[Ossie][Databricks] Add Databricks Unity Catalog Metric View converter

[jbonofre]

Add Databricks Unity Catalog Metric View converter

Adds a bidirectional, offline converter between Apache Ossie semantic models and Databricks Unity Catalog Metric Views (YAML v1.1), filling the Databricks spoke in the hub-and-spoke architecture (DATABRICKS is already a listed vendor in converters/index.md). Validated against the live Databricks Metric View engine.

What's included

• Export (ossie-databricks export): Apache Ossie semantic model -> Metric View (one fact source + a nested joins tree, flattened dimensions, measures).
• Import (ossie-databricks import): Metric View -> Apache Ossie. Metric-View-only features (filter, window, format, rely, cardinality, parameters, materialization) are preserved in custom_extensions[DATABRICKS], so MV -> Apache Ossie -> MV is lossless.
• A unified ossie-databricks CLI and a string-in / string-out Python API.
• A README with a mapping table and limitations.

Design highlights

• Round-trip fidelity via the custom_extensions[DATABRICKS] stash -- the approach the converter guide recommends.
• Join tree, not flat joins. The Apache Ossie relationship graph is reassembled into the Metric View's nested fact + joins tree; a dataset reached by multiple paths (a diamond) is fanned out into one aliased join per path; cyclic graphs are rejected.
• Cardinality from direction. many_to_one / one_to_many is derived from the relationship from/to direction relative to a selectable grain (--source). Multiple facts are supported only narrowly -- when they share a conformed dimension (named as the source); a general galaxy schema (facts each with their own dimensions) or facts sharing no dimension can't map to one Metric View and are rejected with a clear error, never silently mis-converted.
• Key bridge. A primary_key/unique_keys whose columns cover a join key becomes rely: {at_most_one_match: true}, and vice versa.
• On export, joined-table columns are alias-qualified in dimension/measure expressions; fact columns in measures are emitted bare (the DBR idiom).
• Fails loudly, never silently. A join condition an Apache Ossie relationship can't represent (a non-equi or cross join), a cyclic or ambiguous-grain graph, and an unsupported version are rejected with a clear ConversionError; anything dropped (a foreign-vendor extension, an Apache-Ossie-only annotation, a non-DATABRICKS/ANSI_SQL dialect) emits a warning naming the object.

Conventions followed

• Lives at converters/databricks/, packaged like the dbt/gooddata spokes: pyproject.toml (hatchling, Apache-2.0), src/ossie_databricks/, tests/.
• Every source file carries the ASF license header, matching the sibling spokes.
• On export, dialect selection prefers DATABRICKS, falls back to ANSI_SQL.
• The only runtime dependency is PyYAML; Python 3.11+.

Testing

cd converters/databricks
pip install -e . && pip install pytest hypothesis
python -m pytest tests/

Example-based unit tests plus Hypothesis property-based round-trip tests in both directions, with a normalized TPC-DS fixture as a baseline.

Notes

• Targets Databricks Unity Catalog Metric View YAML v1.1.
• Foreign-vendor custom_extensions have no slot in the Metric View YAML; they are dropped on export with a warning.

Credits

@jackstein21 is credited as a co-contributor (see the commit's Co-authored-by trailer), including the property-based round-trip testing approach.

[jbonofre]

@Haoranli503 I suggest to rename everything to Ossie.

Feel free to close this PR to create a new one (now that the transfer is done).

[jbonofre]

@Haoranli503 I would rename this file to ossie_to_metric_view.py now.

[jbonofre]

Here, we flatten every dataset's fields into the Metric View's single dimensions list, it tracks names it has already emitted in a seen_dims set and checks for collisions case-insensitive.

The same pattern applies to measures in _convert_metric (which shares the seen_dims/seen_names set, so a measure can also collide with a dimension:

if name.lower() in seen_names:
    _warn(scope, "measure name collides with another dimension/measure; rename before use")
seen_names.add(name.lower())

So on a name collision it emits a warning to stderr but still appends the duplicate to the output.

I suggest to made this a ConversionError to be consistent with the rest of the converter: the user gets told exactly which field to rename before they hand a broken file to Databricks.

@Haoranli503 ^^