nzpy_extended: High-performance IBM Netezza driver for Python

nzpy_extended is a hard fork of IBM nzpy — enriched with new features, major performance improvements via a native extension, and expanded platform support.

Key differences from upstream nzpy

Windows 11

Measured with the previous C decoder (0.7.x); not re-measured with the Rust decoder.

WSL2 (Linux x86_64)

Measured with nzpy_extended 0.8.0 (Rust decoder) on 2026-10-31 using tools/examples/performance_test.py: 100 000 rows, median of three runs, after the text fast path and the bad-value batch fix. The nzpy_extended columns average the async and sync medians. Run-to-run spread is about 5% for most cells and up to about 15% for a few.

On WSL2 the native extension is about 3.8x faster than pure Python for integer columns, 2.2x for numeric, 1.7x for string, 5.6x for datetime and 3.7x for boolean (ratios of the averaged columns). String columns gain less because the server spends most of the time before the first row arrives: the benchmark query computes its values on the server, so the decoder is about 7% of the end-to-end time for that query (about 15% before the text fast path). A live interleaved check (tools/examples/benchmark_native_vs_python.py, 153 600 mixed rows, five rounds) measured pure Python at about 4.5x the native time.

Note: WSL2 results were obtained on the same Netezza server as Windows 11 (192.168.0.144). ODBC results use a ctypes-based ANSI-ODBC fallback (because pyodbc's Unicode API is incompatible with the Netezza ODBC driver on Linux). This wrapper fetches every cell individually via SQLGetData, so ODBC numbers on WSL2 are understated vs. native pyodbc on Windows (which uses SQLBindCol batch fetching). The real ODBC-vs-native gap on Linux is likely smaller.

macOS ARM64 (Apple M4)

Measured with nzpy_extended 0.8.0 (Rust decoder) on 2026-10-31 using tools/examples/performance_test.py, on a Mac mini (Apple M4, 16 GB RAM) against the same Netezza server at 192.168.0.144. 100 000 rows per query, one run. The nzpy_extended columns average the async and sync results. Full numbers, the native-vs-Python comparison and the bulk-load timings are in docs/benchmarks.md.

If the compiled extension is not available (unsupported platform or Python version), the driver gracefully falls back to pure Python with identical semantics.

Installation

The package is available on PyPI:

pip install nzpy-extended

Pre-built wheels use the Python stable ABI (cp311-abi3), so one wheel per platform installs on Python 3.11 and newer:

Free-threaded CPython builds (3.13t/3.14t) are not covered by these wheels.

For other platforms, pip install builds the native decoder from source with maturin. This requires a stable Rust toolchain (install it with rustup). Without one the install fails — use a supported platform or install Rust first.

Quick Start

Sync — scripts, ETL, Jupyter, Django

import nzpy_extended.sync as nzpy

conn = nzpy.connect(
    user="admin", password="password",
    host="netezza-host", database="mydb",
)

# --- Basic query ---
with conn.cursor() as cur:
    cur.execute("SELECT 1")
    print(cur.fetchone())  # [1]

# --- One-shot convenience (pyodbc pattern) ---
row = conn.execute("SELECT version()").fetchone()
print(row[0])

# --- Parameter binding ---
cur.execute("SELECT id, name FROM users WHERE active = ?", (1,))
for row in cur:
    print(row)

# --- Cursor iteration (batched, arraysize=100) ---
cur.execute("SELECT * FROM large_table")
for row in cur:            # fetches 100 rows per round-trip
    process(row)

# --- Method chaining (PEP 249) ---
rows = cur.execute("SELECT 1, 2, 3").fetchall()

# --- Explicit transaction control ---
conn.autocommit = False
cur.execute("INSERT INTO users (name) VALUES ('Alice')")
cur.execute("INSERT INTO users (name) VALUES ('Bob')")
conn.commit()               # or conn.rollback()
conn.autocommit = True

