The Unix Shell

Pipes and Filters

Overview

Teaching: 15 min
Exercises: 0 min
Questions
  • How can I combine existing commands to do new things?

Objectives
  • Redirect a command’s output to a file.

  • Process a file instead of keyboard input using redirection.

  • Construct command pipelines with two or more stages.

  • Explain what usually happens if a program or pipeline isn’t given any input to process.

  • Explain Unix’s ‘small pieces, loosely joined’ philosophy.

Now that we know a few basic commands, we can finally look at the shell’s most powerful feature: the ease with which it lets us combine existing programs in new ways. We’ll start with a directory called molecules that contains six files describing some simple organic molecules. The .pdb extension indicates that these files are in Protein Data Bank format, a simple text format that specifies the type and position of each atom in the molecule.

$ ls molecules
cubane.pdb    ethane.pdb    methane.pdb
octane.pdb    pentane.pdb   propane.pdb

Let’s go into that directory with cd and run the command wc *.pdb. wc is the “word count” command: it counts the number of lines, words, and characters in files. The * in *.pdb matches zero or more characters, so the shell turns *.pdb into a list of all .pdb files in the current directory:

$ cd molecules
$ wc *.pdb
  20  156 1158 cubane.pdb
  12   84  622 ethane.pdb
   9   57  422 methane.pdb
  30  246 1828 octane.pdb
  21  165 1226 pentane.pdb
  15  111  825 propane.pdb
 107  819 6081 total

Wildcards

* is a wildcard. It matches zero or more characters, so *.pdb matches ethane.pdb, propane.pdb, and every file that ends with ‘.pdb’. On the other hand, p*.pdb only matches pentane.pdb and propane.pdb, because the ‘p’ at the front only matches filenames that begin with the letter ‘p’.

? is also a wildcard, but it only matches a single character. This means that p?.pdb would match pi.pdb or p5.pdb (if we had these two files in the molecules directory), but not propane.pdb. We can use any number of wildcards at a time: for example, p*.p?* matches anything that starts with a ‘p’ and ends with ‘.’, ‘p’, and at least one more character (since the ? has to match one character, and the final * can match any number of characters). Thus, p*.p?* would match preferred.practice, and even p.pi (since the first * can match no characters at all), but not quality.practice (doesn’t start with ‘p’) or preferred.p (there isn’t at least one character after the ‘.p’).

When the shell sees a wildcard, it expands the wildcard to create a list of matching filenames before running the command that was asked for. As an exception, if a wildcard expression does not match any file, Bash will pass the expression as a parameter to the command as it is. For example typing ls *.pdf in the molecules directory (which contains only files with names ending with .pdb) results in an error message that there is no file called *.pdf. However, generally commands like wc and ls see the lists of file names matching these expressions, but not the wildcards themselves. It is the shell, not the other programs, that deals with expanding wildcards, and this is another example of orthogonal design.

Using Wildcards

When run in the molecules directory, which ls command(s) will produce this output?

ethane.pdb methane.pdb

  1. ls *t*ane.pdb
  2. ls *t?ne.*
  3. ls *t??ne.pdb
  4. ls ethane.*

If we run wc -l instead of just wc, the output shows only the number of lines per file:

$ wc -l *.pdb
  20  cubane.pdb
  12  ethane.pdb
   9  methane.pdb
  30  octane.pdb
  21  pentane.pdb
  15  propane.pdb
 107  total

We can also use -w to get only the number of words, or -c to get only the number of characters.

Which of these files is shortest? It’s an easy question to answer when there are only six files, but what if there were 6000? Our first step toward a solution is to run the command:

$ wc -l *.pdb > lengths.txt

The greater than symbol, >, tells the shell to redirect the command’s output to a file instead of printing it to the screen. (This is why there is no screen output: everything that wc would have printed has gone into the file lengths.txt instead.) The shell will create the file if it doesn’t exist. If the file exists, it will be silently overwritten, which may lead to data loss and thus requires some caution. ls lengths.txt confirms that the file exists:

$ ls lengths.txt
lengths.txt

We can now send the content of lengths.txt to the screen using cat lengths.txt. cat stands for “concatenate”: it prints the contents of files one after another. There’s only one file in this case, so cat just shows us what it contains:

$ cat lengths.txt
  20  cubane.pdb
  12  ethane.pdb
   9  methane.pdb
  30  octane.pdb
  21  pentane.pdb
  15  propane.pdb
 107  total

Output Page by Page

We’ll continue to use cat in this lesson, for convenience and consistency, but it has the disadvantage that it always dumps the whole file onto your screen. More useful in practice is the command less, which you use with $ less lengths.txt. This displays a screenful of the file, and then stops. You can go forward one screenful by pressing the spacebar, or back one by pressing b. Press q to quit.

