casanova

If you often find yourself processing CSV files using python, you will quickly notice that, while being more comfortable, csv.DictReader remains way slower than csv.reader:

# To read a 1.5G CSV file:
csv.reader: 24s
csv.DictReader: 84s
casanova.reader: 25s

casanova is therefore an attempt to stick to csv.reader performance while still keeping a comfortable interface, still able to consider headers (even duplicate ones also, something that csv.DictReader is incapable of) etc.

casanova is thus a good fit for you if you need to:

• Stream large CSV files without running out of memory
• Enrich the same CSV files by outputing a similar file, all while adding, filtering and editing cells.
• Have the possibility to resume said enrichment if your process exited
• Do so in a threadsafe fashion, and be able to resume even if your output does not have the same order as the input

casanova also packs exotic utilities able to read csv files in reverse (without loading the whole file into memory and in regular O(n) time), so you can, for instance, fetch useful information to restart some aborted process.

Installation

You can install casanova with pip with the following command:

pip install casanova

If you want to be able to feed CSV files from the web to casanova readers & enrichers you will also need to install at least urllib3 and optionally certifi (if you want secure SSL). Nnote that a lot of python packages, including the popular requests library, already depend on those two, so it is likely you already have them installed anyway:

# Installing them explicitly
pip install urllib3 certifi

# Installing casanova with those implicitly
pip install casanova[http]

Usage

• reader
• reverse_reader
• headers
• writer
• enricher
• threadsafe_enricher
• batch_enricher
• resumers
  • RowCountResumer
  • ThreadSafeResumer
  • BatchResumer
  • LastCellResumer
  • LastCellComparisonResumer
• count
• last_cell
• namedrecord
• set_defaults
• xsv selection mini DSL

reader

Straightforward CSV reader yielding list rows but giving some information about potential headers and their ipositions.

import casanova

with open('./people.csv') as f:

    # Creating a reader
    reader = casanova.reader(f)

    # Getting header information
    reader.fieldnames
    >>> ['name', 'surname']

    reader.headers
    >>> Headers(name=0, surname=1)

    name_pos = reader.headers.name
    name_pos = reader.headers['name']

    'name' in reader.headers
    >>> True

    # Iterating over the rows
    for row in reader:
        name = row[name_pos] # it's better to cache your pos outside the loop
        name = row[reader.headers.name] # this works, but is slower

    # Interested in a single column?
    for name in reader.cells('name'):
        print(name)

    # Need also the current row when iterating on cells?
    for row, name in reader.cells('name', with_rows=True):
        print(row, name, surname)

    # Want to iterate over records
    # NOTE: this has a performance cost
    for name, surname in reader.records('name', 'surname'):
        print(name, surname)

    for record in reader.records(['name', 'age']):
        print(record[0])

    for record in reader.records({'name': 'name', 'age': 1}):
        print(record['age'])

    # No headers? No problem.
    reader = casanova.reader(f, no_headers=True)

# Note that you can also create a reader from a path
with casanova.reader('./people.csv') as reader:
  ...

# And if you need exotic encodings
with casanova.reader('./people.csv', encoding='latin1') as reader:
  ...

# The reader will also handle gzipped files out of the box
with casanova.reader('./people.csv.gz') as reader:
  ...

# If you have `urllib3` installed, casanova is also able to stream
# remote CSV file out of the box
with casanova.reader('https://mydomain.fr/some-file.csv') as reader:
    ...

# The reader will also accept iterables of rows
rows = [['name', 'surname'], ['John', 'Moran']]
reader = casanova.reader(rows)

# And you can of course use the typical dialect-related kwargs
reader = casanova.reader('./french-semicolons.csv', delimiter=';')

