Closed kgreav closed 3 years ago
Hi, thanks for this report. I wonder if it wouldn't it be enough to simply change line 85 to something like
return type.__new__(DataClassInit if options['init'] and user_init else mcs, name, bases, dict_)
This is optimal because it avoids the definition of a slow Python __call__
.
Apologies, I am away for exams and will likely not be able to test and/or merge for a while.
No worries and no rush, I'll test it out - already set up everything in my repo so I'm not beholden to the PR being merged.
OK, after testing your solution is equivalent to mine in function but optimal, however I realised there is one breaking change (applicable to both versions). Personally I don't think it is a bad breaking change as the usage was quite misleading.
In the existing, the following works:
@dataclass(init=False)
class Obj(object):
a: str
b: str
def __init__(self, c, d):
self.a = c
self.b = d
o = Obj("x", "y", "w", "z")
assert o.a = "w"
assert o.b = "z"
whereas in the new version this will fail, and would need to be changed to:
Obj("w", "z")
This is obviously a naive example, and someone could have more contrived examples that make sense, however the worst case would be they'd have to add their classes attributes to their __init__
function which I think is desirable anyway, as from what I've seen you can't currently get a meaningful constructor signature out of this example unless you document it, as it's an amalgamation of __new__
and __init__
.
I think it's fair to call the behaviour you discovered a bug - init
is false and yet the arguments its function would have had are still being consumed. This was an oversight on my part - there is absolutely no reason to use a custom __call__
to modify arguments if there isn't both __init__
and __new__
on the class. If someone is using this obscure behaviour, for some reason, it is a breaking change - but we're still on 0.x
releases so what can they expect!
I took a few moments to push the fix for this. Any problems, just let me know.
Fix released in v0.6.2. Also, apologies, I didn't realise you had incorporated my solution into your PR before I unceremoniously closed it.
Not a problem at all, the quick turnaround more than makes up for it - I didn't get around to fixing/writing tests anyway. Thanks a bunch!
Great work on this lib so far. Much more intuitive than dataclasses and even solves some bugs (e.g in dataclasses you can't define a post_init parameter with the same name as a method on the class).
The only issue I've run into so far is that if you specify
init=False
but you have a custom__init__
, you can't create your class with positional arguments, and you have to always specify keyword arguments. Seems to me that we should just pass all arguments straight through to the user defined__init__
wheninit=False
, and just call__new__
with no args.I've done a PR (#9) that fixes this issue I think without breaking any existing usage, but keen to get your thoughts.