March 8, 2019

2880 words 14 mins read

EntilZha/PyFunctional

EntilZha/PyFunctional

Python library for creating data pipelines with chain functional programming

repo name EntilZha/PyFunctional
repo link https://github.com/EntilZha/PyFunctional
homepage http://www.pyfunctional.org
language Python
size (curr.) 660 kB
stars (curr.) 1631
created 2015-02-05
license MIT License

PyFunctional

TravisCI Coverage by codecov.io ReadTheDocs PyPI version Gitter

Features

PyFunctional makes creating data pipelines easy by using chained functional operators. Here are a few examples of what it can do:

  • Chained operators: seq(1, 2, 3).map(lambda x: x * 2).reduce(lambda x, y: x + y)
  • Expressive and feature complete API
  • Read and write text, csv, json, jsonl, sqlite, gzip, bz2, and lzma/xz files
  • Parallelize “embarrassingly parallel” operations like map easily
  • Complete documentation, rigorous unit test suite, 100% test coverage, and CI which provide robustness

PyFunctional’s API takes inspiration from Scala collections, Apache Spark RDDs, and Microsoft LINQ.

Table of Contents

  1. Installation
  2. Examples
    1. Simple Example
    2. Aggregates and Joins
    3. Reading and Writing SQLite3
    4. Data Interchange with Pandas
  3. Writing to Files
  4. Parallel Execution
  5. Github Shortform Documentation
    1. Streams, Transformations, and Actions
    2. Streams API
    3. Transformations and Actions APIs
    4. Lazy Execution
  6. Contributing and Bug Fixes
  7. Changelog

Installation

PyFunctional is available on pypi and can be installed by running:

# Install from command line
$ pip install pyfunctional

Then in python run: from functional import seq

Examples

PyFunctional is useful for many tasks, and can natively open several common file types. Here are a few examples of what you can do.

Simple Example

from functional import seq

seq(1, 2, 3, 4)\
    .map(lambda x: x * 2)\
    .filter(lambda x: x > 4)\
    .reduce(lambda x, y: x + y)
# 14

# or if you don't like backslash continuation
(seq(1, 2, 3, 4)
    .map(lambda x: x * 2)
    .filter(lambda x: x > 4)
    .reduce(lambda x, y: x + y)
)
# 14

Streams, Transformations and Actions

PyFunctional has three types of functions:

  1. Streams: read data for use by the collections API.
  2. Transformations: transform data from streams with functions such as map, flat_map, and filter
  3. Actions: These cause a series of transformations to evaluate to a concrete value. to_list, reduce, and to_dict are examples of actions.

In the expression seq(1, 2, 3).map(lambda x: x * 2).reduce(lambda x, y: x + y), seq is the stream, map is the transformation, and reduce is the action.

Filtering a list of account transactions

from functional import seq
from collections import namedtuple

Transaction = namedtuple('Transaction', 'reason amount')
transactions = [
    Transaction('github', 7),
    Transaction('food', 10),
    Transaction('coffee', 5),
    Transaction('digitalocean', 5),
    Transaction('food', 5),
    Transaction('riotgames', 25),
    Transaction('food', 10),
    Transaction('amazon', 200),
    Transaction('paycheck', -1000)
]

# Using the Scala/Spark inspired APIs
food_cost = seq(transactions)\
    .filter(lambda x: x.reason == 'food')\
    .map(lambda x: x.amount).sum()

# Using the LINQ inspired APIs
food_cost = seq(transactions)\
    .where(lambda x: x.reason == 'food')\
    .select(lambda x: x.amount).sum()

# Using PyFunctional with fn
from fn import _
food_cost = seq(transactions).filter(_.reason == 'food').map(_.amount).sum()

Aggregates and Joins

The account transactions example could be done easily in pure python using list comprehensions. To show some of the things PyFunctional excels at, take a look at a couple of word count examples.

words = 'I dont want to believe I want to know'.split(' ')
seq(words).map(lambda word: (word, 1)).reduce_by_key(lambda x, y: x + y)
# [('dont', 1), ('I', 2), ('to', 2), ('know', 1), ('want', 2), ('believe', 1)]

In the next example we have chat logs formatted in json lines (jsonl) which contain messages and metadata. A typical jsonl file will have one valid json on each line of a file. Below are a few lines out of examples/chat_logs.jsonl.

{"message":"hello anyone there?","date":"10/09","user":"bob"}
{"message":"need some help with a program","date":"10/09","user":"bob"}
{"message":"sure thing. What do you need help with?","date":"10/09","user":"dave"}
from operator import add
import re
messages = seq.jsonl('examples/chat_logs.jsonl')

# Split words on space and normalize before doing word count
def extract_words(message):
    return re.sub('[^0-9a-z ]+', '', message.lower()).split(' ')


word_counts = messages\
    .map(lambda log: extract_words(log['message']))\
    .flatten().map(lambda word: (word, 1))\
    .reduce_by_key(add).order_by(lambda x: x[1])

Next, lets continue that example but introduce a json database of users from examples/users.json. In the previous example we showed how PyFunctional can do word counts, in the next example lets show how PyFunctional can join different data sources.

