André Boechat
by André Boechat
7 min read



In many situations, unprepared data scientists can spend much more time than necessary on secondary tasks. Although their attention should stay on analyzing data, checking hypothesis, engineering features, etc., they often need to get their hands dirty and code auxiliary scripts and parsers to get the information they need.

Prepared data scientists prefer to use Unix commands and solve their secondary tasks in a few seconds.

This post presents 7 basic Unix commands (maybe eight) that, once incorporated in the day-by-day toolset, can potentially improve the productivity of any data scientist. The beauty of these tools is the ability to easily combine them into powerful commands, executing tasks that could take a full day of coding to get them right.

This is the selected list, probably in the order of most frequent usage:

In addition, it is shown how two auxiliary commands (xargs and man) can improve even further the usability of the 7 commands above.

7 Unix Commands


grep searches for patterns in files or in the standard input. It is really useful to find stuff without opening files in an editor.


Problem: We want to find where the library ggplot was used in any script in the directory r_project.

$ grep -rn ggplot r_project/
r_project/script.R:247:t<-ggplot(base.test.c, aes(x=score, colour=gap_dic, group=gap_dic))
r_project/script.R:265:t<-ggplot(base.test.c[base.test.c$gap_total<40000,], aes(x=gap_total, colour=corte2, group=corte2))
  • option -r stands for recursion
  • option -n tells grep to show the line numbers

Problem: From a big log file, we want to get only the logging messages with the pattern TRAINING -.

$ grep 'TRAINING -' my_app.log

Note that grep is case sensitive by default.

Problem: Given a directory composed of subdirectories whose names contain a date, we want to get the complete name of one or more subdirectories containing a date like 2018_05.

$ ls models_by_date/ | grep 2018_05

Here we use the pipe | operator to send the output of the first command to grep by the standard input 1.

Problem: When listing all installed Python packages, we want to see only the results containing the name gpyopt (usually a single line).

$ pip freeze | grep -i gpyopt

pip freeze returns a list of installed packages to the standard output and grep searches for gpyopt (-i makes it case insensitive) in the standard input.

Problem: Given an arbitrary text file, we want to show only those lines containing a pattern specified by a regex.

$ grep -oP "'[\w]+ == [\d.]+'" python_library/
'numpy == 1.15.0'
'fire == 0.1.3'
'gpyopt == 1.2.5'
'recsys_commons == 0.1.0'

This example uses Perl regular expression to search for packages and versions in a Python setup file.


Prints on the screen (or to the standard output) the contents of files. Simple like that.

$ cat

set -eo pipefail

echo "$BLEU"


find searches for files by specifying many different (and optionally) kinds of parameters. It is also able to execute some simple actions or an entire command line using the resultant files.


Problem: We want to find all files with the extension json in the current directory, including subdirectories.

$ find . -name '*.json'

Problem: All files with extension pyc must be removed from the directory my_library/modules, recursively.

$ find my_library/modules -name '*.pyc' -delete

Problem: In a directory containing multiple projects, we want to find all files that contain the text boto3 (in other words, we are looking for projects using the library boto3).

$ find . -name -type f -exec grep -Hn boto3 {} \;

Here we tell it to search only for files with the exact name, ignoring directories (-type f) and executing grep (-exec) on each one of them (why don’t we just use grep alone?)2. Note that:

  • {} is used to pass a file path as an argument to grep
  • \; marks the end of the command
  • grep parameter -H makes it print the filename


Like working with DataFrames, head and tail print the first and the last lines of files or of the standard input.


Problem: Given a CSV file, we want to quickly look at its header.

$ head -n 1 data.csv

Problem: From a potentially huge log file, we want to read only its last 20 events.

$ tail -n 20 app.log


wc is very useful to count lines, words or even characters in files.


Problem: How many text lines does a file have?

$ wc -l data.csv
624 data.csv

Problem: We want to know the total number of CSV records in a directory containing multiple CSV files.

$ wc -l data_dir/*.csv
102224 data_dir/part-00000-02aa95cd-3907-44c8-87ee-97ff44677349-c000.csv
102513 data_dir/part-00001-02aa95cd-3907-44c8-87ee-97ff44677349-c000.csv
204737 total

We may need to discount the number of header lines (this could also be done just using the shell).


awk uses a programming language (AWK) for text processing. It is powerful and might seem complicated to learn and use. However, there are a few commands that can be used very frequently.


Problem: Given a CSV file, we want to know the number of columns just by analyzing its header.

$ head -n 1 data.csv | awk -F ',' '{print NF}'

First, head sends the first line to the standard output, which is consumed (as the standard input) by awk and broken up into a sequence of fields delimited by ,. NF holds the number of fields in a line.

Problem: Given a big CSV file containing hundreds of columns, we want to have a quick look at the first lines of a specific column (let’s say the third column).

$ head data.csv | awk -F ',' '{print $3}'


shuf generates a random permutation of its inputs. It is very useful to sample data.


Problem: Given a CSV file, we want to take a random sample (50 records) from it and save this sample in another file.

$ cat big_csv.csv | shuf | head -n 50 > sample_from_big_csv.csv

The operator > is used to redirect the standard output to a normal file. Without using it, the result of head would be just printed on the screen.

Problem: Given a directory containing multiple data files, we want to get a random sample of files (5 files) and copy these files to another directory.

$ find origin_dir/ -type f | shuf | head -n 5 | xargs -i cp {} sample_dir/

First find returns a list of files in origin_dir (including this directory name in their paths), then shuf shuffles the list of file paths, head takes the first 5 file paths and, finally, cp copy each of these 5 files to the directory sample_dir (xargs, explained in the next section, is used as an auxiliary command since we couldn’t just use the standard input).

Auxiliary Commands


xargs is a kind of auxiliary program, since its role is to convert the standard input into an argument of another program. This is really useful to make a chain of processing programs that don’t use the standard input.


Problem: We must remove all __pycache__ directories in a given project directory.

$ find my_app -name '__pycache__' -type d | xargs -i rm -r {}

As find action -delete can’t remove nonempty directories, we use rm -r to remove all __pycache__ folders.

Problem: Remove all Git branches that were already fully merged with their upstream branches.

$ git branch | xargs -i git branch -d {}


man is also another auxiliary program. It provides an interface to reference manuals of almost all UNIX commands. The image below shows the result of checking the manual of man itself (man man).


By Kamran Mackey - Arch Linux using the Cinnamon display manager.,GPL.


This post presented 7 Unix commands to increase the productivity of data scientists, including examples of usage and two auxiliary commands to make even more powerful combinations. By describing common problems and possible solutions, the intention is to make the post a useful reference to tackle similar scenarios.

More time we spend exploring and using these commands, more productive we become by using them.

  1. The standard input can be thought as a special kind of file whose contents commands have access to. The pipe operator is able to take the standard output of one command (that would otherwise be printed on the screen) and send it to the standard input of another. 

  2. Using grep recursively would bring all lines of all files containing boto3; the result might be too big to be useful.