# --- Transaction context manager ---
with conn.transaction():
    cur.execute("INSERT INTO users (name) VALUES ('Charlie')")
    # auto-commits on success, auto-rollbacks on exception

# --- Timeout (pyodbc-compatible properties) ---
conn.timeout = 5.0          # default for all new cursors
cur = conn.cursor()         # inherits timeout=5.0
cur.execute("SELECT ...")   # raises OperationalError after 5s

cur.timeout = 10.0          # per-cursor override
cur.execute("SELECT ...", timeout=3.0)  # per-execute override wins

# --- Stored procedures ---
result = cur.callproc("sp_add_numbers", [10, 20])
rows = cur.fetchall()

# --- Server notices (RAISE NOTICE) ---
for msg in cur.messages:
    print("Notice:", msg)

# --- PEP 249 rownumber ---
cur.execute("SELECT * FROM t")
print(cur.rownumber)  # 0
cur.fetchone()
print(cur.rownumber)  # 1

# --- Column metadata ---
desc = cur.description       # 7-tuple after execute()
schema = cur.get_schema_table()  # rich metadata as list[dict]

# --- Cancel / interrupt ---
conn.cancel()                # stops running query, session survives
cur.interrupt()              # alias

# --- Connection state ---
print(conn.closed)           # False
conn.close()
print(conn.closed)           # True

Async — FastAPI, asyncio

import asyncio
import nzpy_extended as nzpy

async def main():
    async with await nzpy.connect(
        user="admin", password="password",
        host="netezza-host", database="mydb",
    ) as conn:
        # --- Basic query ---
        async with conn.cursor() as cur:
            await cur.execute("SELECT id, name FROM users WHERE active = ?", (1,))
            return await cur.fetchall()

        # --- One-shot convenience ---
        row = await (await conn.execute("SELECT version()")).fetchone()

        # --- Async iteration (streaming) ---
        cur = conn.cursor()
        await cur.execute("SELECT * FROM large_table")
        async for row in cur:
            process(row)

        # --- Timeout ---
        await cur.execute("SELECT pg_sleep(999)", timeout=3.0)
        # raises OperationalError after 3s

        # --- Stored procedures ---
        result = await cur.callproc("sp_add_numbers", [10, 20])

        # --- Notices ---
        for msg in cur.messages:
            print("Notice:", msg)

asyncio.run(main())

FastAPI with connection pool

from fastapi import FastAPI, Depends
import nzpy_extended as nzpy
import nzpy_extended.fastapi as nzpy_fastapi

pool = nzpy.NzPool(
    min_size=2, max_size=10,
    host="netezza-host", database="mydb",
    user="admin", password="password",
)

app = FastAPI(lifespan=nzpy_fastapi.lifespan(pool))

@app.get("/users")
async def get_users(conn=Depends(nzpy_fastapi.get_connection)):
    async with conn.cursor() as cur:
        await cur.execute("SELECT * FROM users LIMIT 100")
        return await cur.fetchall()

Sync connection pool

import nzpy_extended.sync as nzpy
from nzpy_extended import SyncPool

pool = SyncPool(
    min_size=2, max_size=10,
    host="netezza-host", database="mydb",
    user="admin", password="password",
)
with pool.connection() as conn:
    conn.execute("SELECT 1").fetchone()
pool.close_all()

Bulk data loading via external table protocol

load_data() inserts rows from a Python iterable or async iterable into a Netezza table using the native external table protocol (REMOTESOURCE 'python'). It supports optional automatic table creation. Lists and tuples are materialized as CSV bytes; generators, async generators, and other iterables are streamed to Netezza in chunks.

