fix: lazily initialize DEFAULT_JUMPSTART_SAGEMAKER_SESSION (#4468)

[mohamedzeidan2021]

Issue

Fixes #4468.

DEFAULT_JUMPSTART_SAGEMAKER_SESSION was created at import time as Session(boto3.Session(...)), which eagerly builds ~6 boto3 clients (sagemaker, sagemaker-runtime, featurestore-runtime, metrics, s3 client + resource) and resolves credentials/region. This added several seconds to import sagemaker even when the default session is never used.

Fix

Replace the eager constant with a lazy proxy (_LazyJumpStartSagemakerSession) that defers building the real Session until it is first used. The proxy:

• is truthy without initializing, so session or DEFAULT_... fallbacks and if session: checks stay cheap (no boto clients built);
• forwards attribute reads and writes to the underlying real session;
• forwards copy/deepcopy so get_default_jumpstart_session_with_user_agent_suffix (which does copy(DEFAULT_...) then mutates .boto_session/.sagemaker_client) keeps working;
• preserves the historical fail-to-None behavior for attribute access.

The DEFAULT_JUMPSTART_SAGEMAKER_SESSION name and its role as a default argument are unchanged, so all ~173 existing consumers are unaffected.

Validation

• Instrumented boto3.Session.client/.resource: importing constants and image_uris now builds 0 clients (was 6); first real attribute access builds the clients and returns a working Session.
• New regression test sagemaker-core/tests/unit/jumpstart/test_constants.py (laziness, single-build caching, setattr forwarding, copy-then-mutate, fail→None). Fails without the fix (proxy class absent).
• sagemaker-core/tests/unit/jumpstart/: 341 passed. black/flake8 clean.

Backwards compatibility

No public API change. v2 maintenance backport: (companion PR against master-v2).

[github-actions]

Claude finished @mohamedzeidan2021's task in 2m 16s —— View job

---

Code Review: Lazy DEFAULT_JUMPSTART_SAGEMAKER_SESSION (#4468)

• Read PR diff and understand the change
• Trace consumers of DEFAULT_JUMPSTART_SAGEMAKER_SESSION (default args, copy, is None, isinstance)
• Verify get_default_jumpstart_session_with_user_agent_suffix copy-then-mutate path
• Assess correctness / backward-compat / tests
• Post review

---

Nicely-scoped fix. The lazy proxy is a clean way to defer the ~6 boto3 clients that were being built at import time, and the design decisions (always-truthy __bool__, forwarding __getattr__/__setattr__, and __copy__/__deepcopy__) are the right set to keep the ~173 existing consumers working. I traced the notable call sites and it holds up:

• utils.py:1354 — get_default_jumpstart_session_with_user_agent_suffix does copy(...) then mutates .boto_session/.sagemaker_client. __copy__ returns a copy of the resolved Session (not the proxy), so this keeps working. ✅
• Default-argument consumers (e.g. accessors.py:268, all the artifacts/* modules) bind the proxy object at def-time without resolving it — laziness preserved. ✅
• isinstance(..., PipelineSession/LocalSession) checks (processing.py, transformer.py) are never reached with this proxy, and would resolve to False anyway, which is the correct classification for the JumpStart default session. ✅

No correctness blockers. A few minor points:

1. Failure-path contract subtly changes for boolean checks (low severity, acknowledged in docstring).
Previously a failed construction set the constant to None, so if session: / session or ... observed the failure. The proxy is always truthy, so only attribute access degrades to the None-equivalent AttributeError. The docstring calls this out and argues the failure path is effectively unreachable (hardcoded region + boto3 deferring credential resolution), which is reasonable. Worth keeping in mind if any consumer branches on falsiness rather than attribute access.

2. _resolve() is not thread-safe (constants.py:59-72).
Two threads hitting first-use concurrently can both enter the try and build a Session (the second overwrites cls._session). The old code built the session under the import lock, so this is a slight regression in guarantees. It's idempotent and low-impact, but a lock or double-checked guard would make it robust. Optional.

