CMPT 732, Fall 2023
This topic is a bit of a tangent: maybe useful for other courses, maybe useful in jobs or job interviews.
Lots of data sets aren't big. In fact, most aren't.
Modern phones have ≥4 GB of memory: if you have less than that, it must be small
. Why use Spark for everything?
Even running locally, Spark has a ≈10 s startup time: any work that takes less than that makes no sense in Spark.
Good reasons to use Spark: [editorial content]
medium datawhere the Spark startup time is worth it when running locally because you will then use multiple cores.
A very good use-case for Spark: ETL work that makes big data small.
Use Spark to extract/aggregate the data you really want to work with. Realize that made your data small
. Move to some small data tools…
Python is one of the most common choices for data science work. (The other is R.)
As a result, there are many very mature data manipulation tools in Python. You should know they exist.
Python's built-in data structures are not very memory-efficient: Python object overhead, references cause bad memory locality, etc.
Data you have will often have fixed types and sizes: exactly what C-style arrays are good at. NumPy provides efficient, typed arrays for Python.
import numpy as np a = np.zeros((10000, 5), dtype=np.float32) print(a) print((a + 6).sum())
[[ 0. 0. 0. 0. 0.] [ 0. 0. 0. 0. 0.] [ 0. 0. 0. 0. 0.] ..., [ 0. 0. 0. 0. 0.] [ 0. 0. 0. 0. 0.] [ 0. 0. 0. 0. 0.]] 300000.0
NumPy can do lots of manipulation on arrays (at C-implemention speeds). e.g.
Pandas provides a DataFrame class for in-memory data manipulation. Pandas DataFrame ≠ Spark DataFrame, but concepts are similar.
import pandas as pd
cities = pd.read_csv('cities.csv')
print(cities)
city population area 0 Vancouver 2463431 2878.52 1 Calgary 1392609 5110.21 2 Toronto 5928040 5905.71 3 Montreal 4098927 4604.26 4 Halifax 403390 5496.31
Similar operate-on-whole-DataFrame API. Slightly different operations. Not lazily evaluated.
cities['area_m2'] = cities['area'] * 1e6 print(cities)
city population area area_m2 0 Vancouver 2463431 2878.52 2.878520e+09 1 Calgary 1392609 5110.21 5.110210e+09 2 Toronto 5928040 5905.71 5.905710e+09 3 Montreal 4098927 4604.26 4.604260e+09 4 Halifax 403390 5496.31 5.496310e+09
Pandas Series (==columns) are stored as NumPy arrays, so you can use NumPy functions if you need to.
print(type(cities['population'].values)) print(cities['population'].values.dtype)
<class 'numpy.ndarray'> int64
If you think of each partition of a Spark DataFrame as a Pandas DataFrame, you'd be technically wrong, but conceptually correct.
That's why Spark's Vectorized UDFs make sense: you get a Pandas Series of a column for each partition and can work on those partition-by-partition.
The Pandas API is similar to Spark: specify operations on columns, operating on all of the data as a single operation. These are equivalent Pandas and Spark operations:
big_cities = cities[cities['population'] > 2000000] two_cols = cities[['city', 'area']]
big_cities = cities.where(cities['population'] > 2000000)
two_cols = cities.select('city', 'area')
The Pandas API has more operations: they tend to be more comprehensive overall, but also aren't limited to things that can be distributed across multiple executor threads/nodes.
e.g. DataFrame.mask, DataFrame.melt, Series.str.extract, Series.to_latex.
Pandas and data science resources:
In Spark 3.2, the Koalas Pandas API on Spark
has been merged into Spark.
But use caution: this is another layer of abstraction on the Spark API. If you have truly-big data sets, it would be easy to get yourself in trouble with expensive operations you can't immediately see. [Best practices in the docs: check execution plans; avoid shuffling; avoid computation on single partition.]
A Pandas DataFrame can be converted to a Spark DataFrame (if you need distributed computation, want to join a Spark DataFrame, etc):
cities_pd = pd.read_csv('cities.csv')
cities_spark = spark.createDataFrame(cities_pd)
cities_spark.show()
+---------+----------+-------+ | city|population| area| +---------+----------+-------+ |Vancouver| 2463431|2878.52| | Calgary| 1392609|5110.21| | Toronto| 5928040|5905.71| | Montreal| 4098927|4604.26| | Halifax| 403390|5496.31| +---------+----------+-------+
… and a Spark DataFrame to Pandas if it will fit in memory in the driver process (if you shrunk the data and want Python/Pandas functionality):
cities_pd2 = cities_spark.toPandas() print(cities_pd2)
city population area 0 Vancouver 2463431 2878.52 1 Calgary 1392609 5110.21 2 Toronto 5928040 5905.71 3 Montreal 4098927 4604.26 4 Halifax 403390 5496.31
This is faster in Spark ≥2.3 if you use the Apache Arrow option.
With NumPy and Pandas, you can do a lot of basic data manipulation operations.
They will likely be faster on small (and medium?) data: no overhead of managing executors or distributing data, but single-threaded.
The SciPy libraries include many useful tools to analyze data. Some examples:
scipy.fftpack)scipy.signal)scipy.linalg)scipy.stats)scipy.ndimage)matplotlib)If those aren't enough, there are SciKits containing much more. e.g.
Scikit-learn is probably going to be useful to you some time: implementations of many machine learning algorithms for Python (and NumPy-shaped data).
Compared to pyspark.ml: older and more battle-tested; includes algorithms that don't distribute well; doesn't do distributed computation.
Maybe the biggest pro-Python argument: it's used for data science and many other things, so libraries you need are implemented in Python.
PyPI is the package repository for Python. You can install packages with the pip command.
pip3 install --user scikit-learn scipy pandas