# --- Sync ---
import nzpy_extended.sync as nzpy
conn = nzpy.connect(...)
count = conn.load_data("my_table", [(1, "Alice"), (2, "Bob")])

# --- Async ---
import nzpy_extended as nzpy
conn = await nzpy.connect(...)
count = await nzpy.load_data(conn, "my_table", [(1, "Alice"), (2, "Bob")])

# --- Auto-infer types from data ---
rows = [(1, "Alice", 100.50), (2, "Bob", 200.75)]
count = conn.load_data("my_table", rows)
# Creates: col1 SMALLINT, col2 VARCHAR(255), col3 NUMERIC

# --- Explicit columns ---
count = conn.load_data(
    table_name="products",
    rows=[(101, "Widget", 9.99)],
    columns=[("id", "INT"), ("name", "VARCHAR(200)"), ("price", "NUMERIC(10,2)")],
)

# --- Generator for large datasets ---
def generate_rows(n):
    for i in range(n):
        yield (i, f"item_{i}")
count = conn.load_data("my_table", rows=generate_rows(50000))

# --- Async generator for streaming ---
async def generate_rows_async(n):
    for i in range(n):
        yield (i, f"item_{i}")
count = await nzpy.load_data(conn, "my_table", rows=generate_rows_async(50000))

Parameters:

Streaming CSV import via load_csv()

load_csv() streams a local CSV file directly into a Netezza table without loading the whole file into memory. It inspects a sample of rows to infer column types when create_if_missing=True, then uses the external table protocol to load the remaining data.

# --- Sync ---
import nzpy_extended.sync as nzpy
conn = nzpy.connect(...)
count = conn.load_csv("my_table", "C:/data/input.csv",
                      delimiter=',', has_header=True, encoding='UTF8')

# --- Async ---
import nzpy_extended as nzpy
conn = await nzpy.connect(...)
count = await nzpy.load_csv(conn, "my_table", "data/input.csv",
                            delimiter=',', has_header=True, encoding='UTF8')

Metadata API (catalog introspection)

The conn.meta object provides async access to Netezza system catalog views — tables, columns, views, procedures, distribution keys, storage stats, sessions, and more. All queries run against the current database; connect to SYSTEM for system-wide objects.

import nzpy_extended as nzpy

async def main():
    conn = await nzpy.connect(user="admin", password="password",
                              host="netezza-host", database="mydb")

    # --- Schemas & databases ---
    schemas = await conn.meta.get_schemas()       # ["ADMIN", "INFORMATION_SCHEMA", ...]
    dbs     = await conn.meta.get_databases()     # ["JUST_DATA", "SYSTEM", ...]
    db_name = await conn.meta.get_current_database()   # "JUST_DATA"

    # --- Tables ---
    tables = await conn.meta.get_tables(schema="ADMIN")
    # [{"schema": "ADMIN", "table_name": "DIMDATE", "owner": "ADMIN",
    #   "objtype": "TABLE", "objid": 123456, "row_count": 500000}, ...]

    tables = await conn.meta.get_tables(
        schema="ADMIN", table_pattern="DIM%", include_system=False,
    )

    # --- Views (includes view definition SQL) ---
    views = await conn.meta.get_views(schema="ADMIN")
    # [{"schema": "ADMIN", "view_name": "V_SALES", "owner": "ADMIN",
    #   "objid": 789012, "definition": "CREATE VIEW V_SALES AS SELECT ..."}, ...]

    # --- Columns ---
    cols = await conn.meta.get_columns("DIMDATE", schema="ADMIN")
    # [{"column_name": "DATEKEY", "ordinal": 1,
    #   "data_type": "DATE", "nullable": "N"}, ...]

    # Dot-notation also works:
    cols = await conn.meta.get_columns("ADMIN.DIMDATE")

    # --- Distribution key ---
    dk = await conn.meta.get_distribution_key("FACT_SALES", schema="ADMIN")
    # ["CUSTOMER_ID"]  or [] for RANDOM distribution

    # --- Table sizes ---
    sizes = await conn.meta.get_table_sizes(schema="ADMIN")
    # [{"schema": "ADMIN", "table_name": "FACT_SALES",
    #   "used_bytes": 500000000, "allocated_bytes": 600000000,
    #   "size_mb": 476, "skew": 1.2}, ...]

    # --- Stored procedures ---
    procs = await conn.meta.get_procedures(schema="ADMIN")
    # [{"schema": "ADMIN", "proc_name": "SP_LOAD_DATA",
    #   "owner": "ADMIN", "signature": "SP_LOAD_DATA(VARCHAR(256))",
    #   "returns": "INTEGER", "source": "CREATE PROCEDURE ..."}, ...]

    # --- Sequences & synonyms ---
    seqs = await conn.meta.get_sequences(schema="ADMIN")
    syns = await conn.meta.get_synonyms(schema="ADMIN")

    # --- Sessions ---
    sessions = await conn.meta.get_sessions()
    # [{"session_id": 123, "username": "ADMIN", "database_name": "JUST_DATA",
    #   "conntime": datetime(...), "priority": 0, "status": "active"}, ...]

    # --- Users & groups ---
    users  = await conn.meta.get_users()
    groups = await conn.meta.get_groups()

    # --- Query history (requires history collection enabled) ---
    history = await conn.meta.get_query_history(limit=50, username="ADMIN")

    # --- Search across tables, views, procedures ---
    results = await conn.meta.search_objects("SALES%", schema="ADMIN")
    # [{"object_type": "TABLE", "schema": "ADMIN",
    #   "object_name": "FACT_SALES", "owner": "ADMIN", "objid": 123}, ...]

asyncio.run(main())

API Reference

Documentation

• docs/ folder in the repository:
  • timeout_and_cancel.md — Timeout and cancel mechanisms
  • sync_api.md — Full sync API reference
  • pep249_compliance.md — PEP 249 compliance table
  • pool.md — Connection pool documentation
  • load_data.md — Bulk data loading
  • metadata_api.md — Catalog introspection API
• GitHub Wiki
• Issue tracker

Requirements

• Python ≥ 3.11 (tested on 3.11 through 3.15)
• CPython (PyPy not supported for native extension, pure-Python fallback only)

Testing

Running the test suite

CI (GitHub Actions)

CI now discovers all unit tests, measures branch coverage and runs native and pure-Python variants on Python 3.11–3.15 across Linux, macOS and Windows. Publishing depends on this quality workflow. Published wheels are built once per platform with the stable ABI. Wheel/sdist import checks run outside the source tree. Use NZ_DEV_DATABASE (preferred) or NZ_DEV_DB for live tests; the preferred name wins when both are set. See CHANGELOG.md for unreleased compatibility fixes.

Pull requests run without a live Netezza database:

• Unit tests (paramstyle, regressions, native/Python parity, buffer pool/stream, pool, csv_import)
• mypy and pyright
• Wheel and sdist import smoke

Integration tests (smoke, full, and differential parity) must be run locally against your Netezza instance.

Local integration

Tests require a running Netezza instance. Set the connection environment variables:

export NZ_DEV_HOST=your_netezza_host
export NZ_DEV_PORT=5480
export NZ_DEV_DB=JUST_DATA
export NZ_DEV_USER=admin
export NZ_DEV_PASSWORD=password

Differential parity tests use the independent JustyBase Netezza driver, not ODBC. Build the Node driver from JustyBase.NetezzaDriver or its Node companion, then point the tests at its compiled package:

export NZ_REFERENCE_NODE_DRIVER=/path/to/justybase_netezza_node_driver
pytest tests/test_odbc_comparison_smoke.py -v
pytest tests/test_odbc_comparison_node.py -m reference_node -v

The bridge in tests/reference_driver.js runs the reference driver in a separate Node process and exchanges JSON-lines, so the two implementations do not share protocol or conversion code. The default local checkout path is ../justybase_netezza_node_driver.

Run all tests locally:

pytest tests/ -v

CI-equivalent (no database):

pytest tests/test_paramstyle.py tests/test_typeobjects.py tests/test_regressions_unit.py \
  tests/test_native_parity_unit.py tests/test_buffer_pool.py tests/test_buffered_stream.py \
  tests/test_csv_import_unit.py tests/test_pool_unit.py -v

See CONTRIBUTING.md for marker profiles (smoke, full, unit, benchmark).