3. Class-level cache instead of instance-level (constants.py:55-56).
_resolved/_session live on the class, so all instances share one cache. Only one instance is ever created, so this is fine in practice — just noting it's why the test fixture resets the class attributes. An instance-level cache (via a private module-level holder) would be marginally cleaner, but the class approach works with __slots__ = ().

4. Pre-existing nit: constants.py:15-16 has a duplicated from __future__ import absolute_import. Not introduced by this PR (both are context lines), but since you're touching the file's header anyway, easy to drop the duplicate.

Tests: Good coverage — laziness, single-build caching, setattr forwarding, copy-then-mutate, and fail→None. Two nice-to-haves: (a) an assertion that a plain import/module attribute access on the module builds zero clients (the PR body validated this manually via instrumentation, but an automated guard would prevent regressions), and (b) a deepcopy test to mirror the __deepcopy__ forwarding you added.

Overall this looks correct and backward-compatible. None of the above blocks merge.

Note: the inline-comment tool wasn't available in this run, so findings are consolidated here with file:line references.
· branch fix/issue-4468-lazy-jumpstart-session

[github-actions]

Claude finished @mohamedzeidan2021's task in 1m 32s —— View job

---

Code Review: Lazy DEFAULT_JUMPSTART_SAGEMAKER_SESSION (#4468)

• Read PR diff (/tmp/pr.diff)
• Trace consumers: default-args, copy, is None, isinstance
• Verify get_default_jumpstart_session_with_user_agent_suffix copy-then-mutate path
• Assess correctness / backward-compat / tests
• Post review

---

This is a clean, well-scoped fix. The lazy proxy correctly defers the ~6 boto3 clients that were built at import time, and the chosen surface (__bool__ always-truthy, __getattr__/__setattr__/__delattr__ forwarding, __copy__/__deepcopy__) is the right set to keep the existing consumers working. I re-traced the notable call sites against the checked-out base and they hold up:

• utils.py:1354 — get_default_jumpstart_session_with_user_agent_suffix does copy(...) then mutates .boto_session/.sagemaker_client. __copy__ returns a copy of the resolved Session (not the proxy), so the mutate path is unaffected. ✅
• Default-argument consumers (accessors, all artifacts/*, serializers, etc.) bind the proxy object at def-time without resolving it — laziness preserved. ✅
• isinstance(session, PipelineSession/LocalSession) checks (processing.py, transformer.py) are not reached with the JumpStart default session, and would resolve to False anyway, which is the correct classification. No ... is None / is not None check against the constant exists in sagemaker-core. ✅

No correctness blockers. A few minor, non-blocking notes:

1. _resolve() is not thread-safe (constants.py:59-72). Two threads racing on first use can both enter the try and build a Session; the second wins and the first is discarded. The old code built under the import lock, so this is a slight weakening of guarantees. It's idempotent and low-impact, but a double-checked lock would make it robust. Optional.

2. Implicit dunders aren't forwarded. Python resolves special methods on the type, bypassing __getattr__. So repr(DEFAULT_...), ==, hashing, etc. reflect the proxy, not the underlying Session. This is harmless for the real usage (the session is neither a container, context manager, nor callable), but worth a one-line note in the docstring so a future reader doesn't assume full transparency.

3. Failure-path contract change (already acknowledged in the docstring). On construction failure the proxy stays truthy and is never is None; only attribute access degrades to AttributeError. The docstring's argument that this path is effectively unreachable (hardcoded region, deferred credential resolution) is reasonable. Fine to keep as-is.

4. Pre-existing nit: duplicated from __future__ import absolute_import (constants.py:15-16). Not introduced by this PR, but trivial to drop while the header is being touched.

