Building a Custom Adapter

Rocky talks to warehouses through a small set of traits in the rocky-adapter-sdk crate. Implementing those traits gives you a working warehouse adapter — the same way rocky-databricks, rocky-snowflake, rocky-bigquery, and rocky-duckdb are wired today.

This guide walks a Rust developer from “I want a ClickHouse adapter” to a compiling skeleton with passing tests in roughly fifteen minutes. The runnable skeleton lives at examples/playground/pocs/07-adapters/06-rust-native-adapter-skeleton/ and is shaped after ClickHouse, but the same shape works for Trino, Redshift, StarRocks, MotherDuck, or any SQL warehouse Rocky doesn’t ship in-tree.

When to reach for the SDK

The adapter SDK is the right tool when:

• The warehouse you need is not in the in-tree adapter list (Databricks, Snowflake, BigQuery, DuckDB).
• You need a forked variant of an existing adapter (e.g. Databricks Serverless on top of rocky-databricks).
• You want to embed Rocky in a tool that owns its own warehouse client and would rather wrap it than spawn rocky as a subprocess.

If your warehouse already ships in-tree, use it directly via [adapter] in rocky.toml. If you want adapters in a non-Rust language (Python, Go, Node), see the process adapter protocol — JSON-RPC over stdio. The POC at pocs/07-adapters/04-custom-process-adapter/ walks that pattern.

The trait surface

Public traits live in engine/crates/rocky-adapter-sdk/src/traits.rs. The two you must implement for any warehouse adapter are WarehouseAdapter and SqlDialect. The rest are opt-in by capability.

Trait	Required?	What it does	Key methods (full surface in rocky-adapter-sdk/src/traits.rs)
WarehouseAdapter	yes	Execute SQL against the warehouse	dialect, execute_statement, execute_query, describe_table, table_exists, close
SqlDialect	yes	Generate warehouse-specific SQL	name, format_table_ref, create_table_as, insert_into, merge_into, describe_table_sql, drop_table_sql, create_catalog_sql, create_schema_sql, row_hash_expr, tablesample_clause, select_clause, watermark_where, insert_overwrite_partition
DiscoveryAdapter	no	Enumerate connectors / tables in a source system	discover
GovernanceAdapter	no	Tags, grants, catalog/schema lifecycle	set_tags, get_grants, apply_grants, revoke_grants
BatchCheckAdapter	no	Batched data-quality queries	batch_row_counts, batch_freshness
LoaderAdapter	no	File ingestion (CSV, Parquet, JSONL)	load, supported_formats
TypeMapper	no	Cross-warehouse type normalization	normalize_type, types_compatible

Each opt-in trait is gated by a flag in AdapterCapabilities. Set the flag, implement the trait, and Rocky’s planner picks up the new behavior automatically.

When each method is called

• execute_statement — every DDL / DML Rocky generates: CREATE TABLE, INSERT INTO, MERGE INTO, ALTER TABLE, DROP TABLE, partition replace.
• execute_query — EXPLAIN, DESCRIBE, row-count assertions, the rocky compile-time SELECT 1 connectivity check.
• describe_table — drift detection (rocky drift), contract validation, the column-list step before generating an incremental insert.
• table_exists — full-refresh-vs-create branching at the start of a materialization.
• dialect() methods — every SQL string Rocky emits is composed by a dialect call. Identifier validation lives here.

Worked example: a ClickHouse-shaped skeleton

The POC at examples/playground/pocs/07-adapters/06-rust-native-adapter-skeleton/ is a compiling, tested starter. To run it:

git clone https://github.com/rocky-data/rocky.git
cd rocky/examples/playground/pocs/07-adapters/06-rust-native-adapter-skeleton
./run.sh

This runs cargo check, the unit tests, and a demo binary that prints the SQL the adapter would have sent to a real warehouse.

Crate layout

adapter/
├── Cargo.toml            # Path-dep on rocky-adapter-sdk; standalone (not in workspace)
├── src/lib.rs            # SkeletonAdapter, SkeletonDialect, MockBackend, tests
└── examples/demo.rs      # End-to-end driver

Adapter struct + manifest

pub struct SkeletonAdapter {
    backend: Arc<dyn Backend>,
    dialect: SkeletonDialect,
}

impl SkeletonAdapter {
    pub fn manifest() -> AdapterManifest {
        AdapterManifest {
            name: "skeleton".into(),
            version: env!("CARGO_PKG_VERSION").into(),
            sdk_version: SDK_VERSION.into(),
            dialect: "skeleton".into(),
            capabilities: AdapterCapabilities {
                warehouse: true,
                discovery: false,
                governance: false,
                batch_checks: false,
                create_catalog: false,  // ClickHouse has no catalogs
                create_schema: true,    // ClickHouse calls these "databases"
                merge: false,           // No MERGE — use incremental instead
                tablesample: true,
                file_load: false,
            },
            auth_methods: vec!["basic".into(), "token".into()],
            config_schema: serde_json::json!({ /* ... */ }),
        }
    }
}

Capability flags are not cosmetic — they gate behavior. merge: false makes Rocky’s planner refuse strategy = "merge" configs against this adapter at validate time rather than failing mid-run. create_catalog: false makes auto_create_catalogs = true surface a clear “warehouse doesn’t support catalogs” error instead of emitting broken SQL.

Backend abstraction

The skeleton hides the actual warehouse client behind a small Backend trait so tests can substitute an in-memory mock:

#[async_trait]
pub trait Backend: Send + Sync {
    async fn execute(&self, sql: &str) -> AdapterResult<()>;
    async fn query(&self, sql: &str) -> AdapterResult<QueryResult>;
    async fn describe(&self, table: &TableRef) -> AdapterResult<Vec<ColumnInfo>>;
    async fn exists(&self, table: &TableRef) -> AdapterResult<bool>;
}

The production impl wraps clickhouse::Client (or reqwest::Client for warehouses without a typed driver). The test impl is MockBackend — a HashMap plus a statement log so tests can assert on the SQL the dialect produced.

Dialect implementation

SqlDialect is where most adapter divergence lives. The skeleton’s format_table_ref shows the two patterns you almost always need: drop arguments your warehouse doesn’t have, and validate every identifier you splice into SQL.

fn format_table_ref(
    &self,
    _catalog: &str,    // ClickHouse has no catalogs — drop on the floor
    schema: &str,
    table: &str,
) -> AdapterResult<String> {
    validate_ident(schema)?;
    validate_ident(table)?;
    Ok(format!("`{schema}`.`{table}`"))
}

Methods worth thinking carefully about:

• merge_into — return AdapterError::not_supported("merge_into") if your warehouse has no MERGE. Rocky’s planner sees the capability flag and won’t generate merge plans, but a defensive impl still helps if someone bypasses the planner.
• insert_overwrite_partition — returns Vec<String> because some warehouses need a multi-statement transaction (Snowflake’s BEGIN; DELETE; INSERT; COMMIT). The runtime executes them in order and rolls back on partial failure.
• row_hash_expr — used for change detection. ClickHouse uses sipHash128(tuple(...)); if you want cross-warehouse comparable hashes, look at how rocky-bigquery and rocky-snowflake agree on a stable encoding.
• watermark_where — the standard incremental filter (col > (SELECT max(col) FROM target)). Validate timestamp_col before splicing.

Auth and connection management

The SDK does not prescribe an auth trait — each adapter wires its own. The two patterns worth copying are in-tree:

• engine/crates/rocky-databricks/src/auth.rs — PAT-first, OAuth M2M fallback. Reads ${DATABRICKS_TOKEN}; if absent, falls through to the client_credentials flow with ${DATABRICKS_CLIENT_ID} / ${DATABRICKS_CLIENT_SECRET}. The auto-detection logic is roughly twenty lines.
• engine/crates/rocky-snowflake/src/auth.rs — multi-method priority: pre-supplied OAuth bearer wins, then RS256 key-pair JWT, then password. Each method reads from a distinct ${SNOWFLAKE_*} variable so config files never carry secrets.

Two rules that apply to every adapter regardless of method:

1. Read credentials at config-parse time, not at adapter-construct time. Rocky substitutes ${VAR} references when parsing rocky.toml. Pull the resolved string out of SkeletonConfig; do not re-read env vars from the adapter constructor or tests will collide on shared state.
2. Pool HTTP clients in the adapter struct. reqwest::Client is internally Arc-counted and cheap to clone — construct it once in SkeletonAdapter::new and clone the handle into every request. Don’t construct a new client per call.

For retry and rate-limiting, look at rocky-adapter-sdk/src/throttle.rs (the AIMD adaptive concurrency helper) and rocky-databricks/src/connector.rs for an is_transient / is_rate_limit retry loop.

Testing your adapter

Two test layers, neither of which needs a live warehouse.

Unit tests with a mock backend

The skeleton’s tests assert on the SQL the dialect generated:

#[tokio::test]
async fn execute_statement_round_trips_to_backend() {
    let backend = Arc::new(MockBackend::new());
    let adapter = SkeletonAdapter::new(backend.clone());

    adapter
        .execute_statement("CREATE TABLE foo (id Int64) ENGINE=Memory")
        .await
        .unwrap();

    let log = backend.statement_log().await;
    assert!(log[0].contains("CREATE TABLE foo"));
}

This style covers everything except real network behavior.

Wiremock for HTTP-backed adapters

For adapters that talk to a REST API, the in-tree pattern is wiremock-based — see how rocky-fivetran/src/client.rs is tested. You stand up a MockServer per test, register expected Match::path("/v1/connectors") handlers, and assert your adapter sends the right verbs against the right paths. CI runs without a real Fivetran account.

The conformance harness

rocky-adapter-sdk::conformance::run_conformance(&manifest) returns a ConformanceResult describing which tests apply (based on declared capabilities) and which were skipped. Today this is a test plan, not a live runner — it prints the matrix but does not yet execute trait calls against your adapter. Treat it as a checklist of behaviors your unit tests should cover. Live execution will land in a future SDK release.

Distributing your adapter

The honest answer for now: fork and merge. The adapter registry is statically registered at compile time — there is no dynamic plugin system today. To ship a new adapter:

1. Fork rocky-data/rocky.
2. Drop your crate into engine/crates/rocky-<name>/.
3. Add it to engine/Cargo.toml workspace members and the CLI’s adapter dispatch.
4. Open a PR upstream. The SDK pins the trait shape so the diff stays small — usually a few hundred lines of crate plus one wiring line in the CLI.

Two looser paths if upstreaming isn’t an option yet:

• Vendor the crate. Keep your fork private, ship Rocky internally with your adapter linked in. The in-tree adapters use this same model — they’re just upstreamed.
• Process adapter (any language). If you want to escape Rust entirely, the JSON-RPC stdio protocol in rocky-adapter-sdk/src/process.rs works today — see pocs/07-adapters/04-custom-process-adapter/ for a working Python adapter against SQLite.

A dynamic registration path (declarative config + crates.io discovery) is on the roadmap but unscheduled. Until it lands, the SDK’s job is to keep the trait surface stable enough that your fork is forward-compatible.

Gotchas worth knowing about

These are real and surface during implementation — flagging them up front so you don’t lose half a day debugging.

• Two trait surfaces exist today. rocky-adapter-sdk/src/traits.rs is the public, marketed contract. rocky-core/src/traits.rs is a richer superset that the in-tree adapters currently use (it adds methods like execute_statement_with_stats, ExplainResult, retention APIs). For new out-of-tree adapters, target the SDK — it’s the contract that will stay stable. The in-tree richer surface is migrating toward the SDK over time.
• Identifier validation is not optional. Anything you splice into SQL must pass [A-Za-z0-9_]+ (or your warehouse’s equivalent). The skeleton’s validate_ident shows the pattern. SQL-injection-bearing string literals were the subject of a real CVE-class fix — don’t reinvent that hole.
• The catalog field in TableRef is always present. Warehouses without catalogs (ClickHouse, Postgres, MySQL) get an empty string. Your dialect’s format_table_ref is responsible for dropping it.
• AdapterError is intentionally type-erased. Use AdapterError::msg(...) for ad-hoc errors, AdapterError::new(my_err) to wrap an std::error::Error, and AdapterError::not_supported("method_name") for capabilities your warehouse doesn’t have. Don’t reach for thiserror inside the trait impl — the SDK boxes everything.

Next steps

• Browse the skeleton POC source — adapter/src/lib.rs is meant to be read top-to-bottom.
• Read the adapter concepts page for the architecture overview.
• Look at the in-tree adapters in engine/crates/rocky-{databricks,snowflake,bigquery,duckdb,fivetran}/ for production patterns.
• File an issue on github.com/rocky-data/rocky if a trait method is missing what you need — the SDK is still young and feedback shapes the roadmap.