What to test, per artifact

Coverage matrix for Databricks artifact types — what matters, at which tier, and which tool asserts it.

Use this matrix to decide what a healthy suite covers for each Databricks artifact type. Tiers refer to the test pyramid.

SQL / warehouse queries

Correctness of logic (joins, window functions, attribution CTEs): local DuckDB tier with committed fixtures. Fast enough to run on every save.
Engine parity: the same statement on DuckDB and Photon must return identical rows — run a parity suite with golden results. Catches dialect divergence (QUALIFY, timestamp zones, null ordering) before production.
Performance smoke (live, optional): statement completes within a budget on a serverless warehouse.

Schema contracts: column names, types, nullability against a committed contract file. Local against DuckDB-created tables; live against information_schema.
Data quality: row counts, uniqueness, null rates, freshness. Express these as BDD Then steps or SQL predicates.
Constraint / partition layout (live): DESCRIBE DETAIL assertions.

End-to-end (live, @dlt): upload fixture files to the ingest location, start a pipeline update, poll to a terminal state, assert the target table contents.
Rescue-data drift: assert _rescued_data is empty after the run — a non-empty rescue column means the ingestion contract drifted.
Expectations: pipeline expectations (data-quality rules) pass; failed expectation counts are zero.

Run-to-terminal (live, @jobs): run-now an existing job or runs/submit a notebook, wait with a timeout, assert SUCCESS and inspect task output.
Task logic: extract pure logic out of notebooks into packages and unit test it locally — the notebook itself should be a thin shell.

Grant assertions (live): a scoped service principal CAN read allowed tables and CANNOT read denied ones (expect PERMISSION_DENIED). Least privilege is a test, not a hope.
Object existence: catalogs, schemas, volumes referenced by config exist.

Credential lifecycle (live, @lakebase): OAuth → credential exchange succeeds, short-lived credentials refresh, connection pool survives the swap.
Migrations: apply idempotently against a scratch branch.

dbt build with tests against a scratch schema (live) — sources must match the tables the ingestion pipeline actually creates.
Source contract: a boundary test asserting dbt _sources.yml table names equal the names emitted by the ingestion code, so the two can never drift.

Golden vectors: freeze expected aggregate/stat outputs for committed fixtures; every engine and every refactor must reproduce them exactly (the same philosophy as assignment conformance vectors in packages/testkit).
Guardrails: SRM χ² and metric-decline checks fire on skewed fixtures and stay silent on balanced ones.