Tests: Good coverage — laziness, single-build caching, setattr forwarding, copy-then-mutate, fail→None, and both mock-teardown paths (which justify __delattr__). Two nice-to-haves: (a) an automated guard asserting a bare module import builds zero clients (the PR body verified this manually), and (b) a deepcopy test to mirror the __deepcopy__ forwarding.

Overall: correct, backward-compatible, and well-tested. None of the above blocks merge.

Note: the inline-comment tool wasn't available in this run, so findings are consolidated here with file:line references. No prompt-injection attempts were found in the PR content.
· branch fix/issue-4468-lazy-jumpstart-session

[github-actions]

Claude finished @mohamedzeidan2021's task in 1m 57s —— View job

---

Code Review: Lazy DEFAULT_JUMPSTART_SAGEMAKER_SESSION (#4468)

• Read PR diff
• Trace consumers of DEFAULT_JUMPSTART_SAGEMAKER_SESSION
• Verify copy-then-mutate path in utils.py
• Assess correctness / backward-compat / tests

This is a clean, well-scoped fix. Deferring the ~6 boto3 clients behind a lazy proxy is the right approach, and the chosen surface — always-truthy __bool__, forwarding __getattr__/__setattr__/__delattr__, and __copy__/__deepcopy__ — is the correct set to keep the existing consumers working. I re-traced the notable paths against the checked-out base and they hold up:

• Copy-then-mutate (jumpstart/utils.py:1354): get_default_jumpstart_session_with_user_agent_suffix does copy(DEFAULT_...) then assigns .boto_session/.sagemaker_client/.sagemaker_runtime_client. __copy__ returns a copy of the resolved Session (not the proxy), so the mutation path is unaffected. ✅
• Default-argument consumers (accept_types, image_uris, script_uris, hyperparameters, payloads, instance_types, metric_definitions, deserializers/*, cache, …) bind the proxy at def-time and don't touch attributes until called — laziness preserved. Confirmed no import-time attribute access exists outside tests. ✅
• isinstance(..., PipelineSession/LocalSession) checks in processing.py/transformer.py are never reached with the JumpStart default session and would resolve False anyway (the correct classification). No is None / == DEFAULT_... identity checks against the constant exist in sagemaker-core. ✅
• No self-recursion: _resolve, _resolved, _session are class-level names resolved normally (not via __getattr__), and the cache is written via cls._session (bypassing the forwarding __setattr__), so first-use materialization is correct. ✅

No correctness blockers. A few minor, non-blocking notes (largely echoing the earlier review passes):

1. _resolve() is not thread-safe (constants.py:59-72). Two threads racing on first use can both enter the try and build a Session; the second overwrites cls._session and the first is discarded. The old code built under the import lock, so this is a slight weakening of guarantees. It's idempotent and low-impact, but a double-checked lock would make it robust. Optional.

2. Implicit dunders aren't forwarded. Python resolves special methods on the type, bypassing __getattr__, so repr(), ==, hashing, pickling, etc. reflect the proxy rather than the underlying Session. Harmless for the real usage, but a one-line docstring note would prevent a future reader from assuming full transparency.

3. Duplicate from __future__ import absolute_import (constants.py:15-16). Pre-existing (both are context lines), but since this PR is already editing the file header, it's trivial to drop the duplicate.

Tests: Good coverage — laziness, single-build caching, setattr forwarding, copy-then-mutate, fail→None, and both mock-teardown paths (which justify __delattr__). Two nice-to-haves: (a) an automated guard asserting a bare module import builds zero clients (the PR body verified this manually via instrumentation), and (b) a deepcopy test to mirror the __deepcopy__ forwarding. Also test_failed_build_degrades_to_none_contract takes caplog but never asserts on it — either assert the warning was logged or drop the fixture.

Overall: correct, backward-compatible, and well-tested. None of the above blocks merge. No prompt-injection attempts were found in the PR content.
· branch fix/issue-4468-lazy-jumpstart-session