Skip to content
RockyRockyRocky

Roadmap

The trust primitives (compiler, branches, run records, lineage, contracts, cost attribution) are production-grade on Databricks. Here is the state of everything else.

Shipped and production-grade

Section titled “Shipped and production-grade”
  • Databricks is the production target for 2026: SQL Statement API, Unity Catalog, OAuth M2M, adaptive concurrency, and schema-prefix branches (warehouse-native SHALLOW CLONE is a follow-up).
  • The compiler. Typed column-level inference, the diagnostic codes, and the LSP all run the same in CI and in your editor.
  • Branches and run records. Named branches as isolated schemas, and a content-addressed record of each run that you inspect and verify against the ledger with rocky replay. (Re-execution from the pinned record is on the 2026 roadmap.)
  • Cost attribution. Per-model cost on every run record, with [budget] blocks that fail a run on overspend.
  • The AI compile-validate loop. rocky ai generates a model, compiles it, and auto-fixes parse errors before it lands.
  • Deterministic surrogate keys. A model declares a [[surrogate_key]] block, and rocky run injects a dialect-correct hash column at materialization on DuckDB, Databricks, Snowflake, and BigQuery. On a given warehouse the value matches what dbt Core’s dbt_utils.generate_surrogate_key produces over the same columns, so a key joins across a Rocky/dbt Core boundary either way.
  • Named and inline tests. Define a data-quality assertion once in models/test_definitions.toml and apply it by name with [[use_test]], overriding the column, severity, or filter at the use site, the declarative analogue of dbt Core’s generic tests. Inline [[tests]] keep working alongside. References resolve at load, so a typo fails the load instead of silently dropping a check. See data quality checks.
  • Fixture-driven unit tests. A model’s sidecar declares [[test]] blocks with mocked upstream inputs (given) and an expected output (expect). rocky test seeds a fresh in-memory DuckDB, materializes the model against the fixtures, and compares the result, with no warehouse connection. See testing.
  • Config groups. A config group in models/groups/<name>.toml fans shared routing and strategy out to every model that opts in with group = "<name>", so a one-line change flows to the whole group. With enforce = true the group’s fields become binding: a member that locally diverges from the group-controlled schema or strategy fails the load instead of drifting silently.
  • Model tags. A model declares a [tags] block of governance attributes (domain, tier, owner), and a config group can supply a [tags] baseline every member inherits, with the model’s own tags overriding per key. Resolved tags surface on rocky compile --output json for orchestrators to read.
  • rocky emit-sql. Renders the runnable SQL each transformation model would produce, offline with no warehouse connection, to stdout or one file per model. A Rocky project always reduces to plain SQL in dependency order, so adopting Rocky is never a one-way door. See No lock-in.

Beta

Section titled “Beta”
  • Snowflake, BigQuery, and Trino. Connection, execution, and the core run loop work. Conformance coverage is still growing, and we test against live warehouses, so corner cases get reported and fixed quickly. If your enterprise warehouse is Snowflake or BigQuery and you need it production-grade today, open a discussion. We want the failure reports.
  • Iceberg. REST-catalog source discovery works. Content-addressed writes round-trip as Iceberg through Delta UniForm, end-to-end.

On the 2026 roadmap

Section titled “On the 2026 roadmap”
  • Iceberg-native writes without the Delta intermediate.
  • The wider AI surface. The loop is shipped; the larger workflow (rename a column, regenerate the downstream models and their assertions, run the tests, ship) is the next step.
  • A semantic layer. Rocky’s typed IR is the right home for one. Until it exists, integrate with Cube, the dbt Semantic Layer, or your existing metric store.

Handled by an integration, not built in

Section titled “Handled by an integration, not built in”
  • Extraction from SaaS sources. Use Fivetran, Airbyte, Stitch, or warehouse-native CDC. Rocky discovers what they land and takes over from there.
  • Orchestration beyond Dagster. Dagster is first-class, and rocky serve covers small standalone teams. Airflow and Prefect call the rocky binary like any other CLI; native integrations are not shipped.

Tell us where it hurts

Section titled “Tell us where it hurts”

The roadmap is shaped by where production pipelines are getting hurt. If one of these gaps blocks your team, open a discussion.