Inconsistent datetime Values with Fixed random_seed in Polyfactory - Expected Behavior?

[saygox]

Hi,
I am grateful for Polyfactory, which has significantly simplified my testing process. I often use a plugin called Syrupy for my tests, which takes snapshots and checks for differences from previous results.

During testing, I noticed that even with a fixed random_seed, the datetime type values were different. I am curious if this is the expected behavior and would appreciate it if we could align our understanding on this.

Below is my testing program:

from datetime import datetime
from pydantic import BaseModel
from polyfactory.factories.pydantic_factory import ModelFactory

def test_datetime(snapshot):
    class A(BaseModel):
        d: datetime

    class A_Factory(ModelFactory):
        __model__ = A
        __random_seed__ = 0

    data = A_Factory.batch(10)

    assert [_.d for _ in data] == snapshot()

Excerpt from the test error:

>       assert [_.d for _ in data] == snapshot()
E       assert [+ received] == [- snapshot]
E           list([
E         -   datetime.datetime(1998, 2, 23, 1, 59, 13),
E         +   datetime.datetime(1998, 2, 23, 1, 59, 16),
E         -   datetime.datetime(2012, 12, 10, 1, 57, 26),
E         +   datetime.datetime(2012, 12, 10, 1, 57, 29),
E         -   datetime.datetime(2022, 5, 29, 19, 14, 51),
E         +   datetime.datetime(2022, 5, 29, 19, 14, 54),

The cause of the changing test results is that the function used to generate the datetime type values is Faker's date_time_between. As discussed in Faker's Issue, this is understood to be intentional, as the difference from the current time (time_delta) is fixed. Here is the relevant link: joke2k/faker#1791.

However, whether it is desirable for the value to change each time with a fixed random_seed for datetime in Polyfactory is a separate issue that I'd like to discuss.

---

There are two methods to standardize the test with the current implementation. One is to fix the current time by adding the freezegun module, and the other is to override the value generation process with get_provider_map.

Method using the freezegun module:

from datetime import datetime
from pydantic import BaseModel
from polyfactory.factories.pydantic_factory import ModelFactory
import freezegun


def test_fz(snapshot):
    with freezegun.freeze_time('2015-10-21 12:34:56'):
        class A(BaseModel):
            d: datetime

        class A_Factory(ModelFactory):
            __model__ = A
            __random_seed__ = 0

        data = A_Factory.batch(10)

    assert [_.d for _ in data] == snapshot()

Overriding the value generation process with get_provider_map:

from datetime import datetime
from pydantic import BaseModel
from polyfactory.factories.pydantic_factory import ModelFactory
from typing import Any, Callable


def test_providermap(snapshot):
    class A(BaseModel):
        d: datetime

    class A_Factory(ModelFactory):
        __model__ = A
        __random_seed__ = 0

        @classmethod
        def get_provider_map(cls) -> dict[Any, Callable[[], Any]]:
            provider_map = super().get_provider_map()
            provider_map.update(
                {
                    datetime: cls.__faker__.date_time,
                }
            )
            return provider_map

    data = A_Factory.batch(10)
    assert [_.d for _ in data] == snapshot()

I look forward to your feedback.

[Goldziher]

Hi,

I think it's fine to add a datetime value Generators module on our end, and stop relying on faker for this purpose.

[saygox]

Thank you for your prompt response.

I understand that opting to discontinue reliance on Faker for datetime value generation is a valid option.
However, I would like to point out that the Faker's date_time() function serves our purpose well.
I'm looking forward to seeing how we can balance these factors moving forward.

Thanks again for your time and consideration.

[Goldziher]

I'm open to ideas.