Storage Formats#

Prerequisites

Intro to DataFrames and Series

Outcomes

Understand that data can be saved in various formats
Know where to get help on file input and output
Know when to use csv, xlsx, feather, and sql formats

Data

Results for all NFL games between September 1920 to February 2017

# Uncomment following line to install on colab
#! pip install 

import pandas as pd
import numpy as np

File Formats#

Data can be saved in a variety of formats.

pandas understands how to write and read DataFrames to and from many of these formats.

We defer to the official documentation for a full description of how to interact with all the file formats, but will briefly discuss a few of them here.

CSV#

What is it? CSVs store data as plain text (strings) where each row is a line and columns are separated by ,.

Pros

Widely used (you should be familiar with it)
Plain text file (can open on any computer, “future proof”)
Can be read from and written to by most data software

Cons

Not the most efficient way to store or access
No formal standard, so there is room for user interpretation on how to handle edge cases (e.g. what to do about a data field that itself includes a comma)

When to use:

A great default option for most use cases

xlsx#

What is it? xlsx is a binary file format used as Excel’s default.

Pros:

Standard format in many industries
Easy to share with colleagues that use Excel

Cons:

Quite slow to read/write large amounts of data
Stores both data and metadata like styling and display information and even plots. This metadata is not always portable to other file formats or programs.

When to use:

When sharing data with Excel
When you would like special formatting to be applied to the spreadsheet when viewed in Excel

Parquet#

What is it? Parquet is a custom binary format designed for efficient reading and writing of data stored in columns.

Pros:

Very fast
Naturally understands all dtypes used by pandas, including multi-index DataFrames
Very common in “big data” systems like Hadoop or Spark
Supports various compression algorithms

Cons:

Binary storage format that is not human-readable

When to use:

If you have “not small” amounts (> 100 MB) of unchanging data that you want to read quickly
If you want to store data in an size-and-time-efficient way that may be accessed by external systems

Feather#

What is it? Feather is a custom binary format designed for efficient reading and writing of data stored in columns.

Pros:

Very fast – even faster than parquet
Naturally understands all dtypes used by pandas

Cons:

Can only read and write from Python and a handful of other programming languages
New file format (introduced in March ‘16), so most files don’t come in this format
Only supports standard pandas index, so you need to reset_index before saving and then set_index after loading

When to use:

Use as an alternative to Parquet if you need the absolute best read and write speeds for unchanging datasets
Only use when you will not need to access the data in a programming language or software outside of Python, R, and Julia

SQL#

What is it? SQL is a language used to interact with relational databases… more info

Pros:

Well established industry standard for handling data
Much of the world’s data is in a SQL database somewhere

Cons:

Complicated: to have full control you need to learn another language (SQL)

When to use:

When reading from or writing to existing SQL databases

NOTE: We can cover interacting with SQL databases in a dedicated lecture – contact us for more information.

Writing DataFrames#

Let’s now talk about saving a DataFrame to a file.

As a general rule of thumb, if we have a DataFrame df and we would like to save to save it as a file of type FOO, then we would call the method named df.to_FOO(...).

We will show you how this can be done and try to highlight some of the items mentioned above.

But, we will not cover all possible options and features — we feel it is best to learn these as you need them by consulting the appropriate documentation.

First, we need some DataFrames to save. Let’s make them now.

Note that by default df2 will be approximately 10 MB.

If you need to change this number, adjust the value of the wanted_mb variable below.

np.random.seed(42)  # makes sure we get the same random numbers each time
df1 = pd.DataFrame(
    np.random.randint(0, 100, size=(10, 4)),
    columns=["a", "b", "c", "d"]
)

wanted_mb = 10  # CHANGE THIS LINE
nrow = 100000
ncol = int(((wanted_mb * 1024**2) / 8) / nrow)
df2 = pd.DataFrame(
    np.random.rand(nrow, ncol),
    columns=["x{}".format(i) for i in range(ncol)]
)

print("df2.shape = ", df2.shape)
print("df2 is approximately {} MB".format(df2.memory_usage().sum() / (1024**2)))

df2.shape =  (100000, 13)
df2 is approximately 9.9183349609375 MB

df.to_csv #

Let’s start with df.to_csv.

Without any additional arguments, the df.to_csv function will return a string containing the csv form of the DataFrame:

# notice the plain text format -- one row per line, columns separated by `'`
print(df1.to_csv())

,a,b,c,d
0,51,92,14,71
1,60,20,82,86
2,74,74,87,99
3,23,2,21,52
4,1,87,29,37
5,1,63,59,20
6,32,75,57,21
7,88,48,90,58
8,41,91,59,79
9,14,61,61,46

If we do pass an argument, the first argument will be used as the file name.

df1.to_csv("df1.csv")

Run the cell below to verify that the file was created.

import os
os.path.isfile("df1.csv")

True

Let’s see how long it takes to save df2 to a file. (Because of the %%time at the top, Jupyter will report the total time to run all code in the cell)

%%time
df2.to_csv("df2.csv")

CPU times: user 1.14 s, sys: 90.7 ms, total: 1.23 s
Wall time: 1.24 s

As we will see below, this isn’t as fastest file format we could choose.

df.to_excel #

When saving a DataFrame to an Excel workbook, we can choose both the name of the workbook (file) and the name of the sheet within the file where the DataFrame should be written.

