[inchang-ing]

Description

MigrationGraph.leaf_nodes(app_label) and root_nodes(app_label) filtered the candidate nodes by app label but still required a node to have no children (or no parents) at all, including those in other apps. An app whose newest migration is depended on from another app therefore had no leaf, and an app whose first migration depends on another app had no root.

With an app label, only same-app edges decide whether that app has a root or a leaf now. The global call (leaf_nodes()/root_nodes() with no argument, used for the full plan) keeps the original any-app behaviour.

Motivation and Context

Fixes #2305.

• migrate/plan for A.__latest__ (and bare A) planned nothing: executor.py looks up graph.leaf_nodes(target.app_label) and got an empty list.
• B.__first__ rolled nothing back for the same reason with graph.root_nodes(target.app_label).
• loader.py resolves __first__/__latest__ dependencies through the same helpers, so a dependency such as ("A", "__latest__") could raise the misleading Dependency on app with no migrations for an app that does have migrations.

How Has This Been Tested?

• New unit tests in tests/migrations/test_migration_graph.py for the reported graph (a.0002 depended on by b.0001): leaf_nodes("a") == [a.0002_second], root_nodes("b") == [b.0001_initial], plus the unchanged global results.
• The app-scoped assertions fail on develop before the change and pass after it.
• pytest tests/migrations/ -> 311 passed, 6 skipped.

Checklist:

• My code follows the code style of this project.
• My change requires a change to the documentation.
• I have updated the documentation accordingly.
• I have added the changelog accordingly.
• I have read the CONTRIBUTING document.
• I have added tests to cover my changes.
• All new and existing tests passed.

---

AI disclosure: prepared with the help of an AI coding assistant. I reviewed the change, reproduced the bug, and ran the tests locally.

[codspeed]

Merging this PR will not alter performance

1 warning

⚠️ 21 benchmarks are dominated by system calls

System calls cannot be simulated consistently and are left out of the measure. The reported value understates the real cost. Switch to the Walltime instrument to measure them.

Learn more about system calls

See affected benchmarks

✅ 24 untouched benchmarks

---

_{Comparing inchang-ing:fix/migration-graph-app-scoped-leaf-root (a677e7e) with develop (15330b9)}

[inchang-ing]

CI is now fully green — check plus test-sqlite/test-postgres/test-mysql/test-mssql across Python 3.10–3.14, plus coverage and CodSpeed, all SUCCESS (runs triggered 2026-10-11). The fix scopes leaf/root migration nodes to same-app edges. Ready for review/merge when you have a chance. Thanks!

[waketzheng]

How about:

    @staticmethod
    def _has_related_node(related: set[Node], app_label: str | None) -> bool:
        # For a single-app request only same-app edges decide whether that app has a root or
        # a leaf: an app whose newest migration is depended on from another app, or whose
        # first migration depends on another app, still has its own target.
        return any(node.key.app_label == app_label for node in related)

    def _filter_boundary_keys(
        self, attr: Literal["parents", "children"], app_label: str | None
    ) -> list[MigrationKey]:
        nodes = self.node_map.values()
        if app_label is None:
            return sorted([node.key for node in nodes if not getattr(node, attr)])
        keys = [
            node.key
            for node in nodes
            if node.key.app_label == app_label
            and not self._has_related_node(node.parents, app_label)
        ]
        return sorted(keys)

    def root_nodes(self, app_label: str | None = None) -> list[MigrationKey]:
        return self._filter_boundary_keys("parents", app_label)

    def leaf_nodes(self, app_label: str | None = None) -> list[MigrationKey]:
        return self._filter_boundary_keys("children", app_label)

[inchang-ing]

Good call — adopted. _filter_boundary_keys now takes the attr and handles the global case once up front, and _has_related_node is a plain same-app check without the app_label is None branch. One correction to the sketch: it hardcoded node.parents, so leaf_nodes() would have kept using parents; it uses getattr(node, attr) now. Rebased onto current develop while at it (the changelog was the only conflict).

[waketzheng]

Use #2307 instead of #2305

[waketzheng]

typing is part of the standard library and is already imported on line 5, so there is no need to put it under TYPE_CHECKING.

[inchang-ing]

Both points are addressed in 3502453: the changelog entry now references #2307, and Literal is imported together with TYPE_CHECKING at module level instead of a TYPE_CHECKING-local import.

pytest tests/migrations/test_migration_graph.py (2 passed) and ruff check pass locally.