Now let’s use the sort command to sort its contents. We will also use the -n flag to specify that the sort is numerical instead of alphabetical. This does not change the file; instead, it sends the sorted result to the screen:

$ sort -n lengths.txt
  9  methane.pdb
 12  ethane.pdb
 15  propane.pdb
 20  cubane.pdb
 21  pentane.pdb
 30  octane.pdb
107  total

We can put the sorted list of lines in another temporary file called sorted-lengths.txt by putting > sorted-lengths.txt after the command, just as we used > lengths.txt to put the output of wc into lengths.txt. Once we’ve done that, we can run another command called head to get the first few lines in sorted-lengths.txt:

$ sort -n lengths.txt > sorted-lengths.txt
$ head -n 1 sorted-lengths.txt
  9  methane.pdb

Using the parameter -n 1 with head tells it that we only want the first line of the file; -n 20 would get the first 20, and so on. Since sorted-lengths.txt contains the lengths of our files ordered from least to greatest, the output of head must be the file with the fewest lines.

Redirecting to the same file

It’s a very bad idea to try redirecting the output of a command that operates on a file to the same file. For example:

$ sort -n lengths.txt > lengths.txt

Doing something like this may give you incorrect results and/or delete the contents of lengths.txt.

If you think this is confusing, you’re in good company: even once you understand what wc, sort, and head do, all those intermediate files make it hard to follow what’s going on. We can make it easier to understand by running sort and head together:

$ sort -n lengths.txt | head -n 1
  9  methane.pdb

The vertical bar, |, between the two commands is called a pipe. It tells the shell that we want to use the output of the command on the left as the input to the command on the right. The computer might create a temporary file if it needs to, or copy data from one program to the other in memory, or something else entirely; we don’t have to know or care.

Nothing prevents us from chaining pipes consecutively. That is, we can for example send the output of wc directly to sort, and then the resulting output to head. Thus we first use a pipe to send the output of wc to sort:

$ wc -l *.pdb | sort -n
   9 methane.pdb
  12 ethane.pdb
  15 propane.pdb
  20 cubane.pdb
  21 pentane.pdb
  30 octane.pdb
 107 total

And now we send the output of this pipe, through another pipe, to head, so that the full pipeline becomes:

$ wc -l *.pdb | sort -n | head -n 1
   9  methane.pdb

This is exactly like a mathematician nesting functions like log(3x) and saying “the log of three times x”. In our case, the calculation is “head of sort of line count of *.pdb”.

Here’s what actually happens behind the scenes when we create a pipe. When a computer runs a program — any program — it creates a process in memory to hold the program’s software and its current state. Every process has an input channel called standard input. (By this point, you may be surprised that the name is so memorable, but don’t worry: most Unix programmers call it “stdin”). Every process also has a default output channel called standard output (or “stdout”). A third output channel called standard error (stderr) also exists. This channel is typically used for error or diagnostic messages, and it allows a user to pipe the output of one program into another while still receiving error messages in the terminal.

The shell is actually just another program. Under normal circumstances, whatever we type on the keyboard is sent to the shell on its standard input, and whatever it produces on standard output is displayed on our screen. When we tell the shell to run a program, it creates a new process and temporarily sends whatever we type on our keyboard to that process’s standard input, and whatever the process sends to standard output to the screen.

Here’s what happens when we run wc -l *.pdb > lengths.txt. The shell starts by telling the computer to create a new process to run the wc program. Since we’ve provided some filenames as parameters, wc reads from them instead of from standard input. And since we’ve used > to redirect output to a file, the shell connects the process’s standard output to that file.

If we run wc -l *.pdb | sort -n instead, the shell creates two processes (one for each process in the pipe) so that wc and sort run simultaneously. The standard output of wc is fed directly to the standard input of sort; since there’s no redirection with >, sort’s output goes to the screen. And if we run wc -l *.pdb | sort -n | head -n 1, we get three processes with data flowing from the files, through wc to sort, and from sort through head to the screen.

