tox.ini

[tox]
requires = tox>=4.24.0
env_list = lint, py311, py312, py313, long, endpoint-import-smoke
labels =
    # Friendly aliases for envs that don't share their name. Multi-env aggregates
    # also live here (`full`). 1-to-1 aliases like `serial = serial` aren't needed
    # — `tox -e <env>` is equivalent and clearer.
    quick = py313
    smoke = endpoint-import-smoke
    full = lint, py311, py312, py313, medium, long, endpoint-import-smoke

[testenv]
# Default: parallel via pytest-xdist, file-level distribution.
# -n auto picks worker count from CPU cores (clamped sanely by xdist).
# --dist=loadfile keeps all tests in one file on the same worker so intra-file
#   ordering / shared fixtures are preserved. Cross-file parallelism is where
#   the speedup comes from — and where the risk lives. If you hit cross-file
#   AWS-resource collisions (same endpoint name, same Parameter Store key, etc.),
#   either lower -n, fix the test to use a unique resource, or drop the offender
#   into a serial group, or use `tox -e serial` to revert to one-at-a-time.
# Drop -x because failures in one worker don't reach others — letting all
#   workers finish gives a fuller picture.
description = Run unit tests in parallel (file-level distribution)
basepython = python3.12
# AWS_* / WORKBENCH_* pass through so both CI (OIDC env credentials + secret
# overrides) and local runs (AWS_PROFILE) reach the tests.
passenv =
    WORKBENCH_*
    AWS_*
    CI
setenv =
    PLOTLY_RENDERER = {env:PLOTLY_RENDERER:browser}
deps =
    pytest
    pytest-sugar
    pytest-xdist
    coverage
    pytest-cov
    .[all]
commands = pytest --assert=plain -n auto --dist=loadfile --durations=0 -m "not medium and not long and not corner_case" {posargs:tests}

[testenv:serial]
description = Run unit tests serially (escape hatch when parallel run hits a resource collision).
# Useful for debugging cross-file races — if `tox` (parallel) flakes intermittently
# and `tox -e serial` passes deterministically, the offender is a shared AWS
# resource between two test files.
basepython = python3.13
deps = {[testenv]deps}
commands = pytest --assert=plain -x -m "not medium and not long and not corner_case" {posargs:tests}

[testenv:verbose]
description = Run tests serially with verbose logging
deps = {[testenv]deps}
commands = pytest --assert=plain -x -m "not medium and not long and not corner_case" -s {posargs:tests}

[testenv:medium]
description = Run medium-tier tests (heavier validation studies, e.g. 3D endpoint comparisons)
deps = {[testenv]deps}
commands = pytest --assert=plain -n 4 --dist=loadfile --durations=0 -m "medium" {posargs:tests}

[testenv:py311]
description = Run tests with Python 3.11 (parallel)
basepython = python3.11
deps = {[testenv]deps}

[testenv:py312]
description = Run tests with Python 3.12 (parallel)
basepython = python3.12
deps = {[testenv]deps}

[testenv:py313]
description = Run tests with Python 3.13 (parallel)
basepython = python3.13
deps = {[testenv]deps}

[testenv:long]
# Parallel via pytest-xdist with file-level distribution (--dist=loadfile keeps
# each file's tests on one worker, preserving intra-file create->use->delete
# ordering; cross-file parallelism is where the speedup comes from). Capped at
# -n 4 to bound concurrent SageMaker endpoint deploys against account instance
# quotas. Drop -x because a failure in one worker doesn't reach the others, so
# letting all workers finish gives a fuller picture.
description = Run long-duration tests (parallel, file-level distribution)
setenv = WORKBENCH_DEBUG = True
commands = pytest --assert=plain -n 4 --dist=loadfile -v -s -m "long" {posargs:tests}

[testenv:lint]
description = Run linters
# Newest supported interpreter so black's AST check matches its py313 target.
basepython = python3.13
skip_install = true
deps =
    black>=22.12
    flake8
commands =
    black --line-length=120 {posargs:.}
    flake8 --exclude='*generated*,.tox/*'

[testenv:format]
description = Run formatters
skip_install = true
deps = black>=22.12
commands = black --line-length=120 {posargs:.}

[testenv:endpoint-import-smoke]
description = Verify the workbench.endpoints contract: every module imports without heavy extras.
# Contract: `workbench.endpoints.*` is THE import surface for model scripts
# running inside SageMaker endpoint containers. Every module under
# workbench/endpoints/ must be importable against the leanest endpoint dep
# manifest in the stack (ci/endpoint_smoke_requirements.txt — the
# intersection of every deployed endpoint container's requirements). If a
# refactor moves a heavy transitive import (sagemaker, aiobotocore, dash,
# cleanlab, umap-learn, matplotlib, mordred, ipython, networkx, xgboost,
# ...) into the import chain of any workbench.endpoints submodule, this
# job fails — telling us before an endpoint deploy does.
#
# This test enumerates the directory dynamically — adding a new module under
# workbench/endpoints/ automatically extends the contract. To exempt a module
# from the contract, it must not live under workbench/endpoints/.
basepython = python3.12
skip_install = true
# Override [testenv].deps (which installs `.[all]` — the full heavy tree). That
# would pollute this deliberately-lean env and defeat the contract. This env is
# built ONLY from the lean smoke requirements + workbench --no-deps (mirroring
# .github/workflows/endpoint-import-smoke.yml).
deps =
commands_pre =
    pip install -r {toxinidir}/ci/endpoint_smoke_requirements.txt
    pip install --no-deps {toxinidir}
commands = python {toxinidir}/ci/endpoint_import_smoke.py

[testenv:template-import-check]
description = Enforce the template import invariant: no top-level workbench.training imports.
# workbench.training modules carry training-only deps (ray/optuna) absent from
# endpoint images. A template may import them only inside __main__ (deferred),
# never at module scope — else the endpoint crashes at load. Static, no install.
basepython = python3.13
skip_install = true
# Override [testenv].deps (`.[all]`) — this check is a stdlib-only line scan; it
# needs nothing installed.
deps =
commands = python {toxinidir}/ci/template_import_check.py