Provides FlickerDataFrame, a wrapper over Pyspark DataFrame to provide a pandas-like API
Project description
🔥flicker
What is flicker?
This python package provides a FlickerDataFrame object. FlickerDataFrame
is a thin wrapper over pyspark.sql.DataFrame. The aim of FlickerDataFrame
is to provide a more Polars-like (not pandas-like) dataframe API.
One way to understand flicker's position is via the following analogy:
keras is to tensorflow as flicker is to pyspark
flicker provides a modern, clean, intuitive, pythonic API over a pyspark
backend. flicker relies completely on pyspark for all distributed
computing work.
How is flicker different than pyspark.pandas (formerly koalas)?
flicker is indeed just an alternative to pyspark.pandas (formerly koalas).
Theoretically, pyspark.pandas can provide similar functionality as flicker but, in practice,
pyspark.pandas suffers from severe performance and usability issues. You can see a detailed example
here.
Flicker, on the other hand, is designed to provide a modern dataframe API. In terms for API design, Flicker is more similar to Polars than to Pandas. Flicker is designed to be just as performant as PySpark itself. And, finally, flicker considers interactive usage (such as exploratory data analysis) as the most important use case.
Installation
flicker is intended to be run with Python>=3.9 and PySpark>=3.4.1. We recommend Python 3.11 and PySpark 3.5.0.
You can install flicker from PyPI:
pip install flicker
If you need to set up Spark on your machine, see pyspark-starter.
Alternatively, you can also build from source.
# Brief instructions. Modify to your requirements.
pip install hatch
git clone https://github.com/ankur-gupta/flicker
cd $REPO_ROOT # typically, ./flicker
hatch build
pip install ./dist/flicker-1.0.0-py3-none-any.whl
Quick Example
flicker aims to simplify some of the common and tedious aspects of a PySpark
dataframe without compromising performance. The following example shows some
of the features of flicker.
from pyspark.sql import SparkSession
from flicker import FlickerDataFrame
# Get a spark session, if needed.
spark = SparkSession.builder.appName('PySparkShell').getOrCreate()
# Set case sensitivity to true.
# https://stackoverflow.com/a/62455158/4383754
spark.conf.set('spark.sql.caseSensitive', True)
# Flicker provides handy factory methods that you can use to create dataframes.
# These factory methods are typically mostly useful to perform quick experiments.
# The easiest one is the `.from_shape` method.
df = FlickerDataFrame.from_shape(spark, nrows=100, ncols=3, names=['a', 'b', 'c'], fill='randn')
# Print the object to see the column names and types
# This returns immediately but doesn't print data.
df
# FlickerDataFrame[a: double, b: double, c: double]
# Best way to get started is to check the shape of the dataframe.
# Spark (and Flicker) already knows the number of columns but the number of rows needs to be
# computed, which can take some time.
df.shape
# (100, 3)
# You can also get the number of rows or columns directly.
df.nrows # returns immediately because nrows is cached
# 100
df.ncols
# 3
# Instead of `df.columns`, use df.names for column names
df.names
# ['a', 'b', 'c']
# dtypes is an OrderedDict
df.dtypes
# OrderedDict([('a', 'double'), ('b', 'double'), ('c', 'double')])
# You can get the dtype for a column name.
df.dtypes['b']
# 'double'
# One of the main features of flicker is the following handy shortcut to view the data.
# Calling a FlickerDataFrame object, returns the first 5 rows as a pandas DataFrame.
# See ?df for more examples on how you can use this to quickly and interactively perform analysis.
df()
# a b c
# 0 -0.593432 0.768301 -0.302519
# 1 -0.11001 0.414888 0.075711
# 2 -0.997298 0.082082 1.080696
# 3 0.299431 -0.072324 -0.055095
# 4 -0.17833 -0.655759 0.252901
# Another cool feature of flicker is pandas-like assignment API. Instead of having to
# use .withColumn(), you can simply assign. For example, if we wanted to create a new
# column that indicates if df['a'] is positive or not, we can do it like this:
df['is_a_positive'] = df['a'] > 0
# See the new column 'is_a_positive'
df # returns immediately
# FlickerDataFrame[a: double, b: double, c: double, is_a_positive: boolean]
# We can now 'call' df to view the first 5 rows.
df()
