Do not cache module lookup results that may become invalid in future (#19044)

Fixes #19037. See my analysis in #19037 for motivation. As a side note,
with this patch the second run behaves as expected (without it mypy
tries to reanalyze a lot of modules during the 2nd run, now it only
rechecks the inputs with errors).

Author: sterliakov

mypy/modulefinder.py

@@ -320,13 +320,21 @@ def find_module(self, id: str, *, fast_path: bool = False) -> ModuleSearchResult
                 use_typeshed = self._typeshed_has_version(id)
             elif top_level in self.stdlib_py_versions:
                 use_typeshed = self._typeshed_has_version(top_level)
-            self.results[id] = self._find_module(id, use_typeshed)
-            if (
-                not (fast_path or (self.options is not None and self.options.fast_module_lookup))
-                and self.results[id] is ModuleNotFoundReason.NOT_FOUND
-                and self._can_find_module_in_parent_dir(id)
-            ):
-                self.results[id] = ModuleNotFoundReason.WRONG_WORKING_DIRECTORY
+            result, should_cache = self._find_module(id, use_typeshed)
+            if should_cache:
+                if (
+                    not (
+                        fast_path or (self.options is not None and self.options.fast_module_lookup)
+                    )
+                    and result is ModuleNotFoundReason.NOT_FOUND
+                    and self._can_find_module_in_parent_dir(id)
+                ):
+                    self.results[id] = ModuleNotFoundReason.WRONG_WORKING_DIRECTORY
+                else:
+                    self.results[id] = result
+                return self.results[id]
+            else:
+                return result
         return self.results[id]
 
     def _typeshed_has_version(self, module: str) -> bool:
@@ -384,11 +392,16 @@ def _can_find_module_in_parent_dir(self, id: str) -> bool:
         while any(is_init_file(file) for file in os.listdir(working_dir)):
             working_dir = os.path.dirname(working_dir)
             parent_search.search_paths = SearchPaths((working_dir,), (), (), ())
-            if not isinstance(parent_search._find_module(id, False), ModuleNotFoundReason):
+            if not isinstance(parent_search._find_module(id, False)[0], ModuleNotFoundReason):
                 return True
         return False
 
-    def _find_module(self, id: str, use_typeshed: bool) -> ModuleSearchResult:
+    def _find_module(self, id: str, use_typeshed: bool) -> tuple[ModuleSearchResult, bool]:
+        """Try to find a module in all available sources.
+
+        Returns:
+            ``(result, can_be_cached)`` pair.
+        """
         fscache = self.fscache
 
         # Fast path for any modules in the current source set.
@@ -424,7 +437,7 @@ def _find_module(self, id: str, use_typeshed: bool) -> ModuleSearchResult:
             else None
         )
         if p:
-            return p
+            return p, True
 
         # If we're looking for a module like 'foo.bar.baz', it's likely that most of the
         # many elements of lib_path don't even have a subdirectory 'foo/bar'.  Discover
@@ -444,6 +457,9 @@ def _find_module(self, id: str, use_typeshed: bool) -> ModuleSearchResult:
             for component in (components[0], components[0] + "-stubs")
             for package_dir in self.find_lib_path_dirs(component, self.search_paths.package_path)
         }
+        # Caching FOUND_WITHOUT_TYPE_HINTS is not always safe. That causes issues with
+        # typed subpackages in namespace packages.
+        can_cache_any_result = True
         for pkg_dir in self.search_paths.package_path:
             if pkg_dir not in candidate_package_dirs:
                 continue
@@ -475,6 +491,7 @@ def _find_module(self, id: str, use_typeshed: bool) -> ModuleSearchResult:
             if isinstance(non_stub_match, ModuleNotFoundReason):
                 if non_stub_match is ModuleNotFoundReason.FOUND_WITHOUT_TYPE_HINTS:
                     found_possible_third_party_missing_type_hints = True
+                    can_cache_any_result = False
             else:
                 third_party_inline_dirs.append(non_stub_match)
                 self._update_ns_ancestors(components, non_stub_match)
@@ -513,7 +530,7 @@ def _find_module(self, id: str, use_typeshed: bool) -> ModuleSearchResult:
                 if verify and not verify_module(fscache, id, path_stubs, dir_prefix):
                     near_misses.append((path_stubs, dir_prefix))
                 else:
-                    return path_stubs
+                    return path_stubs, True
 
             # Prefer package over module, i.e. baz/__init__.py* over baz.py*.
             for extension in PYTHON_EXTENSIONS:
@@ -523,7 +540,7 @@ def _find_module(self, id: str, use_typeshed: bool) -> ModuleSearchResult:
                     if verify and not verify_module(fscache, id, path, dir_prefix):
                         near_misses.append((path, dir_prefix))
                         continue
-                    return path
+                    return path, True
 
             # In namespace mode, register a potential namespace package
             if self.options and self.options.namespace_packages:
@@ -541,7 +558,7 @@ def _find_module(self, id: str, use_typeshed: bool) -> ModuleSearchResult:
                     if verify and not verify_module(fscache, id, path, dir_prefix):
                         near_misses.append((path, dir_prefix))
                         continue
-                    return path
+                    return path, True
 
