SQLAlchemy doesn't mark a parent as dirty when only its children
(``TableColumn`` / ``SqlMetric`` on ``SqlaTable``) are modified.
Continuum's UnitOfWork only creates operations for entities in
``session.dirty``, so a column-only edit produces shadow rows in
``table_columns_version`` but no parent shadow row in
``tables_version``. ``VersionDAO.list_versions`` queries the parent
shadow, so the version dropdown is empty for child-only saves —
exactly the failure mode reported when "I edited a column description
but no version appeared."
Extends ``register_baseline_listener`` with a new before-flush hook
``_force_parent_dirty_on_child_change`` that walks the existing
``_child_to_parent_registry`` and ``attributes.flag_modified(parent,
<first non-excluded versioned column>)`` whenever a versioned child
is dirty / new / deleted but the parent's own scalars haven't been
touched. The flag puts the parent in ``session.dirty`` so Continuum's
UoW creates a parent UPDATE operation; the resulting shadow row's
scalar columns mirror the previous version (only the children
actually changed), and the row exists to anchor the transaction in
the parent's version chain.
``SkipUnmodifiedPlugin._is_no_op_update`` is updated in this commit's
predecessor to recognize the "scalars match but children dirty" case
via ``_has_dirty_versioned_children`` so the forced parent UPDATE
isn't skipped.
Integration coverage: ``test_dataset_column_edit_creates_parent_version``.
Continuum's no-op suppression compared post-flush column values
byte-for-byte against the previous live shadow row. For
``Dashboard.json_metadata`` that produced false-positive version rows
on saves where the user authored nothing — the frontend re-stamps
``map_label_colors`` (regenerated from the ``LabelsColorMap``
singleton) on every save, plus ``chart_configuration`` /
``global_chart_configuration`` / ``show_chart_timestamps`` /
``color_namespace`` (derived from the current chart set), so two
consecutive identical saves produce different bytes for the column.
The diff engine already excluded those keys via
``DASHBOARD_JSON_METADATA_AUDIT_KEYS`` when computing change records;
the skip-plugin diverged.
Adds a ``_COLUMN_NORMALIZERS`` registry keyed on
``(class_name, column_name)`` that maps to a per-column normalizer
applied to both pre- and post-image before equating. The first
entry parses ``Dashboard.json_metadata`` as JSON and drops the
audit-key set before comparing. The same registry is the extension
point for analogous transient fields on charts and datasets.
Promotes ``_DASHBOARD_JSON_METADATA_AUDIT_KEYS`` to a public name
(``DASHBOARD_JSON_METADATA_AUDIT_KEYS``) so the skip-plugin can import
it from ``superset.versioning.diff`` without reaching across a
leading-underscore boundary.
Integration coverage: ``test_map_label_colors_only_change_does_not_create_version``.
Locks in the no-op-suppression behavior implemented by
``SkipUnmodifiedPlugin`` (which lives in ``superset/versioning/factory.py``
shipping with the foundation commit). Five integration tests:
1. Owners-only edit doesn't mint a version row — exercises the
case where every dirty column is an excluded relationship.
2. Re-save with identical scalar values doesn't mint a row —
exercises the json_metadata re-serialise path where
``set_dash_metadata`` rewrites the column to a different byte
sequence with identical parsed content; the plugin must compare
post-flush values against the prior shadow row to detect this.
3. Real scalar change DOES mint a row — guards against the plugin
over-suppressing.
4. Same assertion on a Slice (covers the ``String`` column path on
a different entity type).
5. ``json_metadata`` sub-key edit DOES mint a row — covers the
``MediumText`` column path past the plugin's content-equality
check.
Tests are designed so a column-type change in the parent entities
(e.g. flipping ``json_metadata`` from ``MediumText`` to ``JSON``)
will fail one of these if the plugin's Python ``!=`` comparison
breaks for the new type.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