# First read the json file
users = seq.json('examples/users.json')
#[('sarah',{'date_created':'08/08','news_email':True,'email':'sarah@gmail.com'}),...]

email_domains = users.map(lambda u: u[1]['email'].split('@')[1]).distinct()
# ['yahoo.com', 'python.org', 'gmail.com']

# Join users with their messages
message_tuples = messages.group_by(lambda m: m['user'])
data = users.inner_join(message_tuples)
# [('sarah',
#    (
#      {'date_created':'08/08','news_email':True,'email':'sarah@gmail.com'},
#      [{'date':'10/10','message':'what is a...','user':'sarah'}...]
#    )
#  ),...]

# From here you can imagine doing more complex analysis

CSV, Aggregate Functions, and Set functions

In examples/camping_purchases.csv there are a list of camping purchases. Lets do some cost analysis and compare it the required camping gear list stored in examples/gear_list.txt.

purchases = seq.csv('examples/camping_purchases.csv')
total_cost = purchases.select(lambda row: int(row[2])).sum()
# 1275

most_expensive_item = purchases.max_by(lambda row: int(row[2]))
# ['4', 'sleeping bag', ' 350']

purchased_list = purchases.select(lambda row: row[1])
gear_list = seq.open('examples/gear_list.txt').map(lambda row: row.strip())
missing_gear = gear_list.difference(purchased_list)
# ['water bottle','gas','toilet paper','lighter','spoons','sleeping pad',...]

In addition to the aggregate functions shown above (sum and max_by) there are many more. Similarly, there are several more set like functions in addition to difference.

Reading/Writing SQLite3

PyFunctional can read and write to SQLite3 database files. In the example below, users are read from examples/users.db which stores them as rows with columns id:Int and name:String.

db_path = 'examples/users.db'
users = seq.sqlite3(db_path, 'select * from user').to_list()
# [(1, 'Tom'), (2, 'Jack'), (3, 'Jane'), (4, 'Stephan')]]

sorted_users = seq.sqlite3(db_path, 'select * from user order by name').to_list()
# [(2, 'Jack'), (3, 'Jane'), (4, 'Stephan'), (1, 'Tom')]

Writing to a SQLite3 database is similarly easy

import sqlite3
from collections import namedtuple

with sqlite3.connect(':memory:') as conn:
    conn.execute('CREATE TABLE user (id INT, name TEXT)')
    conn.commit()
    User = namedtuple('User', 'id name')
    
    # Write using a specific query
    seq([(1, 'pedro'), (2, 'fritz')]).to_sqlite3(conn, 'INSERT INTO user (id, name) VALUES (?, ?)')
    
    # Write by inserting values positionally from a tuple/list into named table
    seq([(3, 'sam'), (4, 'stan')]).to_sqlite3(conn, 'user')
    
    # Write by inferring schema from namedtuple
    seq([User(name='tom', id=5), User(name='keiga', id=6)]).to_sqlite3(conn, 'user')
    
    # Write by inferring schema from dict
    seq([dict(name='david', id=7), dict(name='jordan', id=8)]).to_sqlite3(conn, 'user')
    
    # Read everything back to make sure it wrote correctly
    print(list(conn.execute('SELECT * FROM user')))
    
    # [(1, 'pedro'), (2, 'fritz'), (3, 'sam'), (4, 'stan'), (5, 'tom'), (6, 'keiga'), (7, 'david'), (8, 'jordan')]

Writing to files

Just as PyFunctional can read from csv, json, jsonl, sqlite3, and text files, it can also write them. For complete API documentation see the collections API table or the official docs.

Compressed Files

PyFunctional will auto-detect files compressed with gzip, lzma/xz, and bz2. This is done by examining the first several bytes of the file to determine if it is compressed so therefore requires no code changes to work.

To write compressed files, every to_ function has a parameter compression which can be set to the default None for no compression, gzip or gz for gzip compression, lzma or xz for lzma compression, and bz2 for bz2 compression.

Parallel Execution

The only change required to enable parallelism is to import from functional import pseq instead of from functional import seq and use pseq where you would use seq. The following operations are run in parallel with more to be implemented in a future release:

  • map/select
  • filter/filter_not/where
  • flat_map

Parallelization uses python multiprocessing and squashes chains of embarrassingly parallel operations to reduce overhead costs. For example, a sequence of maps and filters would be executed all at once rather than in multiple loops using multiprocessing

Documentation

Shortform documentation is below and full documentation is at docs.pyfunctional.org.

