💾 Archived View for dmerej.info › blog › 0022-docopt-v-argparse.gmi captured on 2022-04-28 at 17:29:22. Gemini links have been rewritten to link to archived content
⬅️ Previous capture (2021-11-30)
-=-=-=-=-=-=-
2016, Sep 24 - Dimitri Merejkowsky License: CC By 4.0
Say you have a program named `foo` that is supposed to be used from the command line, written in Python.
It provides lots of various features, so you decide to group them into subcommands, like `git` or `mercurial` does for instance.
So your program can be called in the following ways:
$ foo bar --baz --num-jobs=4
runs bar(baz=True, num_jobs=4)
$ foo read file.in --verbose
runs read("file.in", verbose=True)
Assuming you are using Python for this, which library should you use to implement command line parsing?
Back in the old days, the answer was getopt
Basically, you would call the `getopt()` function with some weird string, like this:
"""
Usage: foo [options]
Options:
-h, --help Display this help message
-o, --output <PATH> Output path
-v,--verbose Be verbose
"""
import sys
import getopt
def usage():
print(__doc__)
def main():
try:
opts, args = getopt.getopt(sys.argv[1:], "ho:v", ["help", "output="])
except getopt.GetoptError as err:
# print help information and exit:
sys.exit(err) # will print something like "option -a not recognized"
for o, a in opts:
if o in ("-h", "--help"):
usage()
sys.exit(0)
elif o in ("-o", "--output"):
...
Here `ho:v` means you recognize the short options `-h`, `-o` and `-v`, and the colon after the `o` means that `-o` must be followed by an argument (so `foo -o -v` does not work, but `foo -o foo.out` does) Note that you have to use a list of strings to have long options recognized too. (Here `-h` and `--help` will have the same meaning)
Note the call to the `usage()` function. The *usage* of a command line program is simply a string that describes how the program should be called.
Also note how you specify `-o,--output` **three** times: once in the docstring, once it the `getopt()` call and finally once in the `main()` function.
Then came optparse[1].
1: https://docs.python.org/3/library/optparse.html
The idea was to get rid of the weird syntax to define long and short options, and also to generate the help message and usage automatically.
To use `optparse` you create a *parser*, then call the `add_option` method for each pair of long and short options you want to support.
Then you invoke the `parse_args()` method and it returns a tuple of two items:
The first item is a class (`optparse.Values`) that contains the values of the options as attributes.
The second item is simply the list of positional arguments.
Note the difference between *options* and *arguments* here.
When using `optparse`, everything that is optional starts with a dash (`-`), and is called an *option*, and everything that is required simply comes after all the options and is called a *positional argument*, or *argument* for short.
There's no way to have "optional arguments" or "required options".
The `optparse` authors explain why they choose to enforce this behavior in the documentation:
It contains some interesting points about command line API design, so I suggest you read it, even if you won't be using `optparse`.
2: https://docs.python.org/3/library/optparse.html#what-are-options-for
3: https://docs.python.org/3/library/optparse.html#what-are-positional-arguments-for
By default, the attribute name of the parsed option will be computed from its long form.
`--foo-bar=42` gives `opts.foo_bar = 42`.
But you can use `add_option("--input", dest="input_file"`) to override the name of the attribute.
If the option does not take an argument, its parsed value will be a boolean, and you can use `action="store_true"` or `action="store_false"`. You can set its default value by calling `parser.set_defaults()`.
So, if you have a `foo` program that takes a required input file as argument, and an output file as an option, you'll have to write something like:
parser = OptionParser()
parser.add_option("-o", "--output", help="path to the output file")
opts, args = parser.parse_args()
if len(args) != 1:
sys.exit("Missing one required argument")
input_file = args[0]
if opts.output:
output_file = opts.output
else:
output_file = input_file + ".out"
`optparse` has a lot of nice features:
argparse[4] is a direct successor of `optparse`.
