Repository navigation
argparse: make new 'required' argument to add_subparsers default to False instead of True #77290
Description
Activity
wm75 commented
on Mar 20, 2018 wm75mannequinMannequinAuthorI find the True default for 'required' quite cumbersome introduced as a result of bpo-26510.
With existing parsers it can unnecessarily break compatibility between Python3.x versions only to make porting a bit easier for Python2 users.
I think, this late in the life cycle of Python2, within Python3 compatibility should be ranked higher than py2to3 portability.Command line parsing of a package of mine has long used optional subparsers (without me even thinking much about the fact). Now in 3.7, running
python3.7 -m MyPackage
without arguments (the parser is in __main__.py) I get the ill-formatted error message:
__main__.py: error: the following arguments are required:
while my code in 3.3 - 3.6 was catching the empty Namespace returned and printed a help message.
Because the 'required' keyword argument did not exist in < 3.7 there was no simple way for me to write code that is compatible between all 3.x versions. What I ended up doing now is to check sys.argv before trying to parse things, then print the help message, when that only has a single item, just to keep my existing code working.
OTOH, everything would be just fine with a default value of False.
Also that truncated error message should be fixed before 3.7 gets released.asottile commented
on Mar 20, 2018 asottilemannequinMannequinThe intention of the change in bpo-26510 was to pick the least surprising behaviour for the default value of subparsers -- the compatiblity with the behaviour before the regression was introduced in 3.3 was a nice side-effect. As with the rest of positional arguments in argparse, the positional subparsers were changed to required by default.
The main issue addressing the 3.3 regression I believe is https://bugs.python.org/issue9253 and not the one linked.
When I revived the patch, I surveyed a number of open source tools using subparsers (~10-20) and they all fell into the following categories:
- Used the workaround (part of this SO post: https://stackoverflow.com/a/23354355/812183) (most fell into this category)
- crashed with some sort of TypeError (NoneType has no attribute startswith, KeyeError: None, etc.) due to not handling "optional" subparsers
- Manually handled when the subparser was
Noneto raise an argparse error
You can enable a 3.3-3.7 compatible "always optional subparsers" with a similar pattern that was used to manually restore the pre-regression behaviour:
subparsers = parser.add_subparsers(...) subparsers.required = False
I believe the error message issue is already tracked: https://bugs.python.org/issue29298
asottile commented
on Mar 20, 2018 asottilemannequinMannequinGrabbed the wrong SO link, I believe this is the one I meant to link to: https://stackoverflow.com/a/18283730/812183
wm75 commented
on Mar 20, 2018 wm75mannequinMannequinAuthorOn 03/20/2018 04:38 PM, Anthony Sottile wrote:
Anthony Sottile <asottile@umich.edu> added the comment:
The intention of the change in bpo-26510 was to pick the least surprising behaviour for the default value of subparsers -- the compatiblity with the behaviour before the regression was introduced in 3.3 was a nice side-effect. As with the rest of positional arguments in argparse, the positional subparsers were changed to required by default.
Since the 3.3 change happened a long time ago and has been kept through
the next three releases, I never considered it a regression, but rather
thought the original behavior was an oddity of early Python 3s (I've
never written an argparse-based parser in Python 2), so it's interesting
to see this in the larger historical context.Overall, I think "least surprising" is in the eye of the beholder here
and I want to emphasize that I'm all for your change of adding the
keyword argument, just not so convinced about your choice of the default.My main argument for a default of False and against True: having True as
the default will only lead people used to the Python 2 and pre-3.3 way
of things to think that they have code working for Python 3, when, in
fact, that code will break under 3.3-3.6, and, at least, 3.6 will stay
in widespread use for a long time (even Ubuntu 18.04 will still ship
with it as the default python3, so Python3.6 will outlast the life cycle
of Python 2 by a good measure).
Conversely, most projects are dropping Python 3.2 support these days or
have done so already, so nobody really cares about how things worked in
this version (I think it's telling along these lines that your -
corrected - SO link dates back from 2013). So I think, it is
a rather unnecessary incompatibility you are introducing for project
maintainers even if the original issue was a regression.The main issue addressing the 3.3 regression I believe is https://bugs.python.org/issue9253 and not the one linked.
