| name | rocky-new-adapter |
| description | Adding a new warehouse or source adapter crate to the Rocky engine. Use when introducing a new target (e.g. BigQuery, Redshift, ClickHouse) or source (e.g. Airbyte, Singer). Covers crate layout, trait implementation, dialect specifics, auth, conformance tests, and factory registration. |
Adding a new Rocky adapter
Rocky adapters are Cargo crates that implement traits from rocky-adapter-sdk. Existing examples:
| Kind | Crate | What it talks to |
|---|
| Target (warehouse) | rocky-databricks | Databricks SQL Statement Execution API + Unity Catalog |
| Target (warehouse) | rocky-snowflake | Snowflake REST API |
| Target (local) | rocky-duckdb | Embedded DuckDB |
| Source | rocky-fivetran | Fivetran REST API |
When to use this skill
- Adding a new target adapter (warehouse): e.g. BigQuery, Redshift, ClickHouse, Iceberg
- Adding a new source adapter: e.g. Airbyte, Singer, Stitch, database CDC
- Forking an existing adapter to support a cousin backend (e.g. Databricks → Databricks Serverless)
Crate scaffold
Create engine/crates/rocky-<name>/ with a Cargo.toml registered in the workspace (engine/Cargo.toml members list). Use rocky-databricks or rocky-snowflake as a template for a warehouse adapter, rocky-fivetran for a source adapter.
Warehouse adapter files (mirror rocky-databricks/src/)
| File | Responsibility |
|---|
lib.rs | pub mod exports + crate-level docs |
adapter.rs | impl Adapter for <Name>Adapter — the main trait from rocky-adapter-sdk::traits |
dialect.rs | impl SqlDialect — dialect-specific SQL (quoting, type widening rules, insert_overwrite_partition, MV/dynamic-table DDL) |
connector.rs | REST client or native connection — use reqwest for HTTP-based adapters |
auth.rs | Credential handling (PAT, OAuth M2M, key-pair JWT, etc.) with env var auto-detection |
catalog.rs | Catalog/schema CRUD (CREATE/ALTER/DESCRIBE) |
permissions.rs | GRANT/REVOKE + SHOW GRANTS parsing |
batch.rs | Bulk operations (e.g. information_schema UNION ALL batching) |
throttle.rs | Rate limiting / adaptive concurrency if the backend needs it |
types.rs | Native → internal type mapping |
Source adapter files (mirror rocky-fivetran/src/)
| File | Responsibility |
|---|
lib.rs | Exports |
adapter.rs | impl SourceAdapter |
client.rs | Async REST client |
connector.rs | Source discovery + filtering |
schema.rs | Schema config parsing (often nested JSON) |
sync.rs | Sync detection / staleness checks |
Trait implementation
Traits live in engine/crates/rocky-adapter-sdk/src/traits.rs. The key ones:
Adapter — warehouse adapter core: execute_sql, describe_table, list_tables, create_catalog, …SqlDialect — SQL generation specifics. Pay special attention to:
is_safe_type_widening() — allowlist for ALTER TABLE type changes; conservative by defaultinsert_overwrite_partition() — returns Vec<String> so multi-statement transactions work (Snowflake needs BEGIN/DELETE/INSERT/COMMIT)- Materialized views + dynamic tables DDL
SourceAdapter — source discovery and sync staleness detectionAuthenticator — if there's a credential flow worth unifying
Use #[async_trait] for all async trait impls.
Conformance tests
rocky-adapter-sdk/src/conformance.rs hosts a reusable test suite. Wire your adapter into it so rocky test-adapter <name> (the CLI command) validates the baseline behavior: CREATE/DROP, incremental copy, schema drift, permission reconcile.
Test pattern: add an integration test file in your crate that instantiates the adapter and runs the conformance suite. For credential-free local testing, gate behind an env var — see how rocky-databricks skips network tests when DATABRICKS_HOST is unset.
Factory registration
The CLI + config parser needs to know your adapter type string. Search for where existing adapter types are registered (typically in rocky-core/src/config.rs or the CLI's adapter dispatch) and add yours. The config block in rocky.toml will then accept:
[adapter]
type = "<name>"
host = "${MY_HOST}"
Auth patterns to copy
- PAT-first, OAuth M2M fallback — see
rocky-databricks/src/auth.rs. Tries PAT from env, falls back to client_credentials flow if PAT is absent.
- Multi-method priority — see
rocky-snowflake/src/auth.rs. OAuth (pre-supplied) > RS256 key-pair JWT > password.
- Basic auth — see
rocky-fivetran/src/client.rs. Simplest case — base64 of key:secret.
- SQL identifier validation — always route catalog/schema/table names through
rocky-sql/validation.rs before interpolation.
Dialect gotchas
The SqlDialect trait is where adapters actually diverge. Things to get right:
- Quoting: backticks vs double-quotes vs brackets.
rocky-databricks uses backticks for principals; most warehouses use the standard double-quote identifier.
- Safe type widening allowlist:
is_safe_type_widening(from, to) controls whether Rocky will ALTER TABLE or fall back to full refresh. Be conservative — data loss from an unsafe cast is worse than an extra full refresh.
- Partition overwrite atomicity: Databricks has
INSERT INTO ... REPLACE WHERE as a single atomic op. Snowflake and DuckDB need BEGIN; DELETE; INSERT; COMMIT; — hence the Vec<String> return type.
- MV / dynamic table semantics: Databricks materialized views and Snowflake dynamic tables have different refresh models. Read
engine/CLAUDE.md for the canonical DDL strings.
Adapter SDK process helper
rocky-adapter-sdk/src/process.rs provides helpers for out-of-process adapters (spawning an external binary, parsing JSON, handling exit codes). Useful if you're wrapping a dbt-like adapter or a CLI tool instead of speaking a REST API directly.
Shipping checklist
Commit style
feat(engine/rocky-bigquery): scaffold BigQuery target adapter
feat(engine/rocky-bigquery): implement SqlDialect + conformance tests
feat(engine/rocky-bigquery): OAuth service-account auth
docs(engine): add BigQuery adapter reference
