A dozen (or so) ways to start sub-processes in Ruby: Part 2

In the previous article we looked at some basic methods for starting subprocesses in Ruby. One thing all those methods had in common was that they didn’t permit a lot of communication between parent process and child. In this article we’ll examine a few built-in Ruby methods which give us the ability to have a two-way conversation with our subprocesses.

The complete source code for this article can be found at http://gist.github.com/146199.

Method #4: Opening a pipe

As you know, the Kernel#open method allows you to open files for reading and writing (and, with addition of the open-uri library, HTTP sockets as well). What you may not know is that Kernel.open can also open processes as if they were files.

  puts "4a. Kernel#open with |"
  cmd = %Q<|#{RUBY} -r#{THIS_FILE} -e 'hello("open(|)", true)'>
  open(cmd, 'w+') do |subprocess|
    subprocess.write("hello from parent")
    subprocess.close_write
    subprocess.read.split("\n").each do |l|
      puts "[parent] output: #{l}"
    end
    puts
  end
  puts "---"

By passing a pipe (“|”) as the first character in the command, we signal to open that we want to start a process, not open a file. For a command, we’re starting another Ruby process and calling our trusty hello method (see the first article or the source code for this article for the definition of the hello method RUBY and THIS_FILE constants).

open yields an IO object which enables us to communicate with the subprocess. Anything written to the object is piped to the process’ STDIN, and the anything the process writes to its STDOUT can be read back as if reading from a file. In the example above we write a line to the child, read some text back from the child, and then end the block.

Note the call to close_write on line 5. This call is important. Because the OS buffers input and output, it is possible to write to a subprocess, attempt to read back, and wait forever because the data is still sitting in the buffer. In addition, filter-style programs typically wait until they see an EOF on their STDIN to exit. By calling close_write, we cause the buffer to be flushed and an EOF to be sent. Once the subprocess exits, its output buffer wil be flushed and any read calls on the parent side will return.

Also note that we pass “w+” as the file open mode. Just as with files, by default the IO object will be opened in read-only mode. If we want to both write to and read from it, we need to specify an appropriate mode.

Here’s the output of the above code:

4a. Kernel#open with |
[child] Hello, standard error
[parent] output: [child] Hello from open(|)
[parent] output: [child] Standard input contains: "hello from parent"

---

Another way to open a command as an IO object is to call IO.popen:

  puts "4b. IO.popen"
  cmd = %Q<#{RUBY} -r#{THIS_FILE} -e 'hello("popen", true)'>
  IO.popen(cmd, 'w+') do |subprocess|
    subprocess.write("hello from parent")
    subprocess.close_write
    subprocess.read.split("\n").each do |l|
      puts "[parent] output: #{l}"
    end
    puts
  end
  puts "---"

This behaves exactly the same as the Kernel#open version. Which way you choose to use is a matter of preference. The IO.popen version arguably makes it a little more obvious what is going on.

Method #5: Forking to a pipe

This is a variation on the previous technique. If Kernel#open is passed a pipe followed by a dash (“|-“) as its first argument, it starts a forked subprocess. This is like the previous example except that instead of executing a command, it forks the running Ruby process into two processes.

  puts "5a. Kernel#open with |-"
  open("|-", "w+") do |subprocess|
    if subprocess.nil?             # child
      hello("open(|-)", true)
      exit
    else                        # parent
      subprocess.write("hello from parent")
      subprocess.close_write
      subprocess.read.split("\n").each do |l|
        puts "[parent] output: #{l}"
      end
      puts
    end
  end
  puts "---"

Both processes then execute the given block. In the child process, the argument yielded to the block will be nil. In the parent, the block argument will be an IO object. As before, the IO object is tied to the forked process’ standard input and standard output streams.

Here’s the output:

5a. Kernel#open with |-
[child] Hello, standard error
[parent] output: [child] Hello from open(|-)
[parent] output: [child] Standard input contains: "hello from parent"

---

Once again, there is an IO.popen version which does the same thing:

  puts "5b. IO.popen with -"
  IO.popen("-", "w+") do |subprocess|
    if subprocess.nil?             # child
      hello("popen(-)", true)
      exit
    else                        # parent
      subprocess.write("hello from parent")
      subprocess.close_write
      subprocess.read.split("\n").each do |l|
        puts "[parent] output: #{l}"
      end
      puts
    end
  end
  puts "---"

Applications and Caveats

The techniques we’ve looked at in this article are best suited for “filter” style subprocesses, where we want to feed some input to a process and then use the output it produces. Because of the potential for deadlocks mentioned earlier, they are less suitable for running highly interactive subprocesses which require multiple reads and responses.

open/popen also do not give us access to the subprocess’ standard error (STDERR) stream. Any output error generated by the subprocesses will print the same place that the parent process’ STDERR does.

In the upcoming parts of the series we’ll look at some libraries which overcome both of these limitations.

Conclusion

In this article we’ve explored two (or four, depending on how you count it) built-in ways of starting a subprocess and communicating with it as if it were a file. In part 3 we’ll move away from built-ins and on to the facilities provided in Ruby’s Standard Library for starting and controlling subprocesses.