When I revived the patch, I surveyed a number of open source tools using subparsers (~10-20) and they all fell into the following categories:
- Used the workaround (part of this SO post: https://stackoverflow.com/a/23354355/812183) (most fell into this category)
- crashed with some sort of TypeError (NoneType has no attribute startswith, KeyeError: None, etc.) due to not handling "optional" subparsers
- Manually handled when the subparser was
Noneto raise an argparse error
As yet another option, and similar to the third one on your list, I'm
using the set_defaults method of the subparser, and I'm checking whether
the corresponding key is present in the Namespace.You can enable a 3.3-3.7 compatible "always optional subparsers" with a similar pattern that was used to manually restore the pre-regression behaviour:
subparsers = parser.add_subparsers(...)
subparsers.required = FalseAh, right! That's a good option. Didn't realize it would work this way,
too :)But a still think you should consider my above argument.
I believe the error message issue is already tracked: https://bugs.python.org/issue29298
I see, that looks as if it would fix this part. It would be great if it
could get merged into Python 3.7 still.----------
Python tracker <report@bugs.python.org>
<https://bugs.python.org/issue33109\>
asottile commented
on Mar 20, 2018 asottilemannequinMannequinYeah, I picked the default value
Truebecause I couldn't actually find a user of subparsers that wanted optional subparsers. ¯_(ツ)_/¯wm75 commented
on Mar 20, 2018 wm75mannequinMannequinAuthor_wants_ is a bit a strong word, but, at least, you can do a bit a nicer job than the default error, like printing a nicely formatted list of subcommands as you would get it with the main parsers help.
In fact, in my own code I'm actually catching the missing subparser situation, then calling parse_args again with ['--help'] :)bskinn commented
on May 7, 2018 bskinnmannequinMannequinI second Wolfgang's recommendation to change the default back to
False.I started developing CLI apps &c. in Python ~4yrs ago; I dabbled briefly in 2.7, then switched firmly to Python 3. When I started, I was aimed at supporting 3.3 to 3.5; now I'm specifically supporting 3.4 to 3.6, but starting to test my code against the 3.7b versions.
Thus, all I've ever known is the default
required=Falsebehavior. On my one tool currently using subparsers (https://github.com/bskinn/sphobjinv), I made the subparser choice optional, to allow asphobjinv --versioninvocation. So, when 3.7 failed that test (https://github.com/bskinn/sphobjinv/blob/6c1f22e40dc3d129485462aec05adbed2ff40ab8/sphobjinv/test/sphobjinv_cli.py#L419-L422), it seemed like a regression to me.All that said, given that the
subparsers.required = Falsefix is a clean, single line, I have no beef with aTruedefault if that's preferred. However, please include this change in the 3.7 CHANGELOG, at least.I’m sorry I don’t have the time to study this and make a judgment call.
Bringing this to the release manager’s attention.
If the behavior was consistent from 3.3 through 3.6, that is the behavior we should keep going forward in 3.7+ without a deprecation period. (and this does not seem worth deprecating, lets just keep the behavior the same as it was in 3.3-3.6)
That 2.7 is different than >=3.3 here isn't important. There are a lot of things that have conditional behavior differences when using 2 and over time that is becoming increasingly irrelevant.
Libraries often keep compatibility with both 2.7 and 3.4+ today, but it is less common for a command line tool entry point to need to be compatible with both.
asottile commented
on May 15, 2018 asottilemannequinMannequinAccording to the other bugs, the change in 3.3 was an inadvertent regression. The fact that it didn't get fixed for so long is mostly due to the unmaintained state of argparse in the stdlib. The change in behaviour here is the _fix_ of that regression.
Consistency with the rest of the framework to me feels pretty important. subparsers are positional arguments and positional arguments by default are required.
