[2.8] Fix tracking recipe integration test

[YuanTingHsieh]

Summary

• Replace the dynamic _exp_tracking_model import in the experiment tracking recipe smoke tests with the documented dict model config.
• Bundle the example model.py into the server app so model.SimpleNetwork resolves when the job runs.
• Add a dedicated tracking integration backend and route it through the PyTorch CI setup.
• Refresh the integration test README/backend list and stale manual pytest command.

Why

The tests loaded the example model under a synthetic module name and then passed that live nn.Module instance into FedAvgRecipe. During job export, Python source inspection could not resolve the synthetic module and raised:

TypeError: <class '_exp_tracking_model.SimpleNetwork'> is a built-in class

Using model={"class_path": "model.SimpleNetwork", "args": {}} matches the recipe API and avoids relying on importlib state.

[Copilot]

Pull request overview

This PR fixes the experiment tracking recipe integration smoke tests by aligning them with the documented Recipe API model configuration and ensuring required example code is packaged into the exported job. It also introduces a dedicated tracking integration backend and hooks it into the existing integration test runner and PyTorch CI path.

Changes:

• Update experiment tracking recipe tests to pass a dict-based model config (class_path/args) instead of a dynamically imported nn.Module instance, avoiding export-time source inspection failures.
• Bundle the example model.py into the server app so model.SimpleNetwork can be resolved at job runtime.
• Add a new tracking integration backend (pytest-file driven) and route it through ci/run_integration.sh’s PyTorch integration setup; refresh backend list docs.

Reviewed changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
tests/integration_test/test_configs.yml	Adds tracking to pytest_files so it can run dedicated pytest modules.
tests/integration_test/run_integration_tests.sh	Registers tracking as a selectable backend in the integration runner.
tests/integration_test/README.md	Updates documented backend list to include tracking.
tests/integration_test/experiment_tracking_recipes_test.py	Switches to dict model config + bundles model.py into the job’s server app.
ci/run_integration.sh	Routes tracking to the PyTorch integration runner path in CI.

Comments suppressed due to low confidence (1)

tests/integration_test/README.md:29

• README states that only standalone runs explicit pytest files from pytest_files, but this PR also introduces a tracking backend that runs pytest files via the same mechanism. Please update this section to mention tracking as well to avoid misleading instructions.

The backend options are:
`numpy`, `tensorflow`, `pytorch`, `auth`, `preflight`, `cifar`, `stats`, `xgboost`,
`client_api`, `client_api_qa`, `model_controller_api`, `tracking`, and `standalone`.

`preflight` has its own entry file. Most backend options run through
`tests/integration_test/system_test.py`, and `standalone` runs explicit pytest files listed in
`pytest_files` in `tests/integration_test/test_configs.yml`.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[greptile-apps]

Greptile Summary

This PR fixes a TypeError in the experiment tracking recipe smoke tests by replacing dynamic importlib-based model instantiation with a dict-based model config ({"class_path": "model.SimpleNetwork", "args": {}}), and bundling model.py into both server and client apps via _add_model_to_apps.

• Model loading fix: Replaces importlib.util.spec_from_file_location(\"_exp_tracking_model\", ...) with {\"class_path\": \"model.SimpleNetwork\", \"args\": {}} to match the FedAvgRecipe API and avoid synthetic module name resolution failures during job export.
• File distribution fix: Introduces _add_model_to_apps which calls both add_file_to_server and add_file_to_clients, ensuring model.py is available in every site's custom directory at runtime.

Confidence Score: 5/5

Safe to merge — the fix correctly distributes model.py to both server and all client sites, resolving the runtime import failure.

The change is narrow and targeted: it swaps an unreliable importlib trick for a documented dict config, and ensures model.py is copied into every site's custom directory (server via add_file_to_server, clients via add_file_to_clients using ALL_SITES). Both the server-side model instantiation path and the client-side from model import SimpleNetwork import are now satisfied. No logic changes outside the test file.

No files require special attention.

Important Files Changed

Filename	Overview
tests/integration_test/experiment_tracking_recipes_test.py	Fixes synthetic module import error by switching to dict model config and distributes model.py to both server and all client sites.

_{Reviews (3): Last reviewed commit: "Merge branch '2.8' into codex/fix-tracki..." | Re-trigger Greptile}

[YuanTingHsieh]

Scope update after review: this PR has been narrowed to tests/integration_test/experiment_tracking_recipes_test.py only. It no longer adds a tracking integration backend or changes CI runner wiring.

Current fix:

• replace the synthetic _exp_tracking_model import with the recipe dict model config model.SimpleNetwork
• bundle model.py into both server and client apps, since the server persistor and client script both need to import it
• refresh the stale manual pytest filename in the test docstring

Validated locally:

• python3 -m py_compile tests/integration_test/experiment_tracking_recipes_test.py
• git diff upstream/2.8...HEAD --check