         # In namespace mode, re-check those entries that had 'verify'.
         # Assume search path entries xxx, yyy and zzz, and we're
@@ -570,35 +587,35 @@ def _find_module(self, id: str, use_typeshed: bool) -> ModuleSearchResult:
                 for path, dir_prefix in near_misses
             ]
             index = levels.index(max(levels))
-            return near_misses[index][0]
+            return near_misses[index][0], True
 
         # Finally, we may be asked to produce an ancestor for an
         # installed package with a py.typed marker that is a
         # subpackage of a namespace package.  We only fess up to these
         # if we would otherwise return "not found".
         ancestor = self.ns_ancestors.get(id)
         if ancestor is not None:
-            return ancestor
+            return ancestor, True
 
         approved_dist_name = stub_distribution_name(id)
         if approved_dist_name:
             if len(components) == 1:
-                return ModuleNotFoundReason.APPROVED_STUBS_NOT_INSTALLED
+                return ModuleNotFoundReason.APPROVED_STUBS_NOT_INSTALLED, True
             # If we're a missing submodule of an already installed approved stubs, we don't want to
             # error with APPROVED_STUBS_NOT_INSTALLED, but rather want to return NOT_FOUND.
             for i in range(1, len(components)):
                 parent_id = ".".join(components[:i])
                 if stub_distribution_name(parent_id) == approved_dist_name:
                     break
             else:
-                return ModuleNotFoundReason.APPROVED_STUBS_NOT_INSTALLED
+                return ModuleNotFoundReason.APPROVED_STUBS_NOT_INSTALLED, True
             if self.find_module(parent_id) is ModuleNotFoundReason.APPROVED_STUBS_NOT_INSTALLED:
-                return ModuleNotFoundReason.APPROVED_STUBS_NOT_INSTALLED
-            return ModuleNotFoundReason.NOT_FOUND
+                return ModuleNotFoundReason.APPROVED_STUBS_NOT_INSTALLED, True
+            return ModuleNotFoundReason.NOT_FOUND, True
 
         if found_possible_third_party_missing_type_hints:
-            return ModuleNotFoundReason.FOUND_WITHOUT_TYPE_HINTS
-        return ModuleNotFoundReason.NOT_FOUND
+            return ModuleNotFoundReason.FOUND_WITHOUT_TYPE_HINTS, can_cache_any_result
+        return ModuleNotFoundReason.NOT_FOUND, True
 
     def find_modules_recursive(self, module: str) -> list[BuildSource]:
         module_path = self.find_module(module, fast_path=True)

mypy/test/testpep561.py

@@ -145,8 +145,11 @@ def test_pep561(testcase: DataDrivenTestCase) -> None:
                     output.append(line[len(test_temp_dir + os.sep) :].rstrip("\r\n"))
                 else:
                     # Normalize paths so that the output is the same on Windows and Linux/macOS.
-                    line = line.replace(test_temp_dir + os.sep, test_temp_dir + "/")
-                    output.append(line.rstrip("\r\n"))
+                    # Yes, this is naive: replace all slashes preceding first colon, if any.
+                    path, *rest = line.split(":", maxsplit=1)
+                    if rest:
+                        path = path.replace(os.sep, "/")
+                    output.append(":".join([path, *rest]).rstrip("\r\n"))
             iter_count = "" if i == 0 else f" on iteration {i + 1}"
             expected = testcase.output if i == 0 else testcase.output2.get(i + 1, [])
 

test-data/packages/typedpkg_ns_nested/pyproject.toml

@@ -0,0 +1,11 @@
+[project]
+name = 'typedpkg_namespace.nested'
+version = '0.1'
+description = 'Two namespace packages, one of them typed'
+
+[tool.hatch.build]
+include = ["**/*.py", "**/*.pyi", "**/py.typed"]
+
+[build-system]
+requires = ["hatchling==1.18"]
+build-backend = "hatchling.build"

test-data/packages/typedpkg_ns_nested/typedpkg_ns/a/__init__.py

@@ -213,3 +213,23 @@ from typedpkg_ns.a.bbb import bf
 [file dummy.py.2]
 [out]
 [out2]
+
+[case testTypedNamespaceSubpackage]
+# pkgs: typedpkg_ns_nested
+import our
+[file our/__init__.py]
+import our.bar
+import our.foo
+[file our/bar.py]
+from typedpkg_ns.b import Something
+[file our/foo.py]
+import typedpkg_ns.a
+
+[file dummy.py.2]
+
+[out]
+our/bar.py:1: error: Skipping analyzing "typedpkg_ns.b": module is installed, but missing library stubs or py.typed marker
+our/bar.py:1: note: See https://mypy.readthedocs.io/en/stable/running_mypy.html#missing-imports
+[out2]
+our/bar.py:1: error: Skipping analyzing "typedpkg_ns.b": module is installed, but missing library stubs or py.typed marker
+our/bar.py:1: note: See https://mypy.readthedocs.io/en/stable/running_mypy.html#missing-imports