4: https://docs.python.org/3/howto/argparse.html
The idea is to be more flexible that `optparse` and allowing to have "optional arguments" and "required options", because, let's face it, sometimes it's really convenient. For instance, for a dangerous operation, you may want the `--force` option to be required :P
Anyway, `argparse` uses the same API that `optparse`, except that:
Porting from `optparse` to `argparse` is thus quite easy. For instance, the previous example becomes:
parser = ArgumentParser()
# it's an option because it starts with a dash:
parser.add_argument("-o", "--output", help="path to the output file")
# it's a *required* positional argument
parser.add_argument("input", dest="input_file")
args = parser.parse_args()
input_file = args.input_file
if args.output:
output_file = args.output
else:
output_file = input_file + ".out"
But you can have optional positional arguments, by using `add_argument("foo", nargs="?")`, or required options, with `add_argument("--force", action="store_true", required=True)`
Also, `argparse` supports what are called *subparsers*, which means you can have complex API that runs different *commands* depending on the first argument. You do this by calling `parser.add_subparsers()`:
parser = argparse.ArgumentParser()
action_parsers = parser.add_subparsers(title="available actions")
bar_parser = action_parsers.add_parser("bar")
...
read_parser = action_parsers.add_parser("read")
...
This makes it possible to implement the example we started with:
$ foo bar --baz --num-jobs=4 $ foo read file.in --verbose
A nifty hack is to use `set_defaults()` to dispatch the code after parsing, like this:
bar_parser.set_defaults(action="do_bar")
...
read_parser.set_defaults(action="do_read")
...
if args.action == "do_bar":
...
elif args.action == "do_read":
...
By the way, `argparse` started its life as a standalone third-party library, but is now part of Python standard library since Python 2.7.
The official Python documentation for `getopt` and `optparse` both suggest to use `argparse` instead.
`docopt` came with a quite interesting new idea. What if, instead of generating the help message automatically, we *parsed* directly the usage string?
After all, the string usage does have a pretty normalized format. (It's even part of the "IEE Std 1003.1" standard, also knows as ... POSIX)
Instead of writing tons of boring boilerplate (instantiate the parser, call `add_argument` a bunch of times, then `parse_args`), what if you could just call a function using the usage string?
This is how you use `docopt`:
""" Usage: foo bar [--baz] [--num-jobs=<jobs>] foo read <input_file> [--verbose] Options: --baz Do some baz stuff --verbose Be verbose --num-jobs=<jobs> number of jobs [default: 1] """ import docopt opts = docopt.docopt(__doc__) print(opts)
Note that the return value of `docopt` is a simple Python dictionary.
The keys will be `'--baz'`, `'baz'`, or `'<baz>'` depending on the context, and the values will always be strings, except when a subcommand is used.
So for instance you would have:
$ foo bar --num-jobs=42
{'--baz': False,
'--num-jobs': '42',
'--verbose': False,
'<input_file>': None,
'bar': True,
'read': False}
% foo read input.txt
{'--baz': False,
'--num-jobs': '1',
'--verbose': False,
'<input_file>': 'input.txt',
'bar': False,
'read': True}
OK, so by now it should be pretty clear your best option is between `argparse` and `docopt`
Note: there's also other projects like clik[5] that do things differently, but:
Kind of obvious, right? All you have to do is import `docopt` and call `docopt.docopt()` with the usage string.
6: http://xion.io/post/programming/python-dont-use-click.html
The PEP 257[7] says:
7: https://www.python.org/dev/peps/pep-0257
The docstring of a script (a stand-alone program) should be usable as its
"usage" message, printed when the script is invoked with incorrect or missing
arguments (or perhaps with a "-h" option, for "help"). Such a docstring should
document the script's function and command line syntax, environment variables,
and files. Usage messages can be fairly elaborate (several screens full) and
should be sufficient for a new user to use the command properly, as well as a
complete quick reference to all options and arguments for the sophisticated
user.
`docopt` is perfect for this.
You can do this with `argparse` of course, but doing so is not trivial. You have to implement an `argparse` "formatter class", and the names are ugly:
argparse.RawDescriptionHelpFormatter argparse.RawTextHelpFormatter argparse.ArgumentDefaultsHelpFormatter argparse.MetavarTypeHelpFormatter
Ugh.
You can also prevent `argparse` from generating the help message automatically, but then you have to re-implement handling `-h,--help` yourself.
Usage: my_program (--either-this <and-that> | <or-this>)
There isn't any way to do this directly with argparse.
You'll have to do it manually, like so:
parser = argparse.ArgumentParser()
parser.add_argument("--either-this")
parser.add_argument("or_this", nargs="?")
args = args.parse_args():
if not any((args.either_this, args.or_this)):
sys.exit("...")
With `docopt`:
Usage: foo go (--up | --down)
With `argparse`, you must use `add_mutually_exclusive_group()` which is quite verbose ...
If you configure your `argparse.ArgumentParser` from several portions of your code, you risk defining the same option twice, which will lead to this kind of errors:
argparse.ArgumentError: argument -f/--format: conflicting option string: -f
and this gets old fast ...
Although the reference implementation of `docopt` is in Python, the library has been ported to quite a few languages.
The official list is on github
Actually, this has both pros and cons.
`docopt` is able to do its magic because it uses a DSL (domain specific language)
That's why you only need a few lines of code to use it, because the information you need is parsed from an other language, dedicated to this task.
Quite ironically, this is the same approach used by `getopt`. The weird `"ho:v"` string is *also* a DSL ...
But using a DSL comes with a few cons:
For instance, you need *2* spaces between the option name and its description for `docopt` to work:
foo --this-is-long Long description --short Do it short
It's hard to see where the problem is here ...
Also, if `docopt` cannot parse your usage string, you'll get a cryptic error message with no clue about what line is wrong:
""" Usage: naval_fate ship new <name>... naval_fate ship <name> move <x> <y> [--speed=<kn>] naval_fate mine (set|remove <x> <y> [--moored|--drifting] naval_fate ship shoot <x> <y> """ from docopt import docopt opts = docopt(__doc__)
$ python naval_fate.py
docopt.DocoptLanguageError: unmatched '('
That's what you get when creating a new language: you have to take care of many subtle details and problems that comes with it...
See an article from ESR[8] for some discussion around this topic.
8: http://esr.ibiblio.org/?p=7032
But this is a bit philosophical, let's get to more concrete stuff.
`docopt` will *always* print the usage if it can't parse the command line, with no explanation.
`argparse` will generate much more information:
$ naval_fate move invalid choice: 'move' (choose from 'ship', 'mine') $ naval_fate ship new the following arguments are required: names $ naval_fate ship move bounty 5 fortytwo argument y: invalid float value: 'fortytwo'
There's a Pull Request[9] to fix this, though.
9: https://github.com/docopt/docopt/pull/63
Yes, this is a contrived example:
if os.name == "nt":
parser.add_argument("--on-mingw")
Still, you get the idea: since you're using Python to build your parser, you have access to the full features of the language, instead of being constrained by the DSL.
I really like the fact I have a two possible "names" : one for command line API, and an other one for the attribute of the returned value.
For instance, here is a pattern I often use:
def do_foo(raises=True):
ok = try_something()
if not ok and raises:
raise FooFailed()
Using `argparse`, I can do:
parser.add_argument("--ignore-errors", dest="raises")
args = parser.parse_args()
do_foo(raises=args.raises) # I really don't care what the flag is called here
# for all I know, it could be --i-know-what-i-am-doing
With `docopt`, I would have to write:
opts = docopt.docopt(__doc__) do_foo(raises=opts["--ignore-errors"])
which I find less readable, but maybe it's a matter of taste.
An other example is when you need a `--no-foo` option.
I prefer using:
parser.add_argument("--no-foo", dest="foo", action="store_false")
parser.set_defaults(foo=True)
...
if args.foo:
...
rather than `if not opts["--no-foo"]:`
With `docopt`, all you get is a dictionary with strings as keys and values. (And sometimes booleans when you have subparsers)
With argparse, you can force values to be integers, floats, existing files, or even your own type.
For instance:
def valid_filename(value)
if value in [".", ".."]:
raise Error("Invalid name: %s" % value)
bad_chars = r'<>:"/\|?*'
for bad_char in bad_chars:
if bad_char in value: