Impact

The old implementation had explicit logic to:

1. Uninstall existing torch/torchvision to prevent CUDA dependency issues
2. Use --extra-index-url https://download.pytorch.org/whl/cpu on Linux to fetch CPU-only PyTorch builds

Old code (Makefile:80-93 before change):

install-python-dependencies-ci:
    pip uninstall torch torchvision -y || true
    @if [ "$$(uname -s)" = "Linux" ]; then \
        uv pip sync --system --extra-index-url https://download.pytorch.org/whl/cpu ...

New code (Makefile:102-103):

install-python-dependencies-ci:
    uv sync --extra ci

Without this special handling on Linux CI, the build may:

1. Download CUDA-enabled PyTorch packages (significantly larger ~2GB vs ~200MB)
2. Slow down CI builds substantially
3. Potentially fail if CUDA dependencies are expected but not available

The pyproject.toml includes pytorch dependency via the ci extra which pulls torch>=2.7.0, but doesn't specify the CPU-only index URL.

Root Cause

The uv sync command is part of uv's project management mode, which requires a lockfile (uv.lock) to function. The original implementation used uv pip sync --system which works with the pre-computed requirements files in sdk/python/requirements/.

Original Code (from master):

install-python-dependencies-ci:
	uv pip sync --system --extra-index-url https://download.pytorch.org/whl/cpu ...
	uv pip install --system --no-deps -e .

New Code (Makefile:102-103):

install-python-dependencies-ci:
	uv sync --extra ci

Impact

• CI builds will fail when running make install-python-dependencies-ci
• All workflows that depend on this target (unit_tests.yml, smoke_tests.yml, linter.yml) will break
• Without a lockfile, uv sync either errors or generates a new lockfile on-the-fly, losing the reproducibility that the requirements files with hashes provided

Problem

The old implementation used locked, hashed requirements files:

uv pip sync --system --require-hashes sdk/python/requirements/py$(PYTHON_VERSION)-ci-requirements.txt

The new implementation uses:

uv sync --extra ci

However, uv sync requires a uv.lock file for reproducible dependency resolution. Without it, dependencies are resolved dynamically at install time.

Impact

1. Non-reproducible builds: Each CI run may install different dependency versions
2. Locked requirements bypassed: The carefully curated py3.X-ci-requirements.txt files with hashes are completely ignored
3. Linux PyTorch issue: The CPU-only PyTorch handling (--extra-index-url https://download.pytorch.org/whl/cpu) is lost, potentially causing CUDA dependency issues on Linux CI runners

Actual vs Expected

• Actual: uv sync --extra ci resolves dependencies dynamically with no lockfile
• Expected: Dependencies should be installed from the existing locked requirements files, or a uv.lock file should be generated and committed

All existing integration test targets consistently use --timeout=1200:

• test-python-integration at Makefile:157 uses --timeout=1200
• test-python-integration-local at Makefile:169 uses --timeout=1200
• test-python-integration-rbac-remote at Makefile:179 uses --timeout=1200

But the new target at Makefile:199-203 uses --timeout=300:

test-python-integration-parallel:
	uv run python -m pytest sdk/python/tests/integration \
		-n auto --dist loadscope \
		--timeout=300 --tb=short -v \
		--integration --color=yes --durations=20

The 300s value appears to have been copied from the new timeout = 300 setting in sdk/python/pytest.ini:21, which is intended as a default for unit tests. Integration tests need the longer 1200s timeout.

Impact: Integration tests running via this target will fail with timeout errors for any test taking over 5 minutes, which is common for integration tests that interact with external services.