[rsareddy0329]

Issue

unit-tests (sagemaker-serve) is failing on open PRs (e.g. #6508, #6516) — this is environment/dependency churn, not a test or code regression.

The serve unit job runs three tox envs. py310 passes all 2176 tests; py311/py312 fail at pytest collection:

ImportError: C extension: None not built...
ERROR:   py311: commands failed
ERROR:   py312: commands failed

pandas is pulled in transitively (via mlflow) and is unpinned. The newly released pandas 3.1.0 (requires Python >= 3.11) now floats into the py311/py312 envs; its wheels are built against a newer NumPy C-ABI than those envs resolve (numpy==2.1.3), so pandas' compiled extension cannot load and import fails before any test runs. py310 is unaffected because it cannot install 3.x and stays on the working 2.3.3.

Until pandas 3.1.0 was released, py311/py312 resolved pandas 3.0.6 and passed — i.e. this went red purely due to the new release, with no code change.

Description of changes

Cap pandas below 3.1.0 in the serve test extra, restoring the last-known-good configuration:

"pandas>=2.3.0,<3.1.0",

• py311/py312 -> 3.0.6 (what was green before the 3.1.0 release)
• py310 -> 2.3.3 (pandas 3.x requires Python >= 3.11)

A hard ==3.0.6 is intentionally avoided — it would break the py310 env, which cannot install any 3.x.

Testing done

Locally, matching the CI Python versions:

• py312: unconstrained pip install pandas -> 3.1.0 (the broken one); pip install "pandas>=2.3.0,<3.1.0" -> 3.0.6, and import pandas succeeds.
• py310: pip install "pandas>=2.3.0,<3.1.0" -> 2.3.3 (3.x correctly excluded).
• pyproject.toml parses; [test] extra carries pandas>=2.3.0,<3.1.0.

Full validation is the serve unit job on this PR across py310/py311/py312.

By submitting this pull request, I confirm that my contribution is made under the terms of the Apache 2.0 license.

[github-actions]

🤖 AI code review

This PR adds pandas>=2.3.0,<3 to the test extra in sagemaker-serve/pyproject.toml to unblock the py311/py312 serve unit-test CI envs, which were resolving the newly released pandas 3.1.0 and failing at collection time. The change is scoped to the test extra, uses a version range rather than an exact pin, and is consistent with sibling packages (sagemaker-core uses pandas>=2.0.0, <3.0.0, sagemaker-mlops uses pandas<3.0) and the existing conda_in_process.yml pin (pandas>=2.3.0). This also aligns with prior guidance to use ranges rather than exact pins (Source: #4983). No issues found.

_{Reviewed commit 4288634. Automated review; verify before acting.}

[github-actions]

🤖 AI code review

This PR adds a pandas>=2.3.0,<3.1.0 pin to the test extra of sagemaker-serve/pyproject.toml to unblock py311/py312 unit test collection. The change is small, correct, and consistent with existing conventions elsewhere in the repo (e.g. sagemaker-core uses pandas>=2.0.0, <3.0.0). No code bugs found; one minor note about a mismatch between the PR description and the actual constraint.

_{Reviewed commit 372ddb3. Automated review; verify before acting.}

[github-actions]

🟡 Low: Minor: the PR description states the constraint is pandas>=2.3.0,<3 (a hard cap below 3.x), but the actual pin here is pandas>=2.3.0,<3.1.0, which still permits pandas 3.0.x on py311/py312. The inline comment's reasoning is internally consistent (it says 3.0.6 passed and only 3.1.0 broke), so the <3.1.0 cap is intentional — just worth updating the PR description so reviewers aren't confused about which cap was actually applied. Note that team precedent (Source: #4983) discourages hard single-major caps; <3.1.0 is a narrower upper bound and acceptable, but consider whether <3.1.0 vs a future relaxation is tracked anywhere so the temporary cap doesn't linger.