sqlframe-gizmosql

GizmoSQL adapter for SQLFrame - a PySpark-like DataFrame API for GizmoSQL.

Overview

This package provides a GizmoSQL backend for SQLFrame, allowing you to use PySpark-compatible DataFrame operations against a GizmoSQL server. GizmoSQL is a database server that uses DuckDB as its execution engine with an Arrow Flight SQL interface.

As of v1.4.0, sqlframe-gizmosql runs on adbc-driver-gizmosql 2.0, powered by the new native Go GizmoSQL ADBC driver. The API is unchanged, and features like immediate DDL/DML execution, RETURNING support, gizmosql:// connection URIs, and OAuth/SSO are now provided by the shared Go driver library used across all language bindings.

Installation

pip install sqlframe-gizmosql

Requirements

• Python >= 3.10
• GizmoSQL server running and accessible

Quick Start

First, start a GizmoSQL server (see Running GizmoSQL with Docker below), then:

from sqlframe_gizmosql import GizmoSQLSession

# Create a session connected to GizmoSQL
session = GizmoSQLSession.builder \
    .config("gizmosql.uri", "gizmosql://localhost:31337") \
    .config("gizmosql.username", "gizmosql_user") \
    .config("gizmosql.password", "gizmosql_password") \
    .config("gizmosql.tls_skip_verify", True) \
    .getOrCreate()

# Create a DataFrame from a SQL query
df = session.sql("SELECT 1 as id, 'hello' as message")

# Show the results
df.show()

# Use PySpark-like DataFrame API
df2 = session.createDataFrame([
    (1, "Alice", 30),
    (2, "Bob", 25),
    (3, "Charlie", 35),
], ["id", "name", "age"])

# Filter, select, and aggregate
result = df2.filter("age > 25").select("name", "age")
result.show()

# Group by and aggregate
df2.groupBy("age").count().show()

Reading Local Files

spark.read.json(), .csv(), .parquet(), and .load() automatically detect paths that exist on the client filesystem, parse them locally with pyarrow, and bulk-load them into a session-scoped temporary table over ADBC — no server-side file access, no admin gating, and no code changes versus stock PySpark:

df = spark.read.json("local_file.jsonl")   # parsed client-side, bulk-ingested
df.groupBy("age").count().show()

Paths the client can't see (or reads with .option("serverSideRead", True)) fall back to the previous behavior: a server-side read_json()/read_csv()/read_parquet() query against the server's filesystem.

Document-shaped JSON (jsonDocument)

For JSON whose structure varies across records (e.g. MongoDB collection exports, EHR/form documents), typed schema inference is a trap: the inferred schema unions every field ever seen, and both client-side Arrow and server-side DuckDB explode memory materializing millions of mostly-null nested columns (a 100 MB file can OOM a 24 GB client this way). Pass .option("jsonDocument", True) to skip parsing entirely — each line ships as a raw string and lands in a single DuckDB JSON column, in seconds and in constant memory:

from pyspark.sql import functions as F

df = spark.read.option("jsonDocument", True).json("patient_forms.jsonl")  # seconds, any nesting

# Query with the standard PySpark JSON idiom (translated to DuckDB json functions):
typed = df.select(
    F.get_json_object(df["json"], "$._id").alias("form_id"),
    F.get_json_object(df["json"], "$.finalized").cast("boolean").alias("finalized"),
)
typed.groupBy("finalized").count().show()

For heavier use, selectively type the fields you query once, server-side, into a real table/view — from_json() extracts only the fields named in its structure argument and ignores the rest of the document (use json_structure() on a sample row to generate a starting spec):

session.sql("""
    CREATE TABLE patient_forms_typed AS
    SELECT from_json(json, '{"_id": {"_oid": "VARCHAR"}, "finalized": "BOOLEAN"}') AS doc,
           json AS raw
    FROM my_raw_table
""")

Options honored by jsonDocument reads: multiLine (whole file as one document), filename (add a filename column), encoding.

Bulk Ingestion

For bulk-loading in-memory data (e.g. a pyarrow.Table you already have), use session.ingest() instead of session.createDataFrame() or a server-side read_ndjson()/read_json() query. It streams a pyarrow.Table/RecordBatch/RecordBatchReader or pandas.DataFrame to GizmoSQL as columnar Arrow batches over ADBC's adbc_ingest(), reusing the session's existing connection — no server-side file access and no admin gating required, and dramatically faster than either of the alternatives below for bulk data:

import pyarrow.json as paj

# Parse the file into Arrow client-side, then bulk-load it into a real GizmoSQL table.
arrow_table = paj.read_json("large_file.jsonl")
df = session.ingest("patient_forms", arrow_table, mode="create")

# The rest of the PySpark-style workflow is unchanged.
df.show()
session.sql("SELECT COUNT(*) FROM patient_forms").show()

mode controls how existing data is handled:

A Table/RecordBatch of any size is automatically re-batched and streamed as many small Flight messages inside a single ingest stream (the Arrow schema is sent once), so you don't need to size batches by hand around GizmoSQL's default 16 MiB gRPC max message size. Append modes stage through a temporary table plus one server-side INSERT, so a mid-stream failure can never duplicate rows.

For very wide or deeply nested schemas (e.g. JSON with large nested arrays), row-based bisection can hit a wall: the Arrow schema message itself — sent once per Flight stream, before any row data — can already approach the limit on its own, in which case no amount of splitting rows helps. session.ingest() raises a clear error pointing you at the fix: raise the connection's own max message size via the gizmosql.max_msg_size builder config (or the matching activate(..., max_msg_size=...) kwarg):

session = GizmoSQLSession.builder \
    .config("gizmosql.uri", "gizmosql://localhost:31337") \
    .config("gizmosql.max_msg_size", 128 * 1024 * 1024) \
    .getOrCreate()

Use session.ingest() for bulk/large data. It's not a replacement for session.createDataFrame(), which remains the right tool for small, literal in-memory data, or for session.sql("... read_ndjson(...) ..."), which asks the GizmoSQL server to read a file from its own filesystem (useful when the server already has direct access to the file, but slower and gated for non-admin/token sessions on large files).

Configuration

The session can be configured using the builder pattern:

session = GizmoSQLSession.builder \
    .config("gizmosql.uri", "gizmosql://localhost:31337") \
    .config("gizmosql.username", "gizmosql_user") \
    .config("gizmosql.password", "gizmosql_password") \
    .config("gizmosql.tls_skip_verify", True) \
    .getOrCreate()

Using PySpark Imports (activate mode)

You can use the activate() function to enable standard PySpark imports while running on GizmoSQL:

from sqlframe_gizmosql import activate

# Activate GizmoSQL as the backend
activate(
    uri="gizmosql://localhost:31337",
    username="gizmosql_user",
    password="gizmosql_password",
    tls_skip_verify=True  # For self-signed certificates
)

# Now use standard PySpark imports!
from pyspark.sql import SparkSession
from pyspark.sql import functions as F

spark = SparkSession.builder.getOrCreate()

# Create DataFrame and use PySpark-like functions
df = spark.createDataFrame([
    (1, "alice", 100),
    (2, "bob", 200),
    (3, "alice", 150),
], ["id", "name", "amount"])

# Use functions like F.upper, F.sum, F.col, etc.
result = df.select(
    F.col("id"),
    F.upper(F.col("name")).alias("name_upper"),
    F.col("amount")
)
result.show()

# Aggregations
df.groupBy("name").agg(
    F.sum("amount").alias("total"),
    F.count("*").alias("count")
).show()

You can also activate with an existing connection:

from sqlframe_gizmosql import activate, GizmoSQLSession

# Create session first
session = GizmoSQLSession.builder \
    .config("gizmosql.uri", "gizmosql://localhost:31337") \
    .config("gizmosql.username", "gizmosql_user") \
    .config("gizmosql.password", "gizmosql_password") \
    .config("gizmosql.tls_skip_verify", True) \
    .getOrCreate()

# Activate with existing connection
activate(conn=session._conn)

# Use PySpark imports
from pyspark.sql import SparkSession
spark = SparkSession.builder.getOrCreate()

Configuration Options

Connection URIs

The preferred URI scheme is gizmosql://host:port, which is secure by default (gRPC with TLS). Append ?transport=tcp for a plaintext connection. The legacy grpc+tls://, grpc+tcp://, and grpc:// schemes remain fully supported.

You can also connect via an ADBC connection profile — a TOML file that keeps connection details (with {{ env_var(NAME) }} substitution for credentials) out of your code:

session = GizmoSQLSession.builder \
    .config("gizmosql.uri", "profile://my-gizmosql-server") \
    .getOrCreate()

OAuth/SSO Authentication

GizmoSQL supports browser-based OAuth/SSO via auth_type="external". When using external auth, no username or password is needed — a browser window will open for authentication:

from sqlframe_gizmosql import GizmoSQLSession

session = GizmoSQLSession.builder \
    .config("gizmosql.uri", "gizmosql://gizmosql.example.com:31337") \
    .config("gizmosql.auth_type", "external") \
    .config("gizmosql.tls_skip_verify", True) \
    .getOrCreate()

Or with activate mode:

from sqlframe_gizmosql import activate

activate(
    uri="gizmosql://gizmosql.example.com:31337",
    auth_type="external",
    tls_skip_verify=True
)

from pyspark.sql import SparkSession
spark = SparkSession.builder.getOrCreate()

Features

• Full PySpark DataFrame API compatibility via SQLFrame
• Arrow Flight SQL protocol for high-performance data transfer
• Bulk ingestion of Arrow/pandas data via session.ingest() (ADBC adbc_ingest())
• Support for reading/writing various file formats (Parquet, CSV, JSON)
• Window functions
• Aggregations and groupBy operations
• Joins
• UDF registration
• Catalog operations

Observability

The underlying adbc-driver-gizmosql driver (v1.3.0+) emits OpenTelemetry trace spans for Database.Open, Prepare, ExecuteQuery, and ExecuteUpdate. Enable them via the standard OTEL_* environment variables (e.g. OTEL_TRACES_EXPORTER=otlp), or per-connection via the driver's adbc.telemetry.* options. Structured driver logging is available via ADBC_DRIVER_FLIGHTSQL_LOG_LEVEL (debug/info/warn/error). See the adbc-driver-gizmosql README for details.

Running GizmoSQL with Docker

You can run GizmoSQL locally using Docker:

docker run -d \
    --name gizmosql \
    -p 31337:31337 \
    -e GIZMOSQL_USERNAME=gizmosql_user \
    -e GIZMOSQL_PASSWORD=gizmosql_password \
    -e DATABASE_FILENAME=/tmp/test.duckdb \
    -e TLS_ENABLED=1 \
    gizmodata/gizmosql:latest

The gizmosql:// URI scheme uses TLS by default; set gizmosql.tls_skip_verify to True for self-signed certificates.

Development

Setup

# Clone the repository
git clone https://github.com/gizmodata/sqlframe-gizmosql.git
cd sqlframe-gizmosql

# Create a virtual environment
python -m venv .venv
source .venv/bin/activate

# Install dev dependencies
pip install -e ".[dev]"

Running Tests

# Run unit tests
pytest tests/unit

# Run integration tests (requires GizmoSQL server)
pytest tests/integration

Code Quality

# Run linting
ruff check .

# Run formatting
ruff format .

License

Apache License 2.0

Related Projects

• SQLFrame - PySpark-like DataFrame API for multiple SQL backends
• GizmoSQL - Database server using DuckDB with Arrow Flight SQL interface
• sqlmesh-gizmosql - GizmoSQL adapter for SQLMesh