Reducing memory footprint for synchronous S3KeySensor

[jroachgolf84]

This PR aims to reduce the memory footprint for the S3KeySensor. This was done by altering the get_file_metadata method in the S3Hook to yield records in the paginated response, rather than loading them into a single list. The return type for the get_file_metadata method is not an Iterator. An assertion was added to validate this, and all appropriate tests were updated.

closes: #55039

[jroachgolf84]

cc: @dstandish

[jroachgolf84]

@eladkal, since we're changing the return type of get_file_metadata, would this need to be part of a major release?

[dstandish]

mentioned it on slack but just adding here

should check with @eladkal re backcompat issues. changing get_file_metadata to iterator likely means major release. and given that maybe we should update the whole check key funciton logic so that it never has to hold everything in memory, i.e. by making it so that check_fn takes a key and not a list

but thinking about that..... the problem is it's somewhat ambiguous what the behavior should be

for best performance, you would want check_fn to return on first "pass"

but, the behavior that would be most similar to current, would be to return all the files for which check_fn evaluates to true

[jroachgolf84]

cc: @eladkal

[eladkal]

Lets consult first with @o-nikolas and @vincbeck.

[vincbeck]

To keep it backward compatible, could we make get_file_metadata return list | Iterator? We would need to handle both cases in providers/amazon/src/airflow/providers/amazon/aws/sensors/s3.py which I do not really like but that's okay. What do you think?

[jroachgolf84]

I think that's feasible - how would we determine whether to return a list or an Iterator?

[vincbeck]

In the current implementation we would always return an iterator but having list | Iterator as return type annotation would keep it backward compatible

[jroachgolf84]

In the current implementation we would always return an iterator but having list | Iterator as return type annotation would keep it backward compatible

Making the type annotation change now. Both cases are handled in providers/amazon/src/airflow/providers/amazon/aws/sensors/s3.py.

[vincbeck]

Actually sorry but I realized I provided wrong instructions. This does not solve back compat issues ... What we should do instead is deprecate get_file_metadata and create a new one using Iterator and use this one in the sensor. That way we do not need to create a major release because of this.

[jroachgolf84]

Got it - so the plan would be to:

1. Deprecate get_file_metadata with a deprecation warning?
2. Create a new get_file_metadata method that returns an Iterator, and use that in the Sensor.

What is the best way to deprecate a method and create one with the same name? Or should I create a new method with a new name, and use this one in the Sensor?

[vincbeck]

Got it - so the plan would be to:

Deprecate get_file_metadata with a deprecation warning?
Create a new get_file_metadata method that returns an Iterator, and use that in the Sensor.

Correct

What is the best way to deprecate a method and create one with the same name?

To deprecate a method, please emit deprecation warning at the beginning of the method. Example below:

warnings.warn(
            "The property `async_conn` is deprecated. Accessing it in an async context will cause the event loop to block. "
            "Use the async method `get_async_conn` instead.",
            AirflowProviderDeprecationWarning,
            stacklevel=2,
        )

Or should I create a new method with a new name, and use this one in the Sensor?

Yes, you should create a new one with a new name

[jroachgolf84]

@vincbeck, I've updated the PR accordingly!

[vincbeck]

Nice! Thanks for the quick turn around!

[jroachgolf84]

@dstandish, can you take a look at the updated PR?

[jroachgolf84]

@eladkal, can you take a look at this for me?

[jroachgolf84]

@ashb should be all set here!