Assay for Python

Native analytical SQL with Arrow results. Version 0.1.0a1 is an experimental alpha. The packaged fast engine is tested, not proved, and does not include linked verified execution. Separate proof research describes its own narrower scope. Only read Parquet files you trust.

Install the alpha from PyPI:

python -m pip install --only-binary=:all: 'assaydb==0.1.0a1'

A supplied wheel can also be installed with python -m pip install --only-binary=:all: /path/to/assaydb.whl, using its actual filename. No Rust compiler is needed to install a wheel. Python 3.10 or later and PyArrow are required.

import assaydb

with assaydb.connect() as db:
    db.execute("CREATE TABLE events (id INTEGER, label TEXT)")
    db.execute("INSERT INTO events VALUES (9007199254740993, 'hello'), (2, 'world')")
    result = db.sql("SELECT id, label FROM events ORDER BY id")

print(result.to_arrow())
# result.to_pandas()  # Install pandas separately if needed.
# result.to_polars()  # Install polars separately if needed.

Read local files with read_parquet('events.parquet') and read_csv('events.csv') in SQL. CSV has a required header and all columns start as TEXT; use CAST(id AS INTEGER) for integers. Empty fields are empty strings, so use NULLIF(column, '') explicitly when wanted. Quoted commas and newlines are supported. Paths must name local regular files, without URLs or globs; keep files unchanged while queries run. Connections enable reading by default; connect(allow_local_files=False) refuses these functions and is not a directory sandbox. CSV is a bounded import intended for small files; Parquet uses scans.

execute() and sql() execute one statement. Query results provide to_arrow(), to_pandas(), to_polars(), rowcount and is_query. DML results provide affected rowcount; asking for a DML table raises SQLError. Native operations release the GIL, and one connection serializes its operations. Connections own independent temporary catalogs; completed results survive close.

These are not persistent database files or Python DB-API connections. Parameters, transactions and streaming cursors are not implemented by this wrapper. Unsupported SQL raises UnsupportedError, invalid SQL raises SQLError, limits raise ResourceError, and closed connections raise ConnectionClosedError. All derive from AssayError; InternalError reports unexpected engine/storage failures. A caught panic closes the connection. Refusal advice appears in the exception text and .hint, with the original diagnostic in .native_message. Advice never rewrites or retries SQL. REAL casts may round large integers; INTEGER casts truncate fractional values.

Defaults: one worker, 256 MiB accounted query memory, 64 MiB catalog allowance and 64 MiB serialized result limit. Set workers, memory_bytes, catalog_bytes and max_result_bytes explicitly to override them. Catalog allowance must be at most one quarter of query memory. directory= chooses the parent of a private temporary catalog directory. Closing removes only that session's own files. These limits do not cap total process memory.

Arrow IPC preserves engine types, binary values and NULLs, while copying results into Python-owned memory. It is not zero-copy execution. pandas conversion uses Arrow dtypes by default to preserve nullable integers above 2^53; supplying your own types_mapper changes that conversion policy. Polars converts the resulting Arrow table; neither optional dependency executes SQL for Assay.

Current local wheel and rebuilt-sdist checks pass all eleven tests on fresh macOS arm64 Python 3.14, including the nested-INSERT regression. Earlier native CI passed Linux x86_64 and macOS arm64 Python 3.10–3.13 installs; native CI records the exact revision tested by the expanded Python 3.10–3.14 matrix. Package tests and wheel tags do not establish full SQL compatibility.

See the notebook and build instructions.