fix: lazily initialize DEFAULT_JUMPSTART_SAGEMAKER_SESSION (#4468) [v2]

[mohamedzeidan2021]

Issue

Fixes #4468 on the v2 maintenance branch. Companion to #6330 (v3, against master).

At import time DEFAULT_JUMPSTART_SAGEMAKER_SESSION = Session(boto3.Session(...)) eagerly builds ~6 boto3 clients and resolves credentials/region, adding several seconds to import sagemaker even when the default session is never used.

Fix

Same lazy-proxy approach as #6330, ported to v2: _LazyJumpStartSagemakerSession defers Session construction until first use — truthy without initializing, forwards attribute reads/writes and copy/deepcopy, preserves the fail-to-None behavior on attribute access. The public name and its use as a default argument are unchanged.

Validation

• Instrumented boto3.Session.client/.resource: import now builds 0 clients (was 6); first attribute access builds them and yields a working Session.
• New regression test tests/unit/sagemaker/jumpstart/test_constants.py (6 tests); fails without the fix.
• black/flake8 clean.

Backwards compatibility

No public API change; security/bug-fix-only policy respected (behavior for existing inputs is unchanged in the normal path).

[jam-jee]

Summary: The fix itself is fine — src/sagemaker/jumpstart/constants.py is the same proxy as #6330 and the v2 copy(...) call site at utils.py:1304 is handled the same way. The branch, however, is not mergeable as-is.

Blocking

1. GitHub reports CONFLICTING / DIRTY. The branch carries two unrelated commits dated 2026-05-01: 18f22415 "v2 test updates" (rewrites tests/conftest.py GPU fixtures to ml.g4dn.xlarge, edits 6 integ tests, and adds two binaries hello-spark-java.jar / HelloJavaSparkApp.class) and 7412b08e "instance types update". They leak 9 unrelated files into this PR (+240/−69 vs the ~+70/−10 the fix needs) and overlap 88dbcc2e fix(tests): Replace deprecated ml.t2.medium (#6284) already on master-v2, which is the conflict. master-v2...d16e4ed is ahead 4 / behind 27.

   Please git rebase --onto origin/master-v2 7412b08e so only 9be072b4 + d16e4ede remain. The resulting diff should be exactly src/sagemaker/jumpstart/constants.py + tests/unit/sagemaker/jumpstart/test_constants.py.

2. The docs/readthedocs.org:sagemaker failure (No module named 'pkg_resources') is because this base predates 570b04c4 fix: RTD build failure (#6023); the rebase fixes it (v2 PR #6345 built today with the same setuptools and passed).

Should-have

• Carry over test_mock_patch_of_instance_level_attribute_tears_down_cleanly from #6330 — v2 has many tests that patch.object the default session, so that teardown path matters more here than on v3.

Nits (same as #6330)

• _resolve is not thread-safe; two threads racing on first use both build a Session and one is discarded (the old module-level constant was built under the import lock). A threading.Lock is a few lines.
• _resolved = True is set before the try; a KeyboardInterrupt mid-construction leaves the proxy permanently resolved to None. Set it inside both branches.
• Add a __repr__; dunders bypass __getattr__, so Sphinx will render <_LazyJumpStartSagemakerSession object at 0x…> in every sagemaker_session=DEFAULT_... signature.

CI attribution: codestyle-doc-tests (flake8 E501/F401 and black violations in files this PR does not touch; identical on unrelated v2 PR #6337) is pre-existing on master-v2. Unit tests pass on py39-py312.

[lucasjia-aws]

Automated review (1 comment). See inline comment.

[lucasjia-aws]

_resolve() sets cls._resolved = True before the (self-described several-second) Session(...) build completes, with no lock. Under concurrent first use — two threads reading an attribute off the default session — thread B enters _resolve, finds _resolved already True, and returns cls._session, which is still the initial None. __getattr__ then does getattr(None, name) and raises AttributeError on the success path, not just the documented failure path. The old eager init ran under the import lock, so this race is new. Guard the build with a lock and publish _resolved last, inside the lock:

import threading

class _LazyJumpStartSagemakerSession:
    __slots__ = ()
    _resolved = False
    _session = None
    _lock = threading.Lock()

    @classmethod
    def _resolve(cls):
        if not cls._resolved:
            with cls._lock:
                if not cls._resolved:
                    try:
                        cls._session = Session(
                            boto3.Session(region_name=JUMPSTART_DEFAULT_REGION_NAME)
                        )
                    except Exception as e:  # pylint: disable=W0703
                        cls._session = None
                        JUMPSTART_LOGGER.warning(...)
                    cls._resolved = True
        return cls._session

Note __slots__ = () is on the instance; a class-level _lock attribute is unaffected by it.