pure delayed gives wrong result for dataclass methods #8376
Description
What happened:
from dask import delayed
from dataclasses import dataclass, field
@dataclass(frozen=True)
class A:
param: float = field(repr=False)
def get_param(self):
return self.param
def get_delayed_param(self, *args, **kwargs):
return delayed(self.get_param, pure=True)(*args, **kwargs)
(A(1).get_delayed_param() - A(0).get_delayed_param()).compute()
Out[2]: 0This is an incorrect result.
Apparently, A(1).get_delayed_param().key is erratically the same as A(0).get_delayed_param().key. It seems like the tokenize fails and falls back on something that gives the same result for both as str(A(1)) == str(A(0)).
What you expected to happen:
The correct result is obtained without delayed:
A(1).get_param() - A(0).get_param()
Out[3]: 1One also gets the correct result when not using dataclasses
class B:
def __init__(self, param):
self.param = param
def __repr__(self):
return 'B()'
def get_param(self):
return self.param
def get_delayed_param(self, *args, **kwargs):
return delayed(self.get_param, pure=True)(*args, **kwargs)
(B(1).get_delayed_param() - B(0).get_delayed_param()).compute()
Out[4]: 1Here, we get the correct result even though str(B(1)) == str(B(0)).
Environment:
- Dask version: 2021.10.0
- Python version: 3.9.4.final.0
- Operating System: Windows 10
- Install method (conda, pip, source): pip