We do this by passing the workbook name as the first argument and the sheet name as the second argument as follows.

df1.to_excel("df1.xlsx", "df1")

pandas also gives us the option to write more than one DataFrame to a workbook.

To do this, we need to first construct an instance of pd.ExcelWriter and then pass that as the first argument to df.to_excel.

Let’s see how this works.

with pd.ExcelWriter("df1.xlsx") as writer:
    df1.to_excel(writer, "df1")
    (df1 + 10).to_excel(writer, "df1 plus 10")

The

with ... as ... :

syntax used above is an example of a context manager.

We don’t need to understand all the details behind what this means (google it if you are curious).

For now, just recognize that particular syntax as the way to write multiple sheets to an Excel workbook.

Warning

Saving df2 to an excel file takes a very long time.

For that reason, we will just show the code and hard-code the output we saw when we ran the code.

%%time
df2.to_excel("df2.xlsx")

 Wall time: 25.7 s

pyarrow.feather.write_feather #

As noted above, the feather file format was developed for very efficient reading and writing between Python and your computer.

Support for this format is provided by a separate Python package called pyarrow.

This package is not installed by default. To install it, copy/paste the code below into a code cell and execute.

!pip install pyarrow

The parameters for pyarrow.feather.write_feather are the DataFrame and file name.

Let’s try it out.

import pyarrow.feather
pyarrow.feather.write_feather(df1, "df1.feather")

%%time
pyarrow.feather.write_feather(df2, "df2.feather")

CPU times: user 30.1 ms, sys: 8.6 ms, total: 38.7 ms
Wall time: 20.5 ms

An example timing result:

format	time
csv	2.66 seconds
xlsx	25.7 seconds
feather	43 milliseconds

As you can see, saving this DataFrame in the feather format was far faster than either CSV or Excel.

Reading Files into DataFrames#

As with the df.to_FOO family of methods, there are similar pd.read_FOO functions. (Note: they are in defined pandas, not as methods on a DataFrame.)

These methods have many more options because data storage can be messy or wrong.

We will explore these in more detail in a separate lecture.

For now, we just want to highlight the differences in how to read data from each of the file formats.

Let’s start by reading the files we just created to verify that they match the data we began with.

# notice that index was specified in the first (0th -- why?) column of the file
df1_csv = pd.read_csv("df1.csv", index_col=0)
df1_csv.head()

	a	b	c	d
0	51	92	14	71
1	60	20	82	86
2	74	74	87	99
3	23	2	21	52
4	1	87	29	37

df1_xlsx = pd.read_excel("df1.xlsx", "df1", index_col=0)
df1_xlsx.head()

	a	b	c	d
0	51	92	14	71
1	60	20	82	86
2	74	74	87	99
3	23	2	21	52
4	1	87	29	37

# notice feather already knows what the index is
df1_feather = pyarrow.feather.read_feather("df1.feather")
df1_feather.head()

	a	b	c	d
0	51	92	14	71
1	60	20	82	86
2	74	74	87	99
3	23	2	21	52
4	1	87	29	37

With the pd.read_FOO family of functions, we can also read files from places on the internet.

We saved our df1 DataFrame to a file and posted it online.

Below, we show an example of using pd.read_csv to read this file.

df1_url = "https://raw.githubusercontent.com/QuantEcon/lecture-datascience.myst/main/lectures/pandas/df1.csv"
df1_web = pd.read_csv(df1_url, index_col=0)
df1_web.head()

	a	b	c	d
0	51	92	14	71
1	60	20	82	86
2	74	74	87	99
3	23	2	21	52
4	1	87	29	37

Practice#

Now it’s your turn…

In the cell below, the variable url contains a web address to a csv file containing the result of all NFL games from September 1920 to February 2017.

Your task is to do the following:

Use pd.read_csv to read this file into a DataFrame named nfl
Print the shape and column names of nfl
Save the DataFrame to a file named nfl.xlsx
Open the spreadsheet using Excel on your computer

If you finish quickly, do some basic analysis of the data. Try to do something interesting. If you get stuck, here are some suggestions for what to try:

Compute the average total points in each game (note, you will need to sum two of the columns to get total points).
Repeat the above calculation, but only for playoff games.
Compute the average score for your favorite team (you’ll need to consider when they were team1 vs team2).
Compute the ratio of “upsets” to total games played. An upset is defined as a team with a lower ELO winning the game.

url = "https://raw.githubusercontent.com/fivethirtyeight/nfl-elo-game/"
url = url + "3488b7d0b46c5f6583679bc40fb3a42d729abd39/data/nfl_games.csv"

# your code here --- create more cells if necessary

Cleanup#

If you want to remove the files we just created, run the following cell.

def try_remove(file):
    if os.path.isfile(file):
        os.remove(file)

for df in ["df1", "df2"]:
    for extension in ["csv", "feather", "xlsx"]:
        filename = df + "." + extension
        try_remove(filename)

Storage Formats#

File Formats#

CSV#

xlsx#

Parquet#

Feather#

SQL#

Writing DataFrames#

df.to_csv#

df.to_excel#

pyarrow.feather.write_feather#

Reading Files into DataFrames#

Practice#

Cleanup#

df.to_csv #

df.to_excel #

pyarrow.feather.write_feather #