dosi-engine

Python bindings for osi-engine: load an OSI semantic model once, then list metrics/dimensions, compile metric queries to dialect SQL, and execute them against a warehouse — all in-process.

from dosi_engine import Engine, QueryError

engine = Engine(
    model_path="model.yaml",
    connections={"warehouse": {"type": "snowflake", "account": "...", "password": "..."}},
)

engine.metrics()                     # [{"name", "kind", "datasets", ...}]
engine.compile({"metrics": ["order_count"],
                "group_by": [{"field": "orders.status"}]},
               dialect="snowflake")  # {"dialect": "snowflake", "sql": "..."}
engine.execute({"metrics": ["order_count"]},
               connection="warehouse", timeout_secs=60)
#   {"dialect", "sql", "columns", "rows": [{col: val}], "row_count"}

The list/compile/execute payloads are the same machine contract as osi --format json and the REST API; errors raise ModelError / QueryError / ExecuteError (see dosi_engine.errors) carrying the same structured fields (code, candidates, hint) as the CLI/REST JSON error contract. QueryError.candidates lists valid alternatives for unknown/ambiguous names, so agentic callers can retry without parsing prose.

Semantics worth knowing

• One Engine = one compiled model. Construction runs load → validate → compile and fails fast with ModelError. Engines are thread-safe; compile and execute release the GIL.
• Connections are an in-memory name -> datasource settings mapping in the Datus agent.yml datasources: vocabulary. File discovery and loading belong to the Rust CLI/server; the Python binding never reads or writes a connections file.
• execute without connection runs on the runtime connections mapping's default: true profile when one exists, else on local DuckDB (db_path= file, or in-memory).
• Timeouts propagate cancellation to cooperative executors. Athena uses a bounded cleanup window and reports confirmation in execution_info. Other drivers may keep running after the caller's deadline. Successful Athena results and execution errors retain one diagnostic entry per query.
• DuckDB runs in process and ships inside the wheel — nothing to install for local execution. Warehouse drivers are compiled in too (exec-all), so no client libraries either.

Building

Requires a Python ≥ 3.12 interpreter discoverable by pyo3; on hosts whose default python3 is older, set PYO3_PYTHON:

cd crates/dosi-py
PYO3_PYTHON=$(command -v python3.12) uvx maturin build --release -o ../../target/wheels

Development loop against a venv: uvx maturin develop (with the venv active). Rust-side tests: cargo test -p dosi-py; Python smoke tests (after maturin develop): pytest tests/python/.