Streams API

All of PyFunctional streams can be accessed through the seq object. The primary way to create a stream is by calling seq with an iterable. The seq callable is smart and is able to accept multiple types of parameters as shown in the examples below.

# Passing a list
seq([1, 1, 2, 3]).to_set()
# [1, 2, 3]

# Passing direct arguments
seq(1, 1, 2, 3).map(lambda x: x).to_list()
# [1, 1, 2, 3]

# Passing a single value
seq(1).map(lambda x: -x).to_list()
# [-1]

seq also provides entry to other streams as attribute functions as shown below.

# number range
seq.range(10)

# text file
seq.open('filepath')

# json file
seq.json('filepath')

# jsonl file
seq.jsonl('filepath')

# csv file
seq.csv('filepath')
seq.csv_dict_reader('filepath')

# sqlite3 db and sql query
seq.sqlite3('filepath', 'select * from data')

For more information on the parameters that these functions can take, reference the streams documentation

Transformations and Actions APIs

Below is the complete list of functions which can be called on a stream object from seq. For complete documentation reference transformation and actions API.

Function Description Type
map(func)/select(func) Maps func onto elements of sequence transformation
starmap(func)/smap(func) Apply func to sequence with itertools.starmap transformation
filter(func)/where(func) Filters elements of sequence to only those where func(element) is True transformation
filter_not(func) Filters elements of sequence to only those where func(element) is False transformation
flatten() Flattens sequence of lists to a single sequence transformation
flat_map(func) func must return an iterable. Maps func to each element, then merges the result to one flat sequence transformation
group_by(func) Groups sequence into (key, value) pairs where key=func(element) and value is from the original sequence transformation
group_by_key() Groups sequence of (key, value) pairs by key transformation
reduce_by_key(func) Reduces list of (key, value) pairs using func transformation
count_by_key() Counts occurrences of each key in list of (key, value) pairs transformation
count_by_value() Counts occurrence of each value in a list transformation
union(other) Union of unique elements in sequence and other transformation
intersection(other) Intersection of unique elements in sequence and other transformation
difference(other) New sequence with unique elements present in sequence but not in other transformation
symmetric_difference(other) New sequence with unique elements present in sequence or other, but not both transformation
distinct() Returns distinct elements of sequence. Elements must be hashable transformation
distinct_by(func) Returns distinct elements of sequence using func as a key transformation
drop(n) Drop the first n elements of the sequence transformation
drop_right(n) Drop the last n elements of the sequence transformation
drop_while(func) Drop elements while func evaluates to True, then returns the rest transformation
take(n) Returns sequence of first n elements transformation
take_while(func) Take elements while func evaluates to True, then drops the rest transformation
init() Returns sequence without the last element transformation
tail() Returns sequence without the first element transformation
inits() Returns consecutive inits of sequence transformation
tails() Returns consecutive tails of sequence transformation
zip(other) Zips the sequence with other transformation
zip_with_index(start=0) Zips the sequence with the index starting at start on the right side transformation
enumerate(start=0) Zips the sequence with the index starting at start on the left side transformation
cartesian(*iterables, repeat=1) Returns cartesian product from itertools.product transformation
inner_join(other) Returns inner join of sequence with other. Must be a sequence of (key, value) pairs transformation
outer_join(other) Returns outer join of sequence with other. Must be a sequence of (key, value) pairs transformation
left_join(other) Returns left join of sequence with other. Must be a sequence of (key, value) pairs transformation
right_join(other) Returns right join of sequence with other. Must be a sequence of (key, value) pairs transformation
join(other, join_type='inner') Returns join of sequence with other as specified by join_type. Must be a sequence of (key, value) pairs transformation
partition(func) Partitions the sequence into elements which satisfy func(element) and those that don’t transformation
grouped(size) Partitions the elements into groups of size size transformation
sorted(key=None, reverse=False)/order_by(func) Returns elements sorted according to python sorted transformation
reverse() Returns the reversed sequence transformation
slice(start, until) Sequence starting at start and including elements up to until transformation
head() / first() Returns first element in sequence action
head_option() Returns first element in sequence or None if its empty action
last() Returns last element in sequence action
last_option() Returns last element in sequence or None if its empty action
len() / size() Returns length of sequence action
count(func) Returns count of elements in sequence where func(element) is True action
empty() Returns True if the sequence has zero length action
non_empty() Returns True if sequence has non-zero length action
all() Returns True if all elements in sequence are truthy action
exists(func) Returns True if func(element) for any element in the sequence is True action
for_all(func) Returns True if func(element) is True for all elements in the sequence action
find(func) Returns the element that first evaluates func(element) to True action
any() Returns True if any element in sequence is truthy action
max() Returns maximal element in sequence action
min() Returns minimal element in sequence action
max_by(func) Returns element with maximal value func(element) action
min_by(func) Returns element with minimal value func(element) action
sum()/sum(projection) Returns the sum of elements possibly using a projection action
product()/product(projection) Returns the product of elements possibly using a projection action
average()/average(projection) Returns the average of elements possibly using a projection action
aggregate(func)/aggregate(seed, func)/aggregate(seed, func, result_map) Aggregate using func starting with seed or first element of list then apply result_map to the result action
fold_left(zero_value, func) Reduces element from left to right using func and initial value zero_value action
fold_right(zero_value, func) Reduces element from right to left using func and initial value zero_value action
make_string(separator) Returns string with separator between each str(element) action
dict(default=None) / to_dict(default=None) Converts a sequence of (Key, Value) pairs to a dictionary. If default is not None, it must be a value or zero argument callable which will be used to create a collections.defaultdict action
list() / to_list() Converts sequence to a list action
set() / to_set() Converts sequence to a set action
to_file(path) Saves the sequence to a file at path with each element on a newline action
to_csv(path) Saves the sequence to a csv file at path with each element representing a row action
to_jsonl(path) Saves the sequence to a jsonl file with each element being transformed to json and printed to a new line action
to_json(path) Saves the sequence to a json file. The contents depend on if the json root is an array or dictionary action
to_sqlite3(conn, tablename_or_query, *args, **kwargs) Save the sequence to a SQLite3 db. The target table must be created in advance. action
to_pandas(columns=None) Converts the sequence to a pandas DataFrame action
cache() Forces evaluation of sequence immediately and caches the result action
for_each(func) Executes func on each element of the sequence action

Lazy Execution

Whenever possible, PyFunctional will compute lazily. This is accomplished by tracking the list of transformations that have been applied to the sequence and only evaluating them when an action is called. In PyFunctional this is called tracking lineage. This is also responsible for the ability for PyFunctional to cache results of computation to prevent expensive re-computation. This is predominantly done to preserve sensible behavior and used sparingly. For example, calling size() will cache the underlying sequence. If this was not done and the input was an iterator, then further calls would operate on an expired iterator since it was used to compute the length. Similarly, repr also caches since it is most often used during interactive sessions where its undesirable to keep recomputing the same value. Below are some examples of inspecting lineage.

def times_2(x):
    print(x)
    return 2 * x
elements = seq(1, 1, 2, 3, 4).map(times_2).distinct()
elements._lineage
# Lineage: sequence -> map(times_2) -> distinct

l_elements = elements.to_list()
# Prints: 1
# Prints: 1
# Prints: 2
# Prints: 3
# Prints: 4

elements._lineage
# Lineage: sequence -> map(times_2) -> distinct -> cache

l_elements = elements.to_list()
# The cached result is returned so times_2 is not called and nothing is printed

Files are given special treatment if opened through the seq.open and related APIs. functional.util.ReusableFile implements a wrapper around the standard python file to support multiple iteration over a single file object while correctly handling iteration termination and file closing.

Road Map Idea

  • SQL based query planner and interpreter
  • _ lambda operator
  • Prepare for 1.0 next release

Contributing and Bug Fixes

Any contributions or bug reports are welcome. Thus far, there is a 100% acceptance rate for pull requests and contributors have offered valuable feedback and critique on code. It is great to hear from users of the package, especially what it is used for, what works well, and what could be improved.

To contribute, create a fork of PyFunctional, make your changes, then make sure that they pass when running on TravisCI (you may need to sign up for an account and link Github). In order to be merged, all pull requests must:

  • Pass all the unit tests
  • Pass all the pylint tests, or ignore warnings with explanation of why its correct to do so
  • Achieve 100% test coverage on coveralls.io)
  • Edit the CHANGELOG.md file in the Next Release heading with changes

Contact

Gitter for chat

Supported Python Versions

PyFunctional supports and is tested against Python 2.7, 3.5, 3.6, and PyPy

Changelog

Changelog

About me

To learn more about me (the author) visit my webpage at pedrorodriguez.io.

I am a PhD student in Computer Science at the University of Colorado at Boulder. My research interests include large-scale machine learning, distributed computing, and adjacent fields. I completed my undergraduate degree in Computer Science at UC Berkeley in 2015. I have previously done research in the UC Berkeley AMPLab with Apache Spark, worked at Trulia as a data scientist, and worked as a data scientist at Oracle Data Cloud this past summer.

I created PyFunctional while using Python extensively at Trulia, and finding that I missed the ease of use for manipulating data that Spark RDDs and Scala collections have. The project takes the best ideas from these APIs as well as LINQ to provide an easy way to manipulate data when using Scala is not an option or PySpark is overkill.

Contributors

These people have generously contributed their time to improving PyFunctional

comments powered by Disqus