphext on Twitter: https://twitter.com/graphext