[Bug] root logger has FileHandler after using hydra_task_runner

[nai62]

🐛 Bug

Once hydra_task_runner is called from a unit test, a FileHandler seems to be registered to the root logger of logging. When a logger tries to log something from another test, it fails because of the staled FileHandler.

I could successfully make the tests pass by manually removing the FileHandler, i.e., adding getLogger().handlers.pop() at the end of test_1(). But it doesn't seem to be an elegant solution. Probably is it possible to remove additional handlers in TaskTestFunction.__exit__() or somewhere?

To reproduce

from logging import getLogger


def test_1(hydra_restore_singletons, hydra_task_runner):
    with hydra_task_runner(
        calling_file="test_hydra.py",
        calling_module=None,
        config_path=".",
        config_name="config.yaml",
    ):
        logger = getLogger(__name__)
        logger.info('test_1')


def test_2():
    logger = getLogger(__name__)
    logger.info('test_2')

Then run pytest.

** Stack trace/error message **

E       FileNotFoundError: [Errno 2] No such file or directory: '/tmp/[TEMPORARY DIRECTORY]/test_hydra.log'

System information

• Hydra Version : I tried both 1.0.0rc2 and master.
• Python version : 3.8.3

[omry]

Thanks for reporting.
Can you share more info about your what you are using the task_runner fixture for?
Up to this point it was only used by Hydra itself and Hydra plugins.

[nai62]

Actually I wanted to unit test my app that uses hydra. I first tried Compose API and found that hydra.utils.get_original_cwd() failed with them. So I thought it might be better to initialize hydra as much similar way to actual training as possible, and tried to use hydra_task_runner while patching TaskTestFunction.__call__. Is there any better way?

[omry]

1. The compose API very intentionally does not change the working directory. so hydra.runtime.cwd is not even defined except for ".".

2. get_original_cwd() is also relying on a a Hydra config singleton which is not populated in compose.
   You could initialize the HydraConfig singleton yourself. I don't have a code sample but look how it's done normally in the codebase.

You can pass return_hydra_conf=true to compose to get the Hydra config node.

[nai62]

Thank you for your kind explanation! I didn't notice HydraConfig singleton has to be initialized. I'll give it a try.
So it seems I don't need hydra_task_runner anymore. May I close this issue?

[omry]

Well, normally it is initialized automatically by @hydra.main(). It just occurred to me after seeing your question that initializing it might be a solution in your situation.

The problem you are seeing is likely due to the TaskTestFunction shutting down the logging on exit. During the run of the task the logging subsystem is initialized and it's shutdown after the task.
I need to think about it a bit to see if there is a way around it. so far this has not been an issue for the tests in Hydra.