[shunping]

• gcsio.py: Guard google-cloud-storage imports behind a flag instead of failing at module import time when GCP extras are not installed. This allows GCSFileSystem to be imported and registered in FileSystem.get_all_subclasses(), deferring RuntimeError until GcsIO is instantiated without a client.

• s3io.py: Explicitly import boto3 inside the try block so BOTO3_INSTALLED is set to False when boto3 is missing (previously from apache_beam.io.aws.clients.s3 import boto3_client succeeded even without boto3, so BOTO3_INSTALLED has no effect).

fixes #37445

[shunping]

r: @Abacn
cc: @hjtran

[github-actions]

Stopping reviewer notifications for this pull request: review requested by someone other than the bot, ceding control. If you'd like to restart, comment assign set of reviewers

[shunping]

I got a bunch of test failures. Looks like our test suites often use whether import apache_beam.io.gcp.gcsio can run successfully as the criteria of the present of gcs dependencies. For example:

beam/sdks/python/apache_beam/options/pipeline_options_test.py

Line 50 in 73bb356

import apache_beam.io.gcp.gcsio # pylint: disable=unused-import

Our change in this PR breaks these tests and we have to fix them.

[Abacn]

Thanks, LGTM

Looks like our test suites often use whether import apache_beam.io.gcp.gcsio can run successfully as the criteria of the present of gcs dependencies.

This is the only thing a bit worrying. Current approach (Beam 2.77.0) is fragile admittedly. Hopefully this pattern isn't taken by users downstream