Fix breaking-changes checker for packages absent from Azure Artifacts feed (#48152)

* Fix breaking-changes checker for packages absent from Azure Artifacts feed

The breaking-changes detector resolves the previous released version via
PyPIClient. In CI, PIP_INDEX_URL points at the curated Azure Artifacts feed,
which is not a full mirror of public PyPI. A package released on public PyPI
(e.g. azure-mgmt-datatransfer) but absent from that feed caused
get_ordered_versions to raise KeyError, crashing changelog generation.

Add a force_pypi option to PyPIClient so callers can always query public PyPI,
and use it in _resolve_pypi_version so version resolution targets pypi.org.

* Potential fix for pull request finding

Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>

* Assert breaking-change checker resolves versions with force_pypi=True

Pin that _resolve_pypi_version constructs PyPIClient(force_pypi=True) so a
revert of that line is caught by the test suite instead of silently
restoring the original CI failure.

---------

Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>

Author: msyyc

eng/tools/azure-sdk-tools/pypi_tools/pypi.py

@@ -20,15 +20,19 @@ class PyPIClient:
     By default, reads ``PIP_INDEX_URL`` to decide the backend:
     * If the URL contains ``pkgs.dev.azure.com`` → Azure Artifacts REST API.
     * Otherwise → PyPI JSON API (``https://pypi.org``).
+
+    Pass ``force_pypi=True`` to ignore ``PIP_INDEX_URL`` and select the PyPI
+    JSON backend. Requests use ``host`` (public PyPI by default), so an
+    explicitly supplied host is still honored.
     """
 
-    def __init__(self, host="https://pypi.org"):
+    def __init__(self, host="https://pypi.org", force_pypi=False):
         index_url = os.environ.get("PIP_INDEX_URL", "")
 
         # Lazy import to avoid circular deps at module level.
         from pypi_tools.azdo import parse_pip_index_url, AzureArtifactsClient
 
-        azdo_cfg = parse_pip_index_url(index_url) if index_url else None
+        azdo_cfg = None if force_pypi else (parse_pip_index_url(index_url) if index_url else None)
 
         if azdo_cfg is not None:
             self._backend = "azdo"

eng/tools/azure-sdk-tools/tests/test_pypi_client.py

@@ -286,3 +286,55 @@ def test_filter_packages_for_compatibility(self, mock_sys):
         filtered = client.get_ordered_versions(WELL_KNOWN_PACKAGE, True)
         unfiltered = client.get_ordered_versions(WELL_KNOWN_PACKAGE, False)
         assert len(filtered) < len(unfiltered)
+
+
+# ---------------------------------------------------------------------------
+# Backend selection — force_pypi bypasses the Azure Artifacts feed
+# ---------------------------------------------------------------------------
+
+
+class TestBackendSelection:
+    """The AzDO feed is curated and not a full PyPI mirror, so callers that need
+    public PyPI (e.g. the breaking-change checker) pass force_pypi=True.
+    """
+
+    def _with_index_url(self, index_url, **kwargs):
+        old = os.environ.get("PIP_INDEX_URL")
+        try:
+            os.environ["PIP_INDEX_URL"] = index_url
+            return PyPIClient(**kwargs)
+        finally:
+            if old is not None:
+                os.environ["PIP_INDEX_URL"] = old
+            elif "PIP_INDEX_URL" in os.environ:
+                del os.environ["PIP_INDEX_URL"]
+
+    def test_azdo_index_url_selects_azdo_backend_by_default(self):
+        client = self._with_index_url(AZDO_FEED_URL)
+        assert client._backend == "azdo"
+
+    def test_force_pypi_overrides_azdo_index_url(self):
+        client = self._with_index_url(AZDO_FEED_URL, force_pypi=True)
+        assert client._backend == "pypi"
+
+    def test_force_pypi_get_ordered_versions_queries_pypi_json(self):
+        # A package present on public PyPI but absent from the curated feed must
+        # still resolve when force_pypi=True, because get_ordered_versions goes
+        # through project() (pypi.org JSON API) rather than the AzDO feed.
+        project_data = {
+            "info": {"version": "1.0.0b1"},
+            "releases": {
+                "1.0.0b1": [
+                    {
+                        "packagetype": "sdist",
+                        "url": "https://example.test/pkg-1.0.0b1.tar.gz",
+                    }
+                ],
+            },
+        }
+        client = self._with_index_url(AZDO_FEED_URL, force_pypi=True)
+
+        with patch.object(PyPIClient, "project", return_value=project_data):
+            versions = client.get_ordered_versions("azure-mgmt-datatransfer")
+
+        assert versions == [Version("1.0.0b1")]

scripts/breaking_changes_checker/detect_breaking_changes.py

@@ -735,12 +735,18 @@ def _resolve_pypi_version(package_name: str, latest_pypi_version: bool) -> str:
     """
     from pypi_tools.pypi import PyPIClient
 
-    client = PyPIClient()
+    # Force the public PyPI backend: in CI ``PIP_INDEX_URL`` points at the curated
+    # Azure Artifacts feed, which is not a full mirror of PyPI. A package can be
+    # released on public PyPI while absent from that feed, so query pypi.org
+    # directly to resolve the version to compare against.
+    client = PyPIClient(force_pypi=True)
     try:
         if latest_pypi_version:
             return str(client.get_ordered_versions(package_name)[-1])
         return str(client.get_relevant_versions(package_name)[1])
-    except IndexError:
+    except (IndexError, KeyError):
+        # IndexError: the package exists but has no relevant/previous version.
+        # KeyError: the package is brand new and not published on PyPI yet.
         _LOGGER.warning(f"No relevant version for {package_name} on PyPi. Exiting...")
         exit(0)
 

scripts/breaking_changes_checker/tests/test_code_report_changelog.py

@@ -433,7 +433,7 @@ def test_use_apistub_changelog_resolves_stable_from_pypi_and_current_from_local(
     checker.report_changes.return_value = ""
     checker.breaking_changes = []
 
-    with mock.patch("pypi_tools.pypi.PyPIClient", return_value=pypi_client), mock.patch.object(
+    with mock.patch("pypi_tools.pypi.PyPIClient", return_value=pypi_client) as pypi_client_cls, mock.patch.object(
         detect_breaking_changes, "build_report_from_apistub", return_value={}
     ) as build_report, mock.patch.object(
         detect_breaking_changes, "compare_report_dicts", return_value=checker
@@ -454,6 +454,11 @@ def test_use_apistub_changelog_resolves_stable_from_pypi_and_current_from_local(
 
     assert build_report.call_count == 2, "Expected separate apistub reports for current and stable"
 
+    # The resolver must force the public PyPI backend: in CI PIP_INDEX_URL points
+    # at the curated Azure Artifacts feed, which is not a full mirror of PyPI.
+    # Reverting force_pypi=True would restore the original CI failure, so pin it.
+    pypi_client_cls.assert_called_once_with(force_pypi=True)
+
     calls_by_label = {call.kwargs["label"]: call for call in build_report.call_args_list}
     assert set(calls_by_label) == {"current", "stable"}