5 remaining items
I tried to use add_subparsers() with required=True and have found it not usable.
import argparse parser = argparse.ArgumentParser(prog='PROG') subparsers = parser.add_subparsers(required=True) parser_a = subparsers.add_parser('a') parser_b = subparsers.add_parser('b') parser.parse_args([])
The result:
Traceback (most recent call last): File "<stdin>", line 1, in <module> File "/home/serhiy/py/cpython/Lib/argparse.py", line 1745, in parse_args args, argv = self.parse_known_args(args, namespace) File "/home/serhiy/py/cpython/Lib/argparse.py", line 1777, in parse_known_args namespace, args = self._parse_known_args(args, namespace) File "/home/serhiy/py/cpython/Lib/argparse.py", line 2012, in _parse_known_args ', '.join(required_actions)) TypeError: sequence item 0: expected str instance, NoneType found
Seems that not only the default value should be changed, but the whole PR 3027 should be reverted.
asottile commented
on May 22, 2018 asottilemannequinMannequinThat's a separate issue (also a bug introduced by the bad 3.3 patch): https://bugs.python.org/issue29298
I have an open PR to fix it as well but it has not seen review action: #3680
Wouldn't be better to first fix this bug, and only after that add the 'required' parameter?
Adding it introduced yet one bug: when pass arguments as positional, the 'help' argument will be swallowed. You can add new parameters only after existing positional parameters.
asottile commented
on May 22, 2018 asottilemannequinMannequinThe bug is orthogonal, you can trigger it without the
required=keyword argument via the (currently suggested) monkeypatch workaround which restores the pre-3.3 behaviour:import argparse parser = argparse.ArgumentParser() subp = parser.add_subparsers() subp.add_parser('test') subp.required = True parser.parse_args()
$ python3 test.py Traceback (most recent call last): File "test.py", line 7, in <module> parser.parse_args() File "/usr/local/Cellar/python/3.6.5/Frameworks/Python.framework/Versions/3.6/lib/python3.6/argparse.py", line 1730, in parse_args args, argv = self.parse_known_args(args, namespace) File "/usr/local/Cellar/python/3.6.5/Frameworks/Python.framework/Versions/3.6/lib/python3.6/argparse.py", line 1762, in parse_known_args namespace, args = self._parse_known_args(args, namespace) File "/usr/local/Cellar/python/3.6.5/Frameworks/Python.framework/Versions/3.6/lib/python3.6/argparse.py", line 1997, in _parse_known_args ', '.join(required_actions)) TypeError: sequence item 0: expected str instance, NoneType found
Also note that when
destis specified it works fine:import argparse parser = argparse.ArgumentParser() subp = parser.add_subparsers(dest='cmd') subp.add_parser('test') subp.required = True parser.parse_args()
$ python3 test.py usage: test.py [-h] {test} ... test.py: error: the following arguments are required: cmdPlease ignore the last paragraph. It was my mistake, all add_subparsers() parameters are keyword-only, and _SubParsersAction is a privale class.
Is there then no pathway for actually fixing the bug? aka how can I get
required=Trueto be the default.There may very well be but, unfortunately, dealing with this newly-identified 3.x compatibility issue takes precedence for 3.7.0. In general, for cpython, all other things being equal, the status quo wins and, in this case, that means not breaking 3.6->3.7 compatibility without a good reason.
The underlying problem here, IMHO, is that we are essentially an all-volunteer project and at the moment argparse does not have an active core developer to review and champion change requests. If a core developer does want to take up the existing queue of argparse issues, we *might* come to the conclusion that making another incompatible change is the overall right thing to do. But, until that happens, the least bad option is to not make things worse. How to get more core developer interest in argparse issues is a whole 'nother matter and out-of-scope for this issue. Sorry, I wish I had a better answer. Thanks, Anthony and everyone else here, for your input.
Lucas-C commented
on May 21, 2021 Lucas-CmannequinMannequinReporting a duplicate / superseder with the following bug:
parser = argparse.ArgumentParser() subparsers = parser.add_subparsers(required=True) subparsers.add_parser('foo') parser.parse_args()
Raising:
TypeError: sequence item 0: expected str instance, NoneType found
It has already been reported in https://bugs.python.org/issue29298
with an interesting solution by Greg Minshall.Lucas-C commented
on May 21, 2021 Lucas-CmannequinMannequinSorry, the fix was by Mathias Ettinger:
elif isinstance(argument, _SubParsersAction): return '{%s}' % ','.join(map(str, argument.choices))I submitted a PR with this patch and a corresponding unit test:
#26278
Note: these values reflect the state of the issue at the time it was migrated and might not reflect the current state.
Show more details
GitHub fields:
bugs.python.org fields: