dbt-frostlake

A dbt adapter for Frostlake, so a dbt project runs locally with zero warehouse cost.

It is a standalone adapter: built directly on dbt-adapters, with its own macros and dialect rules. No other adapter is installed or imported. The transport is frostlake-connector over frostlake, which speak Frostlake's HTTP protocol.

Install

pip install dbt-frostlake

That brings dbt-core, frostlake-connector and frostlake with it. No JVM is needed on the client.

Profile

my_project:
  target: dev
  outputs:
    dev:
      type: frostlake
      host: localhost      # Frostlake HTTP server
      port: 18082
      database: demo_db
      schema: public
      threads: 1

account, user and password are optional: they are accepted so a profile carried over from another warehouse still loads, and ignored — the server has no login.

Start a server for it

The engine is published to Maven Central, so no build is required:

mvn dependency:get -Dartifact=dev.frostlake:frostlake-db:0.2.0
java -cp "<frostlake-db jar>:<its runtime deps>" dev.frostlake.http.DatabaseHttpServer 18082

Requires JDK 17 and a Frostlake engine 0.2.0 or newer. Create the target database once it is up:

CREATE DATABASE IF NOT EXISTS demo_db;

Smoke-tested surface

Against a booted DatabaseHttpServer: dbt debug, dbt seed (typed CSV loads incl. DECIMAL), dbt run (view + table materializations), dbt test (unique, not_null), dbt build, dbt snapshot (check strategy, repeat runs), incremental models (first and subsequent runs), and dbt docs generate (information_schema catalog) — all green.

Dialect notes

Frostlake resolves unquoted identifiers uppercase, so the adapter turns dbt's default quoting off, and seed columns are unquoted unless a project sets quote_columns.

Two engine limitations are worked around in the macros: ALTER VIEW … RENAME TO rejects a qualified target (tables accept one), and CASCADE is accepted on DROP SCHEMA but not on DROP TABLE/DROP VIEW.

A runnable example project lives in frostlake-dbt-demo, a sibling repository: two seeds, two staging views, a table model and five data tests.

Tests

export JAVA_HOME=/path/to/jdk17
export FROSTLAKE_CLASSPATH="/path/to/frostlake-db.jar:<engine deps>"
python3 test/test_adapter.py

Covers the plugin wiring dbt reads on load (adapter and credentials classes, the empty plugin dependency list, the shipped include path), the credential surface and its unique_field connection key, and open() against a live server — handle usability, database/schema context, query tags, and reuse of an already-open connection. The plugin and credential tests need no server; the connection tests skip without FROSTLAKE_CLASSPATH.

Set FL_CORPUS to the engine's testkit directory (an absolute path) and the same run also replays the engine's language-neutral SQL corpus through the adapter's open() (testkit_runner.py); without it, that test skips:

FL_CORPUS=/path/to/frostlake/engine/src/test/resources/testkit python3 test/test_adapter.py

Still to do

Grind dbt's standard acceptance suite (dbt-tests-adapter) green; every failure is either a facade gap or a Frostlake SQL gap worth fixing.