Redirects and Pipes

This simple idea is why Unix has been so successful. Instead of creating enormous programs that try to do many different things, Unix programmers focus on creating lots of simple tools that each do one job well, and that work well with each other. This programming model is called “pipes and filters”. We’ve already seen pipes; a filter is a program like wc or sort that transforms a stream of input into a stream of output. Almost all of the standard Unix tools can work this way: unless told to do otherwise, they read from standard input, do something with what they’ve read, and write to standard output.

The key is that any program that reads lines of text from standard input and writes lines of text to standard output can be combined with every other program that behaves this way as well. You can and should write your programs this way so that you and other people can put those programs into pipes to multiply their power.

Redirecting Input

As well as using > to redirect a program’s output, we can use < to redirect its input, i.e., to read from a file instead of from standard input. For example, instead of writing wc ammonia.pdb, we could write wc < ammonia.pdb. In the first case, wc gets a command line parameter telling it what file to open. In the second, wc doesn’t have any command line parameters, so it reads from standard input, but we have told the shell to send the contents of ammonia.pdb to wc’s standard input.

Nelle’s Pipeline: Checking Files

We want to check the dimensions of our image files The ‘file’ command gives us information about files

$ cd ../photos
$ file *.jpg

The output is 1520 lines that look like this:

1411.jpg: JPEG image data, JFIF standard 1.01, aspect ratio, density 1x1,
segment length 16, baseline, precision 8, 1000x678, frames 3
1412.jpg: JPEG image data, JFIF standard 1.01, aspect ratio, density 1x1,
segment length 16, baseline, precision 8, 1000x700, frames 3
1413.jpg: JPEG image data, JFIF standard 1.01, aspect ratio, density 1x1,
segment length 16, baseline, precision 8, 1000x680, frames 3
1414.jpg: JPEG image data, JFIF standard 1.01, aspect ratio, density 1x1,
segment length 16, baseline, precision 8, 1000x704, frames 3
1415.jpg: JPEG image data, JFIF standard 1.01, aspect ratio, density 1x1,
segment length 16, baseline, precision 8, 1000x696, frames 3
1416.jpg: JPEG image data, JFIF standard 1.01, aspect ratio, density 1x1,
segment length 16, baseline, precision 8, 1000x697, frames 3
1417.jpg: JPEG image data, JFIF standard 1.01, aspect ratio, density 1x1,
segment length 16, baseline, precision 8, 1000x692, frames 3
1418.jpg: JPEG image data, JFIF standard 1.01, aspect ratio, density 1x1,
segment length 16, baseline, precision 8, 1000x679, frames 3
1419.jpg: JPEG image data, JFIF standard 1.01, aspect ratio, density 1x1,
segment length 16, baseline, precision 8, 1000x684, frames 3
1420.jpg: JPEG image data, JFIF standard 1.01, aspect ratio, density 1x1,
segment length 16, baseline, precision 8, 699x1000, frames 3
... ...

We can use the awk command to segment a response by a separator

$ echo "1,2,3" | awk -F "," '{print $3,$1}'
3 1

So to sort the first image dimensions from smallest to largest

$ file *.jpg | awk -F"," '{print $8,$1 }' | sort -n
 699x1000 1420.jpg: JPEG image data
 1000x678 1411.jpg: JPEG image data
 1000x679 1418.jpg: JPEG image data
 1000x680 1413.jpg: JPEG image data
 1000x684 1419.jpg: JPEG image data
 1000x692 1417.jpg: JPEG image data
 1000x696 1415.jpg: JPEG image data
 1000x697 1416.jpg: JPEG image data
 1000x700 1412.jpg: JPEG image data
 1000x704 1414.jpg: JPEG image data

Extension Challenge

Use the commands above to sort on the second dimension

Solution

file *.jpg | awk -F”,” ‘{print $8,$1 }’ | awk -F”x” ‘{print $2,$1 }’ | sort -n

What Does >> Mean?

What is the difference between:

$ echo hello > testfile01.txt

and:

$ echo hello >> testfile02.txt

Hint: Try executing each command twice in a row and then examining the output files.