# a b c is_a_positive
# 0 -0.593432 0.768301 -0.302519 False
# 1 -0.11001 0.414888 0.075711 False
# 2 -0.997298 0.082082 1.080696 False
# 3 0.299431 -0.072324 -0.055095 True
# 4 -0.17833 -0.655759 0.252901 False
# These features can intermixed in nearly every imaginable way. Here are some quick examples.
# Example 1: show the first 2 rows of the dataframe that has only 'a' and 'c' names selected.
df[['a', 'c']](2)
# a c
# 0 -0.593432 -0.302519
# 1 -0.11001 0.075711
# Example 2: Filter the data to select only the rows that have a positive value in column 'a' and
# show the first 3 rows of the filtered dataframe.
df[df['is_a_positive']](3)
# a b c is_a_positive
# 0 0.299431 -0.072324 -0.055095 True
# 1 0.338228 -0.48378 -1.168131 True
# 2 0.578432 -1.223312 -0.546291 True
# Example 3: Show first 2 rows that have a positive product of 'a' and 'b'
df[(df['a'] * df['b']) > 0][['a', 'b']](2)
# a b
# 0 -0.17833 -0.655759
# 1 -0.054472 -0.82237
# You can also get some basic column operations done
df['a'] # returns immediately
# FlickerColumn<'a'>
# Show first 2 values of column 'a'
df[['a']](2)
# a
# 0 -0.593432
# 1 -0.11001
# Describe the distribution of column 'a'
df[['a']].describe()
# a
# summary
# count 100
# mean -0.024628
# stddev 0.980973
# min -2.752549
# max 2.477625
# Get the value counts for 'is_a_positive' column
df['is_a_positive'].value_counts() # returns immediately
# FlickerDataFrame[is_a_positive: boolean, count: bigint]
# See the first 5 rows of the above dataframe by 'calling' it
df['is_a_positive'].value_counts()()
# is_a_positive count
# 0 False 57
# 1 True 43
# Normalize the counts
df['is_a_positive'].value_counts(normalize=True)()
# is_a_positive count
# 0 False 0.57
# 1 True 0.43
Use the underlying PySpark DataFrame or Column
If flicker isn't enough, you can always use the underlying PySpark DataFrame.
Here are a few examples.
from pyspark.sql import SparkSession
from flicker import FlickerDataFrame
from flicker.udf import type_udf
# Get a spark session, if needed.
spark = SparkSession.builder.appName('PySparkShell').getOrCreate()
spark.conf.set('spark.sql.caseSensitive', True)
# Create a more complicated dataframe using one of the factory methods
data = [(1, 'spark', 2.4, {}), (2, 'flicker', None, {'key': 1})]
column_names = ['a', 'b', 'c', 'd']
df = FlickerDataFrame.from_rows(spark, rows=data, names=column_names)
df
# FlickerDataFrame[a: bigint, b: string, c: double, d: map<string,bigint>]
# Typically, NaNs get converted into None(s) but be careful about None vs NaN issues.
df()
# a b c d
# 0 1 spark 2.4 {}
# 1 2 flicker None {'key': 1}
# Always best to extract the FlickerDataFrame into pure python to verify if something is a NaN
# or None.
df.take()
# [{'a': 1, 'b': 'spark', 'c': 2.4, 'd': {}},
# {'a': 2, 'b': 'flicker', 'c': None, 'd': {'key': 1}}]
# `._df` contains the underlying PySpark DataFrame.
# Note that `df._df` is immutable but `df` is mutable.
type(df._df) # pyspark.sql.dataframe.DataFrame
# You can use call any of the underlying methods of `df._df`. Since, `df._df` is immutable,
# you don't have to worry about any `df._df.method()` call modifying `df`.
df._df.show()
# +---+-------+----+----------+
# | a| b| c| d|
# +---+-------+----+----------+
# | 1| spark| 2.4| {}|
# | 2|flicker|NULL|{key -> 1}|
# +---+-------+----+----------+
# You can destructure a FlickerDataFrame by accessing the underlying columns
d = df['d']
d # FlickerColumn<'d'>
type(d) # flicker.column.FlickerColumn
# As before, you can access the underlying pyspark.sql.Column
d._column # Column<'d'>
type(d._column) # pyspark.sql.column.Column
# You can call UDFs on the underlying pyspark.sql.Column
# For example, you can use the built-in flicker.udf.type_udf
df['d_type'] = type_udf(d._column)
# Note the new column 'd_type'
df
# FlickerDataFrame[a: bigint, b: string, c: double, d: map<string,bigint>, d_type: string]
# Use PySpark functions to compute the frequency table based on type of column 'd'
df._df.groupBy(['d_type']).count().show()
# +------+-----+
# |d_type|count|
# +------+-----+
# | dict| 2|
# +------+-----+
# You can, of course, just use the value_counts method.
df['d_type'].value_counts()()
# d_type count
# 0 dict 2
# You can always convert a PySpark DataFrame into a FlickerDataFrame
# after you've performed the native PySpark operations. This way, you can
# continue to enjoy the benefits of FlickerDataFrame. Converting a
# PySpark DataFrame into a FlickerDataFrame is always fast irrespective of
# dataframe size.
df_freq_table = FlickerDataFrame(df._df.groupBy(['d_type']).count())
df_freq_table()
# d_type count
# 0 dict 2
Why not use pyspark.pandas (formerly koalas)?
Koalas was a pandas API over Apache Spark,
which was officially included in PySpark as
pyspark.pandas.
Koalas is now deprecated and directs users towards pyspark.pandas.
You can see the documentation for pyspark.pandas here.
While pyspark.pandas is the official, user-friendly dataframe API, there are three noticeable issues when working with
it.
pyspark.pandascrashes with frequentjava.lang.OutOfMemoryErrorerrors even whenpyspark.sql.DataFrameis capable of handling the same data sizepyspark.pandas's design runs unnecessary spark queries even for non-data-dependent tasks (such as getting the documentation), making interactive use too cumbersome for real-life usepyspark.pandasinherits all the design choices, whether good or bad, from pandas
We demonstrate the above issues with a real-life example. You may need to set up your system for PySpark before you can run the example yourself; see pyspark-starter.
For this example, we will use the publicly available Speedtest by Ookla dataset on AWS S3. This dataset has ~186M rows. The following code snippet was run on an Apple Macbook Pro (M2 Max, 32 GB).
# pyspark.pandas.DataFrame: Frequent OOM errors
import pyspark.pandas as ps
pdf = ps.read_parquet('s3a://ookla-open-data/parquet/performance/type=*/year=*/quarter=*/*.parquet')
type(pdf) # pyspark.pandas.frame.DataFrame
print(f'{pdf.shape[0] / 1e6: .1f}M') # 185.8M
# Just getting one row fails
pdf.head(1)
# [Stage 5:=======> (9 + 12) / 70]
# java.lang.OutOfMemoryError: Java heap space
# ConnectionRefusedError: [Errno 61] Connection refused
# Python session is now unusable: we cannot run any other query now.
pdf.head(1) # fails immediately
# ConnectionRefusedError: [Errno 61] Connection refused
# We cannot even get the shape
pdf.shape # fails immediately
# ConnectionRefusedError: [Errno 61] Connection refused
# We cannot reload the data either. Python session needs to be restarted.
pdf = ps.read_parquet('s3a://ookla-open-data/parquet/performance/type=*/year=*/quarter=*/*.parquet')
# ConnectionRefusedError: [Errno 61] Connection refused
After encountering the OOM error, the entire python session is useless. We now need to restart the python session if we want to run more code. These OOM errors are too frequent and require frequent python restarts.
While one might assume that this data is simply too big for our machine, that's not the case.
We can easily load the same data as a Flicker (or pyspark.sql.DataFrame) dataframe, as shown below.
# flicker.FlickerDataFrame or pyspark.sql.DataFrame: No OOM errors
# See https://github.com/ankur-gupta/pyspark-starter to set up your system for Pyspark and AWS S3.
from pyspark.sql import SparkSession
from flicker import FlickerDataFrame
spark = SparkSession.builder.appName('PySparkShell').getOrCreate()
df = FlickerDataFrame(spark.read.parquet('s3a://ookla-open-data/parquet/performance/type=*/year=*/quarter=*/*.parquet'))
print(f'{df.shape[0] / 1e6: .1f}M') # 185.8M
type(df) # flicker.dataframe.FlickerDataFrame
df() # no OOM error
# [Stage 5:> (0 + 1) / 1]
# quadkey tile avg_d_kbps avg_u_kbps avg_lat_ms tests devices
# 0 0022133222312322 POLYGON((-160.02685546875 70.6435894914449, -1... 8763 3646 45 1 1
# 1 0022133222330013 POLYGON((-160.032348632812 70.6399478155463, -... 9195 3347 43 1 1
# 2 0022133222330023 POLYGON((-160.043334960938 70.6363054807905, -... 6833 3788 42 1 1
# 3 0022133222330100 POLYGON((-160.02685546875 70.6417687358462, -1... 8895 3429 43 2 2
# 4 0022320121121332 POLYGON((-166.739501953125 68.3526207780586, -... 4877 935 45 3 2
df.shape # returns immediately because nrows is cached
# (185832935, 7)
Let's give pyspark.pandas.DataFrame another shot. In a new python session, let's try some basic computation.
# pyspark.pandas.DataFrame: Runs unnecessary spark queries
import pyspark.pandas as ps
pdf = ps.read_parquet('s3a://ookla-open-data/parquet/performance/type=*/year=*/quarter=*/*.parquet')
pdf.columns # returns immediately
# Index(['quadkey', 'tile', 'avg_d_kbps', 'avg_u_kbps', 'avg_lat_ms', 'tests', 'devices'],
# dtype='object')
# Just printing a column errors out. This is because pyspark.pandas tries to load too much
# data at every operation.
pdf['tests'] # OOM error
# [Stage 2:=======> (9 + 12) / 70]
# java.lang.OutOfMemoryError: Java heap space
# ConnectionRefusedError: [Errno 61] Connection refused
# As before, this python session is now useless.
pdf['tests']
# ConnectionRefusedError: [Errno 61] Connection refused
As the above snippet shows, pyspark.pandas runs a spark query in order to mimic pandas-like behavior.
This causes pyspark.pandas to load too much data at nearly every operation, which
exacerbates the OOM error problem. Even if there is no OOM error, the repeated loading of too much data slows down
everything, even when we don't actually want to perform a data-heavy computation. Here is an example.
# In yet another new python session
# pyspark.pandas.DataFrame: Runs unnecessary spark queries
import pyspark.pandas as ps
pdf = ps.read_parquet('s3a://ookla-open-data/parquet/performance/type=*/year=*/quarter=*/*.parquet')
# Note that we didn't actually call the `.value_counts` method yet.
pdf['tests'].value_counts
# [Stage 2:=======> (9 + 12) / 70]
# java.lang.OutOfMemoryError: Java heap space
# ConnectionRefusedError: [Errno 61] Connection refused
We have to restart our python session yet again. This is what makes pyspark.pandas impractical for interactive use.
Interactive use is arguably the most important reason for the existence of a pandas-like API on Spark.
Sometimes, we are able to execute some operations before we eventually run out of memory.
# In yet another new python session
import pyspark.pandas as ps
pdf = ps.read_parquet('s3a://ookla-open-data/parquet/performance/type=*/year=*/quarter=*/*.parquet')
type(pdf['tests']) # pyspark.pandas.series.Series
pdf['tests'].value_counts()
# [Stage 2:=======> (9 + 12) / 71]
# 1 59630189
# 2 25653143
# 3 15311171
# 4 10511920
# 5 7820975
# 6 6146086
# 7 4972223
# ...
# 982 205
# 986 204
# Name: tests, dtype: int64
# Showing only the first 1000
pdf['tests'] == 0
# [Stage 5:===============> (19 + 12) / 70]
# java.lang.OutOfMemoryError: Java heap space
# ConnectionRefusedError: [Errno 61] Connection refused
The OOM error can happen even when the printable output is known to be small.
# In yet another new python session
import pyspark.pandas as ps
pdf = ps.read_parquet('s3a://ookla-open-data/parquet/performance/type=*/year=*/quarter=*/*.parquet')
pdf['devices'].head(1) # no OOM error
# 0 1
# Name: devices, dtype: int64
# Getting the documentation requires executes a spark query, which may or may not fail.
# Even without a failure, we must wait for the spark query to run just to see the documentation.
# This kills interactivity.
?pdf.apply
# [Stage 2:=======> (9 + 12) / 70]
# java.lang.OutOfMemoryError: Java heap space
# py4j.protocol.Py4JNetworkError: Error while sending or receiving
# Your machine may not see an error yet, but it'll still take a long time before you can see the documentation.
# Even if you encounter, you should still see the documentation after the long time.
# However, you'll now have to reload your python session.
pdf['devices'].head(1)
# ConnectionRefusedError: [Errno 61] Connection refused
pyspark.pandas is designed in a way that requires the execution of a spark query even if all you want to do is get
documentation of a method. This "feature" kills interactivity. You'll have to wait a long time for the query to execute
before you'll see the documentation. If the query finishes successfully, you'll just have had to wait a long time, which
can be annoying. If the query runs into an OOM error, you will still be able to see the documentation (after the long
wait) but your python session will now need to be restarted.
Flicker and PySpark's original interface (pyspark.sql.DataFrame) does not suffer from these problems.
Obviously, both flicker.FlickerDataFrame and pyspark.sql.DataFrame can encounter OOM errors, but they don't suffer
from design-induced OOM that are frequently with pyspark.pandas.
Neither pyspark.pandas nor pyspark.sql.DataFrame run a spark query just to access documentation.
# flicker.FlickerDataFrame or pyspark.sql.DataFrame: No spark query to access documentation
# See https://github.com/ankur-gupta/pyspark-starter to set up your system for Pyspark and AWS S3.
from pyspark.sql import SparkSession
from flicker import FlickerDataFrame
spark = SparkSession.builder.appName('PySparkShell').getOrCreate()
df = FlickerDataFrame(spark.read.parquet('s3a://ookla-open-data/parquet/performance/type=*/year=*/quarter=*/*.parquet'))
df # returns immediately
# FlickerDataFrame[quadkey: string, tile: string,
# avg_d_kbps: bigint, avg_u_kbps: bigint, avg_lat_ms: bigint,
# tests: bigint, devices: bigint]
df.names
# ['quadkey', 'tile', 'avg_d_kbps', 'avg_u_kbps', 'avg_lat_ms', 'tests', 'devices']
df['tests'] # returns immediately
# FlickerColumn<'tests'>
df[['tests']]() # no OOM error
# tests
# 0 1
# 1 1
# 2 1
# 3 2
# 4 3
df['tests'].value_counts() # returns immediately
# FlickerDataFrame[tests: bigint, count: bigint]
df['tests'].value_counts()() # no OOM error
# [Stage 7:=======> (9 + 12) / 71]
# tests count
# 0 1 59630189
# 1 2 25653143
# 2 3 15311171
# 3 4 10511920
# 4 5 7820975
df['tests'] == 0 # returns immediately
# FlickerColumn<'(tests = 0)'>
(df['tests'] == 0).value_counts() # returns immediately
# FlickerDataFrame[(tests = 0): boolean, count: bigint]
(df['tests'] == 0).value_counts()() # no OOM error
# [Stage 14:=======> (9 + 12) / 71]
# (tests = 0) count
# 0 False 185832935
?df.head # no spark query, no OOM error
# Signature: df.head(n: 'int | None' = 5) -> 'FlickerDataFrame'
# Docstring: <no docstring>
# File: .../flicker/src/flicker/dataframe.py
# Type: method
df[['devices']](1) # no OOM error
# devices
# 0 1
df[['devices']].value_counts() # still no OOM error
# [Stage 11:=======> (10 + 12) / 71]
# devices count
# 0 1 92155855
# 1 2 30108598
# 2 3 15262519
# 3 4 9404369
# 4 5 6455647
# This is a more time-consuming query, but if the original `spark.sql.DataFrame` can run it,
# then so can `flicker.FlickerDataFrame`.
(df['quadkey'].astype('Long') % 5).value_counts()()
# [Stage 31:=======> (9 + 12) / 71]
# (CAST(quadkey AS BIGINT) % 5) count
# 0 0 46468834
# 1 2 46456313
# 2 1 46454483
# 3 3 46453305
# Weird! No 'quadkey' % 5 == 4.
# We can keep going. No OOM error. No need to restart our python session.
We can keep going. There is always a chance that we will encounter an OOM for some operation.
For example, we will get an expected OOM error if we tried to convert the entire ~186M-row flicker.DataFrame into
a pandas.DataFrame. But, OOM errors are much less frequent than with pyspark.pandas.
In our experiments, it is more likely that your AWS S3 token will expire before you encounter an OOM error. And,
unlike, pyspark.pandas, flicker does not run unnecessary spark queries. This makes interactive use very efficient and predictable.
Status
flicker is actively being developed. While flicker is ready for use, please note that
some API may be changed in the future. Also, it is very likely that you will need a function that has not yet
written in flicker. In such cases, you can always use the underlying PySpark DataFrame to do
every operation that PySpark supports. Please consider filing an issue for
missing functions, bugs, or unintuitive API. Happy sparking!
License
flicker is available under Apache License 2.0.
flicker depends on other python packages listed in requirements.txt
which have their own licenses. flicker releases do not bundle any code from
the dependencies directly.
The documentation is made using Material for MkDocs theme which has MIT License.
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
