argparse set_defaults on subcommands should override top level set_defaults

[bethard]

BPO	9351
Nosy	@ezio-melotti, @bitdancer, @akheron, @remram44, @shihai1991
Files	issue9351.patch

^{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:

assignee = None
closed_at = None
created_at = <Date 2010-07-23.14:06:50.789>
labels = ['type-bug', 'library']
title = 'argparse set_defaults on subcommands should override top level set_defaults'
updated_at = <Date 2021-10-27.21:44:47.104>
user = 'https://bugs.python.org/bethard'

bugs.python.org fields:

activity = <Date 2021-10-27.21:44:47.104>
actor = 'paul.j3'
assignee = 'none'
closed = False
closed_date = None
closer = None
components = ['Library (Lib)']
creation = <Date 2010-07-23.14:06:50.789>
creator = 'bethard'
dependencies = []
files = ['27795']
hgrepos = []
issue_num = 9351
keywords = ['patch']
message_count = 21.0
messages = ['111324', '173602', '174147', '174206', '174208', '206993', '219332', '229619', '229620', '229968', '229983', '230056', '230084', '233041', '244768', '244786', '274369', '342154', '350638', '373665', '405130']
nosy_count = 12.0
nosy_names = ['bethard', 'ezio.melotti', 'r.david.murray', 'nailor', 'python-dev', 'petri.lehtinen', 'paul.j3', 'mikn', 'remram', 'Changaco', 'smparkes', 'shihai1991']
pr_nums = []
priority = 'normal'
resolution = None
stage = 'needs patch'
status = 'open'
superseder = None
type = 'behavior'
url = 'https://bugs.python.org/issue9351'
versions = ['Python 2.7', 'Python 3.4', 'Python 3.5']

[bethard]

If you use set_defaults on a subparser, but a default exists on the top level parser, the subparser defaults are ignored:

>>> parser = argparse.ArgumentParser()
>>> xparser = parser.add_subparsers().add_parser('X')
>>> parser.set_defaults(foo=1)
>>> xparser.set_defaults(foo=2)
>>> parser.parse_args(['X'])
Namespace(foo=1)

This is counter to what people probably expect, that the subparser, when selected, would override the top level parser.

The behavior is because of the following code in parse_known_args:

        for dest in self._defaults:
            if not hasattr(namespace, dest):
                setattr(namespace, dest, self._defaults[dest])

This happens before the subparser sees the namespace object, and so the subparser sees that no defaults need to be filled in.

[nailor]

I've attached a proposed fix where the subparser parses to an empty namespace and that namespace is then merged with the parent namespace

[akheron]

+        for key in vars(subnamespace):
+            setattr(namespace, key, getattr(subnamespace, key))

There might be even more clever ways to achieve this, but what about at least saying "for key, value in vars(subnamespace).items()", and then using the value accordingly, to avoid the getattr() call?

[nailor]

Yeah, I tried figuring out something more clever, as this, in the current form, has a bit too hackish feeling in it, but I couldn't find a proper tool for the job.

Anyway, attached a patch with the getattr removed.

[akheron]

LGTM. Steven?

[mikn]

Just wanted to drop in here to let you know that I hit this behaviour recently when writing a small tool using both configparser and argparse and the workaround proves rather annoying (custom namespace object or similar).

Would be awesome if this moved forward with the proposed patch!

[paulj3]

This is not getting much attention for several reasons:

• there's quite a backlog of argparse patches and issues

• 'set_defaults' is not commonly used. Setting default in 'add_argument' seems more common.

• defining the same argument for both the parser and subparser can create other difficulties.

• in Bethard's example, 'set_default' sets an attribute that has no connection to any argument. It's a cleaver trick that few users will think to use, and probably won't be of much value.

The idea proposed here of using a subnamespace for the subparser is interesting. It reminds me of a StackOverflow question about implementing nested namespaces.

http://stackoverflow.com/questions/18668227

[python-dev]

New changeset e9cb45ccf42b by R David Murray in branch '3.4':
bpo-9351: set_defaults on subparser is no longer ignored if set on parent.
https://hg.python.org/cpython/rev/e9cb45ccf42b

New changeset b35a811d4420 by R David Murray in branch 'default':
Merge: bpo-9351: set_defaults on subparser is no longer ignored if set on parent.
https://hg.python.org/cpython/rev/b35a811d4420

New changeset 1a3143752db2 by R David Murray in branch '2.7':
bpo-9351: set_defaults on subparser is no longer ignored if set on parent.
https://hg.python.org/cpython/rev/1a3143752db2

[bitdancer]

Thanks, Jyrki.

[paulj3]

A possible problem with this patch is that it overrides any Namespace given by the user. In the test example:

    parser.parse_args(['X'], namespace=argparse.Namespace(foo=3))

the result is 'foo=2', the setdefault from the subparser. Previously, and with a single level parser, the result would be 'foo=3'. This is also true for an ordinary argument default in the subparser.

It could be argued that a user is unlikely to use both the 'namespace=...' parameter, and 'setdefault' for the same variable, especially one that isn't an argument 'dest'. But the fact that the custom namespace does not override regular subparser argument defaults bothers me.

Also, should an untested change like this be applied to 3.4 and 2.7? This is not a backward compatible bug fix.

The test case also touches on a real bug in the recent releases - subparsers are now optional. 'parser.parse_args([])' runs without complaint about the missing 'X'. Actually it is convenient for testing this case. But it is still an incompatibility with earlier behavior

http://bugs.python.org/issue9253

[bitdancer]

Yeah, I hesitated a bit about the backports, but didn't visualize the scenario you are suggesting and thought it would be safe. Perhaps it should be backed out of 2.7 and 3.4.

For 3.5, do you have any thoughts about how to make namespace= play nicely with this problem?

[paulj3]

I'm exploring modifying that patch by adding a 'subnamespace' attribute to the subparsers Action object. This would give the user, or the main parser control over the namespace that is passed to a subparser.

For example:

        subnamespace = getattr(self, 'subnamespace', False)
        if subnamespace is None:
            pass
        elif not  subnamespace:
            subnamespace = namespace
        # if None, subparser will use a fresh one
        subnamespace, arg_strings = parser.parse_known_args(arg_strings, subnamespace)
        for key, value in vars(subnamespace).items():
            setattr(namespace, key, value)

As written here, no value (or any False) would use the main parser namespace, as done with existing code.

A 'None' value, would use the a new empty namespace, as proposed in this patch.

Or the user could directly define the namespace, e.g.

    sp = parser.add_subparsers()
    sp.subnamespace = Namespace(foo=4)

That could also be convenient if the user is using a custom Namespace class.

'parse_known_args' could also set this attribute, based on whether the user has given it a namespace or not.

[bitdancer]

If I understand you correctly, that would mean that if the namespace keyword is not used, we'd have the fixed behavior, but if the namespace keyword is used, we'd have the backward compatible behavior? If I'm understanding correctly, that sounds like a good solution to me (coupled with backing this out of the maint versions).

[Changaco]

The broken patch made it into the 2.7.9 release, causing argparse to silently ignore arguments, cf http://bugs.python.org/issue23058

[paulj3]

Another example of this patch causing backward compatibility problems

http://bugs.python.org/issue24251

[remram44]

To me this is much more than a compatibility problem. The way it worked before made a lot of sense, and just felt like the "correct" solution to accept a flag in multiple places.

Having a --verbose flag is something everybody should consider (Python has a decent builtin logging module), and anybody providing it would definitely want to accept it before and after subcommands (or at least, for every subcommand).

The only way right now is to not only create different arguments with add_argument(), for each parser, but you also need to provide different destination names (and then do something shitty like verbosity = args.verb_main+args.verb_subcommand). This bug makes argparse completely unusable for any real-life application that uses subparsers (in addition to breaking existing programs). And it breaks silently too, simply amazing!

Of course there is very little point in fixing this now. Since this affects multiple released versions of Python, I have to use a work-around anyway (until I can move from argparse to something that won't decide to break someday for the hell of it).

[paulj3]

In http://bugs.python.org/issue27859 I've demonstrated how a subclass of _SubParsersAction can be used to revert the behavior to pre 2.7.9, passing the main namespace to the subparser.

The only other change is to the parser registry:

    parser.register('action', 'parsers', MyParserAction)

So there's no need to change the argparse.py file itself.

[paulj3]

A variation on the problem I reported in

https://bugs.python.org/issue9351#msg229968

is that a custom Namespace class as documented in

https://docs.python.org/3/library/argparse.html#the-namespace-object

will not be used by the subparsers. Only the main parser uses that custom Namespace object; the subparsers continue to use the default 'argparse.Namespace()'.

[shihai1991]

How about use a flag(such USING_OUT_NAMESPACE) to identify we use namespace or not?

For example:

subnamespace, arg_strings = parser.parse_known_args(arg_strings, None)
for key, value in vars(subnamespace).items():
    if USING_OUT_NAMESPACE and not hasattr(namespace, key):
        setattr(namespace, key, value)

[paulj3]

I just realized if the subparser argument used

default=argparse.SUPPRESS

it would not clobber values (default or user) set by the main parser.

(a 'store_true' would have to be replaced with a 'store_const' to do this)

[paulj3]

A new patch, https://bugs.python.org/issue45235 has clobbered this patch.

It has also exposed the inadequate unittesting for the case(s) where the 'dest' of main namespace, subparser namespace, user provided namespace overlap.

[serhiy-storchaka]

This change looks very wrong to me. It breaks several implicit rules.

• Command line arguments take precedence over default values. But defaults in subparser override now user specified values for the main parser.
• Namespace values take precedence over default values. But defaults in subparser override now namespace values.

The order of precedence between defaults of the main parser and subparser is a nuance. Of course, there are arguments for making a subclass defaults having higher priority than the main parser defaults, but if it is impossible to implement, this is just a documentation issue.

I tried to solve this issue in other way, by setting default values after parsing, when they are not set by actions and was not pre-populated in a namespace. This solves the original issue and many issues caused by previous solution, but it has its own cost. The default values no longer set in a namespace when user actions are called. This is impossible, it conflicts with the requirement of this issue. This breaks one of existing tests.

On other hand, it opens a way for more interesting issue -- lazily evaluated default values.

We may add an option to chose between these behavior. I need to investigate what other behavior is changed.