More on Wildcards

Sam has a directory containing calibration data, datasets, and descriptions of the datasets:

2015-10-23-calibration.txt
2015-10-23-dataset1.txt
2015-10-23-dataset2.txt
2015-10-23-dataset_overview.txt
2015-10-26-calibration.txt
2015-10-26-dataset1.txt
2015-10-26-dataset2.txt
2015-10-26-dataset_overview.txt
2015-11-23-calibration.txt
2015-11-23-dataset1.txt
2015-11-23-dataset2.txt
2015-11-23-dataset_overview.txt

Before heading off to another field trip, she wants to back up her data and send some datasets to her colleague Bob. Sam uses the following commands to get the job done:

$ cp *dataset* /backup/datasets
$ cp ____calibration____ /backup/calibration
$ cp 2015-____-____ ~/send_to_bob/all_november_files/
$ cp ____ ~/send_to_bob/all_datasets_created_on_a_23rd/

Help Sam by filling in the blanks.

Solution

$ cp *calibration.txt /backup/calibration
$ cp 2015-11-* ~/send_to_bob/all_november_files/
$ cp *-23-dataset* ~send_to_bob/all_datasets_created_on_a_23rd/

Piping Commands Together

In our current directory, we want to find the 3 files which have the least number of lines. Which command listed below would work?

  1. wc -l * > sort -n > head -n 3
  2. wc -l * | sort -n | head -n 1-3
  3. wc -l * | head -n 3 | sort -n
  4. wc -l * | sort -n | head -n 3

Solution

Option 4 is the solution. The pipe character | is used to feed the standard output from one process to the standard input of another. > is used to redirect standard output to a file. Try it in the data-shell/molecules directory!

Why Does uniq Only Remove Adjacent Duplicates?

The command uniq removes adjacent duplicated lines from its input. For example, the file data-shell/data/salmon.txt contains:

coho
coho
steelhead
coho
steelhead
steelhead

Running the command uniq salmon.txt from the data-shell/data directory produces:

coho
steelhead
coho
steelhead

Why do you think uniq only removes adjacent duplicated lines? (Hint: think about very large data sets.) What other command could you combine with it in a pipe to remove all duplicated lines?

Solution

$ sort salmon.txt | uniq

Removing Unneeded Files

Suppose you want to delete your processed data files, and only keep your raw files and processing script to save storage. The raw files end in .dat and the processed files end in .txt. Which of the following would remove all the processed data files, and only the processed data files?

  1. rm ?.txt
  2. rm *.txt
  3. rm * .txt
  4. rm *.*

Solution

  1. This would remove .txt files with one-character names
  2. This is correct answer
  3. The shell would expand * to match everything in the current directory, so the command would try to remove all matched files and an additional file called .txt
  4. The shell would expand *.* to match all files with any extension, so this command would delete all files

Wildcard Expressions

Wildcard expressions can be very complex, but you can sometimes write them in ways that only use simple syntax, at the expense of being a bit more verbose.
Consider the directory data-shell/north-pacific-gyre/2012-07-03 : the wildcard expression *[AB].txt matches all files ending in A.txt or B.txt. Imagine you forgot about this.

  1. Can you match the same set of files with basic wildcard expressions that do not use the [] syntax? Hint: You may need more than one expression.

  2. The expression that you found and the expression from the lesson match the same set of files in this example. What is the small difference between the outputs?

  3. Under what circumstances would your new expression produce an error message where the original one would not?

Solution

$ ls *A.txt
$ ls *B.txt
  1. The output from the new commands is separated because there are two commands.
  2. When there are no files ending in A.txt, or there are no files ending in B.txt.

Which Pipe?

The file contains 586 lines of data formatted as follows:

2012-11-05,deer
2012-11-05,rabbit
2012-11-05,raccoon
2012-11-06,rabbit
...

Assuming your current directory is data-shell/data/, what command would you use to produce a table that shows the total count of each type of animal in the file?

  1. grep {deer, rabbit, raccoon, deer, fox, bear} animals.txt | wc -l
  2. sort animals.txt | uniq -c
  3. sort -t, -k2,2 animals.txt | uniq -c
  4. cut -d, -f 2 animals.txt | uniq -c
  5. cut -d, -f 2 animals.txt | sort | uniq -c
  6. cut -d, -f 2 animals.txt | sort | uniq -c | wc -l

Solution

Option 5. is the correct answer. If you have difficulty understanding why, try running the commands, or sub-sections of the pipelines (make sure you are in the data-shell/data directory).

Key Points