# Readers can also be closed if you want to avoid context managers
reader.close()

Arguments

• input_file str or Path or file or Iterable[list[str]]: input file given to the reader. Can be a path that will be opened for you by the reader, a file handle or even an arbitrary iterable of list rows.
• no_headers bool, optional [False]: set to True if input_file has no headers.
• encoding str, optional [utf-8]: encoding to use to open the file if input_file is a path.
• dialect str or csv.Dialect, optional: CSV dialect for the reader to use. Check python standard csv module documentation for more info.
• quotechar str, optional: quote character used by CSV parser.
• delimiter str, optional: delimiter characted used by CSV parser.
• prebuffer_bytes int, optional: number of bytes of input file to prebuffer in attempt to get a total number of lines ahead of time.
• total int, optional: total number of lines to expect in file, if you already know it ahead of time. If given, the reader won't prebuffer data even if prebuffer_bytes was set.
• multiplex casanova.Multiplexer, optional: multiplexer to use. Read this for more information.
• strip_null_bytes_on_read bool, optional [False]: before python 3.11, the csv module will raise when attempting to read a CSV file containing null bytes. If set to True, the reader will strip null bytes on the fly while parsing rows.
• reverse bool, optional [False]: whether to read the file in reverse (except for the header of course).

Properties

• total int, optional: total number of lines in the file, if known through prebuffering or through the total kwarg.
• headers casanova.Headers, optional, optional: CSV file headers if no_headers=False.
• empty bool: whether the given file was empty.
• fieldnames list[str], optional: list representing the CSV file headers if no_headers=False.
• row_len int: expected number of items per row.

Methods

• rows: returns an iterator over the reader rows. Same as iterating over the reader directly.
• cells: take the name of a column or its position and returns an iterator over values of the given column. Can be given with_rows=True if you want to iterate over a value, row tuple instead if required.
• enumerate: resuming-safe enumeration over rows yielding index, row tuples. Takes an optional start kwarg like builtin enumerate.
• enumerate_cells: resuming-safe enumeration over cells yielding index, cell or index, row, cell if given with_rows=True. Takes an optional start kwarg like builtin enumerate.
• wrap: method taking a list row and returning a RowWrapper object to wrap it.
• close: cleans up the reader resources manually when not using the dedicated context manager. It is usually only useful when the reader was given a path and not an already opened file handle.

Multiplexing

Sometimes, one column of your CSV file might contain multiple values, separated by an arbitrary separator character such as |.

In this case, it might be desirable to "multiplex" the file by making a reader emit one copy of the line with each of the values contained by a cell.

To do so, casanova exposes a special Multiplexer object you can give to any reader like so:

import casanova

# Most simple case: a column named "colors", separated by "|"
reader = casanova.reader(
    input_file,
    multiplex=casanova.Multiplexer('colors')
)

# Customizing the separator:
reader = casanova.reader(
    input_file,
    multiplex=casanova.Multiplexer('colors', separator='$')
)

# Renaming the column on the fly:
reader = casanova.reader(
    input_file,
    multiplex=casanova.Multiplexer('colors', new_column='color')
)

reverse_reader

A reverse CSV reader might sound silly, but it can be useful in some scenarios. Especially when you need to read the last line from an output file without reading the whole thing first, in constant time.

It is mostly used by casanova resumers and it is unlikely you will need to use them on your own.

import casanova

# people.csv looks like this
# name,surname
# John,Doe,
# Mary,Albert
# Quentin,Gold

with open('./people.csv', 'rb') as f:
    reader = casanova.reverse_reader(f)

    reader.fieldnames
    >>> ['name', 'surname']

    next(reader)
    >>> ['Quentin', 'Gold']

headers

A class representing the headers of a CSV file. It is useful to find the row position of some columns and perform complex selection.

import casanova

# Headers can be instantiated thusly
headers = casanova.headers(['name', 'surname', 'age'])

# But you will usually use a reader or an enricher's one:
headers = casanova.reader(input_file).headers

# Accessing a column through attributes
headers.surname
>>> 1

# Accessing a column by indexing:
headers['surname']
>>> 1

# Getting a column
headers.get('surname')
>>> 1
headers.get('not-found')
>>> None

# Getting a duplicated column name
casanova.headers(['surname', 'name', 'name'])['name', 1]
>>> 2

casanova.headers(['surname', 'name', 'name']).get('name', index=1)
>>> 2

# Asking if a column exists:
'name' in headers:
>>> True

# Retrieving fieldnames:
headers.fieldnames
>>> ['name', 'surname', 'age']

# Iterating over headers
for col in headers:
    print(col)

# Couting columns:
len(headers)
>>> 3

# Retrieving the nth header:
headers.nth(1)
>>> 'surname'

# Wraping a row
headers.wrap(['John', 'Matthews', '45'])
>>> RowWrapper(name='John', surname='Matthews', age='45')

# Selecting some columns (by name and/or index)):
headers.select(['name', 2])
>>> [0, 2]

# Selecting using xsv mini DSL:
headers.select('name,age')
>>> [0, 2]
headers.select('!name')
>>> [1, 2]

For more info about xsv mini DSL, check this part of the documentation.

writer

casanova also exports a csv writer. It can automatically write headers when needed and is able to resume some tasks.

import casanova

with open('output.csv') as f:
    writer = casanova.writer(f, fieldnames=['name', 'surname'])

    writer.writerow(['John', 'Davis'])

    # If you want to write headers yourself:
    writer = casanova.writer(f, fieldnames=['name', 'surname'], write_header=False)

    writer.writeheader()

Arguments

• output_file file or casanova.Resumer: target file.
• fieldnames Iterable[str], optional: column names.
• strip_null_bytes_on_write bool, optional [False]: whether to strip null bytes when writing rows. Note that on python 3.10, there is a bug that prevents a csv.writer will raise an error when attempting to write a row containing a null byte.
• dialect csv.Dialect or str, optional: dialect to use to write CSV.
• delimiter str, optional: CSV delimiter.
• quotechar str, optional: CSV quoting character.
• quoting csv.QUOTE_*, optional: CSV quoting strategy.
• escapechar str, optional: CSV escaping character.
• lineterminator str, optional: CSV line terminator.
• write_header bool, optional [True]: whether to automatically write header if required (takes resuming into account).

Properties

• headers casanova.Headers, optional, optional: CSV file headers if fieldnames were provided
• fieldnames list[str], optional: provided fieldnames.

Resuming

A casanova.writer is able to resume through a LastCellResumer.

enricher

casanova enrichers are basically a smart combination of both a reader and a writer.

It can be used to transform a given CSV file. This means you can transform its values on the fly, select some columns to keep from input and add new ones very easily.

Note that enrichers inherits from both casanova.reader and casanova.writer and therefore keep both their properties and methods.

import casanova

with open('./people.csv') as input_file, \
     open('./enriched-people.csv', 'w') as output_file:

    enricher = casanova.enricher(input_file, output_file)

    # The enricher inherits from casanova.reader
    enricher.fieldnames
    >>> ['name', 'surname']

    # You can iterate over its rows
    name_pos = enricher.headers.name

    for row in enricher:

        # Editing a cell, so that everyone is called John now
        row[name_pos] = 'John'
        enricher.writerow(row)

    # Want to add columns?
    enricher = casanova.enricher(f, of, add=['age', 'hair'])

    for row in enricher:
        enricher.writerow(row, ['34', 'blond'])

    # Want to keep only some columns from input?
    enricher = casanova.enricher(f, of, add=['age'], select=['surname'])

    for row in enricher:
        enricher.writerow(row, ['45'])

    # Want to select columns to keep using xsv mini dsl?
    enricher = casanova.enricher(f, of, select='!1-4')

    # You can of course still use #.cells etc.
    for row, name in enricher.cells('name', with_rows=True):
        print(row, name)

Arguments

• input_file file or str: file object to read or path to open.
• output_file file or Resumer: file object to write.
• no_headers bool, optional [False]: set to True if input_file has no headers.
• encoding str, optional [utf-8]: encoding to use to open the file if input_file is a path.
• add Iterable[str|int], optional: names of columns to add to output.
• select Iterable[str|int]|str, optional: selection of columns to keep from input. Can be an iterable of column names and/or column positions or a selection string writting in xsv mini DSL.
• dialect str or csv.Dialect, optional: CSV dialect for the reader to use. Check python standard csv module documentation for more info.
• quotechar str, optional: quote character used by CSV parser.
• delimiter str, optional: delimiter characted used by CSV parser.
• prebuffer_bytes int, optional: number of bytes of input file to prebuffer in attempt to get a total number of lines ahead of time.
• total int, optional: total number of lines to expect in file, if you already know it ahead of time. If given, the reader won't prebuffer data even if prebuffer_bytes was set.
• multiplex casanova.Multiplexer, optional: multiplexer to use. Read this for more information.
• reverse bool, optional [False]: whether to read the file in reverse (except for the header of course).
• strip_null_bytes_on_read bool, optional [False]: before python 3.11, the csv module will raise when attempting to read a CSV file containing null bytes. If set to True, the reader will strip null bytes on the fly while parsing rows.
• strip_null_bytes_on_write bool, optional [False]: whether to strip null bytes when writing rows. Note that on python 3.10, there is a bug that prevents a csv.writer will raise an error when attempting to write a row containing a null byte.
• writer_dialect csv.Dialect or str, optional: dialect to use to write CSV.
• writer_delimiter str, optional: CSV delimiter for writer.
• writer_quotechar str, optional: CSV quoting character for writer.
• writer_quoting csv.QUOTE_*, optional: CSV quoting strategy for writer.
• writer_escapechar str, optional: CSV escaping character for writer.
• writer_lineterminator str, optional: CSV line terminator for writer.
• write_header bool, optional [True]: whether to automatically write header if required (takes resuming into account).

Properties

• total int, optional: total number of lines in the file, if known through prebuffering or through the total kwarg.
• headers casanova.Headers, optional, optional: CSV file headers if no_headers=False.
• empty bool: whether the given file was empty.
• fieldnames list[str], optional: list representing the CSV file headers if no_headers=False.
• row_len int: expected number of items per row.
• output_headers casanova.Headers, optional, optional: output CSV headers if no_headers=False.
• output_fieldnames list[str], optional: list representing the output CSV headers if no_headers=False.
• added_count int: number of columns added to the output.

Resuming

A casanova.enricher is able to resume through a RowCountResumer or a LastCellComparisonResumer.

threadsafe_enricher

Sometimes, you might want to process multiple input rows concurrently. This can mean that you will emit rows in an arbitrary order, different from the input one.

This is fine, of course, but if you still want to be able to resume an aborted process efficiently (using the `ThreadSafeResumer), your output will need specific additions for it to work, namely a column containing the index of an output row in the original input.

casanova.threadsafe_enricher makes it simpler by providing a tailored writerow method and iterators always provided the index of a row safely.

Note that such resuming is only possible if one row in the input will produce exactly one row in the output.

import casanova

with open('./people.csv') as f, \
     open('./enriched-people.csv', 'w') as of:

    enricher = casanova.threadsafe_enricher(f, of, add=['age', 'hair'])

    for index, row in enricher:
        enricher.writerow(index, row, ['67', 'blond'])

    for index, value in enricher.cells('name'):
        ...

    for index, row, value in enricher.cells('name', with_rows=True):
        ...

Arguments

Everything from casanova.enricher plus:

• index_column str, optional [index]: name of the automatically added index column.

Resuming

A casanova.threadsafe_enricher is able to resume through a ThreadSafeResumer.

batch_enricher

Sometimes, you might want to process a CSV file and paginate API calls per row. This means that each row of your input file should produce multiple new lines, which will be written in batch each time one call from the API returns.

Sometimes, the pagination might be quite long (think collecting the Twitter followers of a very popular account), and it would not be a good idea to accumulate all the results for a single row before flushing them to file atomically because if something goes wrong, you will lose a lot of work.

But if you still want to be able to resume if process is aborted, you will need to add some things to your output. Namely, a column containing optional "cursor" data to resume your API calls and an "end" symbol indicating we finished the current input row.

import casanova

with open('./twitter-users.csv') as input_file, \
     casanova.BatchResumer('./output.csv') as output_file:

    enricher = casanova.batch_resumer(input_file, output_file)

    for row in enricher:
        for results, next_cursor in paginate_api_calls(row):

            # NOTE: if we reached the end, next_cursor is None
            enricher.writebatch(row, results, next_cursor)

Arguments

Everything from casanova.enricher plus:

• cursor_column str, optional [cursor]: name of the cursor column to add.
• end_symbol str, optional [end]: unambiguous (from cursor) end symbol to mark end of input row processing.

Resuming

A casanova.batch_enricher is able to resume through a BatchResumer.

resumers

Through handy Resumer classes, casanova lets its enrichers and writers resume an aborted process.

Those classes must be used as a wrapper to open the output file and can assess whether resuming is actually useful or not for you.

All resumers act like file handles, can be used as a context manager using the with keyword and can be manually closed using the close method if required.

Finally know that resumers should work perfectly well with multiplexing

RowCountResumer

The RowCountResumer works by counting the number of line of the output and skipping that many lines from the input.

It can only work in 1-to-1 scenarios where you only emit a single row per input row.

It works in O(2n) => O(n) time and O(1) memory, n being the number of already processed rows.

It is only supported by casanova.enricher.

import casanova

with open('input.csv') as input_file, \
     casanova.RowCountResumer('output.csv') as resumer:

    # Want to know if we can resume?
    resumer.can_resume()

    # Want to know how many rows were already done?
    resumer.already_done_count()

    # Giving the resumer to an enricher as if it was the output file
    enricher = casanova.enricher(input_file, resumer)

ThreadSafeResumer

casanova exports a threadsafe resumer that allows row to be processed concurrently and emitted in a different order.

In this precise case, couting the rows is not enough and we need to be smarter.

One way to proceed is to leverage the index column added by the threadsafe enricher to compute a set of already processed row while reading the output. Then we can just skip the input rows whose indices are in this set.

The issue here is that this consumes up to O(n) memory, which is prohibitive in some use cases.

To make sure this still can be done while consuming very little memory, casanova uses an exotic data structure we named a "contiguous range set".

This means we can resume operation in O(n + log(h) * n)) => O(log(h) * n) time and O(log(h)) memory, n being the number of already processed rows and h being the size of the largest hole in the sorted indices of those same rows. Note that most of the time h << n since the output is mostly sorted (albeit not at a local level).

You can read more about this data structure in this blog post.

Note finally this resumer can only work in 1-to-1 scenarios where you only emit a single row per input row.

It is supported by casanova.threadsafe_enricher only.

import casanova

with open('input.csv') as input_file, \
     casanova.ThreadSafeResumer('output.csv') as resumer:

    # Want to know if we can resume?
    resumer.can_resume()

    # Want to know how many rows were already done?
    resumer.already_done_count()

    # Giving the resumer to an enricher as if it was the output file
    enricher = casanova.threadsafe_enricher(input_file, resumer)

# If you want to use casanova ContiguousRangeSet for whatever reason
from casanova import ContiguousRangeSet

BatchResumer

todo...

LastCellResumer

Sometimes you might write an output CSV file while performing some paginated action. Said action could be aborted and you might want to resume it where you left it.

The LastCellResumer therefore enables you to resume writing a CSV file by reading its output's last row using a casanova.reverse_reader and extracting the value you need to resume in constant time and memory.

It is only supported by casanova.writer.

import casanova

with casanova.LastCellResumer('output.csv', value_column='user_id') as resumer:

    # Giving the resumer to a writer as if it was the output file
    writer = casanova.writer(resumer)

    # Extracting last relevant value if any, so we can properly resume
    last_value = resumer.get_state()

LastCellComparisonResumer

In some scenarios, it is possible to resume the operation of an enricher if you can know what was the last value of some column emitted in the output.

Fortunately, using casanova.reverse_reader, one can read the last line of a CSV file in constant time.

As such the LastCellComparisonResumer enables you to resume the work of an enricher in O(n) time and O(1) memory, with n being the number of already done lines that you must quickly skip when repositioning yourself in the input.

Note that it only works when the enricher emits a single line per line in the input and when the considered column value is unique across the input file.

It is only supported by casanova.enricher.

import casanova

with open('input.csv') as input_file, \
     casanova.LastCellComparisonResumer('output.csv', value_colum='user_id') as resumer:

    # Giving the resumer to an enricher as if it was the output file
    enricher = casanova.enricher(input_file, resumer)

count

casanova exposes a helper function that one can use to quickly count the number of lines in a CSV file.

import casanova

count = casanova.count('./people.csv')

# You can also stop reading the file if you go beyond a number of rows
count = casanova.count('./people.csv', max_rows=100)
>>> None # if the file has more than 100 rows
>>> 34   # else the actual count

# Any additional kwarg will be passed to the underlying reader as-is
count = casanova.count('./people.csv', delimiter=';')

last_cell

casanova exposes a helper function using a reverse_reader to read only the last cell value from a given column of a CSV file.

import casanova

last_cell = casanova.last_cell('./people.csv', column='name')
>>> 'Quentin'

# Will return None if the file is empty
last_cell = casanova.last_cell('./empty.csv', column='name')
>>> None

# Any additional kwarg will be passed to the underlying reader as-is
last_cell = casanova.last_cell('./people.csv', column='name', delimiter=';')

namedrecord

casanova's namedrecord is basically an enhanced & CSV-aware version of python namedtuple.

from casanova import namedrecord

Record = namedrecord(
    'Record',
    ['title', 'urls', 'is_accessible', 'data'],
    defaults=[True, None]
)

Record.fieldnames
>>> ['title', 'urls', 'is_accessible']

example = Record('Le Monde', ['https://lemonde.fr', 'https://www.lemonde.fr'])

# It works exactly like a namedtuple would, but with perks:
example
>>> Record(title='Le Monde', urls=['https://lemonde.fr', 'https://www.lemonde.fr'], is_accessible=True)

# You can read by index:
example[0]
>>> 'Le Monde'

# You can read its attributes:
example.title
>>> 'Le Monde'

# You can access it like a dict:
example['title']
>>> 'Le Monde'

# You can use #.get:
example.get('what?')
>>> None

# You can return it as a plain dict:
example.as_dict()
>>> {
    'title': 'Le Monde',
    ...
}

# You can format it as a CSV row (notice the list serialization):
example.as_csv_row():
>>> ['Le Monde', 'https://lemonde.fr|https://www.lemonde.fr', 'true']

# You can tweak formatting if needed:
example.as_csv_row(
    plural_separator='$',
    none_value='null',
    true_value='yes',
    false_value='',
)
>>> ['Le Monde', 'https://lemonde.fr$https://www.lemonde.fr', 'yes']

# You can also set default formatting options at the record level
Record = namedrecord(
    'Record',
    ['title'],
    plural_separator='$',
    ...
)

# You can format as a CSV dict row (suitable for csv.DictWriter, if required):
example.as_csv_dict_row():
>>> {
    'title': 'Le Monde',
    'urls': 'https://lemonde.fr|https://www.lemonde.fr',
    'is_accessible': 'true'
}

# You can also embed json data if you feel crazy:
Record = namedrecord(
    'Record',
    ['title', 'data'],
    json=['data']
)

example = Record('With JSON', {'hello': 'world'})

example.as_csv_row()
>>> ['With JSON', '{"hello": "world"}']

set_defaults

casanova.set_defaults lets you edit global defaults for casanova:

import casanova

casanova.set_defaults(strip_null_bytes_on_read=True)

# As a context manager:
with casanova.temporary_defaults(strip_null_bytes_on_read=True):
    ...

Arguments

• strip_null_bytes_on_read bool, optional [False]: should readers and enrichers strip null bytes on read?
• strip_null_bytes_on_write bool, optional [False]: should writers and enrichers strip null bytes on write?
• prebuffer_bytes int, optional: default prebuffer bytes for readers and enrichers.

xsv selection mini DSL

xsv, a command line tool written in Rust to handle csv files, uses a clever mini DSL to let users specify column selections.

casanova has a working python implementation of this mini DSL that can be used by the headers.select method and the enrichers select kwargs.

Here is the gist of it (copied right from xsv documentation itself):

Select one column by name:
    * name

Select one column by index (1-based):
    * 2

Select the first and fourth columns:
    * 1,4

Select the first 4 columns (by index and by name):
    * 1-4
    * Header1-Header4

Ignore the first 2 columns (by range and by omission):
    * 3-
    * '!1-2'

Select the third column named 'Foo':
    * 'Foo[2]'

Re-order and duplicate columns arbitrarily:
    * 3-1,Header3-Header1,Header1,Foo[2],Header1

Quote column names that conflict with selector syntax:
    * '"Date - Opening","Date - Actual Closing"'