Upgrading to 0.4.0

If you connect with securityLevel=2 and relied on automatic fallback to an unencrypted session when SSL fails, you must now opt in explicitly:

await nzpy.connect(
    ...,
    securityLevel=2,
    ssl={"ssl_allow_fallback": True},  # only if you need legacy behaviour
)

See CHANGELOG.md for the full list of changes.

Native decoder / pure Python parity

The native DBOS decoder (nzpy_extended.rust_ext, source in rust/) and the pure-Python decoder must produce identical results for all data types.

Unit tests (tests/test_native_parity_unit.py) — compare the native decoders against the Python reference on generated rows: appliance-shaped, mutated, and out-of-spec widths. No database required.

pytest tests/test_native_parity_unit.py -v

Integration tests (tests/test_native_parity_integration.py) — run real SQL queries through both code paths and verify results match. Requires a database.

pytest tests/test_native_parity_integration.py -v

Verification script — runs the decoder tests against the selected native module and in pure-Python mode:

python tools/verify_native_parity.py

Disabling the native decoder at runtime:

Set the environment variable NZPY_EXTENDED_NO_NATIVE=1 to force pure-Python mode even when the native module is installed. Useful for debugging or verifying fallback correctness.

NZPY_EXTENDED_NO_NATIVE=1 python -c "from nzpy_extended import _native; print(_native._HAVE_NATIVE)"  # False

If the native module cannot be imported, the driver emits a RuntimeWarning and falls back to pure Python, which is much slower. Setting NZPY_EXTENDED_NO_NATIVE=1 chooses pure Python deliberately and silences that warning. The older name NZPY_EXTENDED_NO_CEXT was removed in 0.9.

Reproducing benchmark results

The per-type benchmark table above is generated by tools/examples/performance_test.py. The raw results from both machines, the native-vs-Python comparison and the load_data timings (via tools/examples/benchmark_load_data.py) are in docs/benchmarks.md.

To compare the native and pure-Python decoders on your own appliance, run the live interleaved benchmark. It creates a uniquely named table, drops it afterwards, and should only be run against a disposable database:

python tools/examples/benchmark_native_vs_python.py --doublings 9 --rounds 5 --output native_vs_python.json

Prerequisites

• A running Netezza instance with a table named JUST_DATA..FACTPRODUCTINVENTORY (or adjust SOURCE_TABLE in the script)
• Python ≥ 3.11
• Install the required packages:

pip install nzpy_extended

The official IBM driver is also tested for comparison (pip install nzpy).

Optional (for ODBC comparison):

• pip install pyodbc — with NetezzaSQL ODBC driver installed (rows labeled pyodbc)
• Set NZ_ODBC_DRIVER if your ODBC driver uses a different name (default: NetezzaSQL). On Linux/WSL2 the name is defined in /etc/odbcinst.ini, e.g.:

  export NZ_ODBC_DRIVER=NetezzaSQL  # Linux/WSL2
  

• Linux/WSL2: pyodbc's Unicode API is incompatible with the Netezza ODBC driver. The benchmark auto-detects this and falls back to a ctypes-based ANSI-ODBC wrapper. No additional configuration needed.

Steps

1. Set connection environment variables:

set NZ_HOST=your_netezza_host     # Windows
set NZ_PORT=5480
set NZ_USER=admin
set NZ_PASSWORD=password
set NZ_DATABASE=JUST_DATA

# or on Linux/macOS:
export NZ_HOST=your_netezza_host
export NZ_PORT=5480
export NZ_USER=admin
export NZ_PASSWORD=password
export NZ_DATABASE=JUST_DATA

The script has no host, user or password defaults. Set NZ_DEV_HOST, NZ_DEV_PORT, NZ_DEV_USER, NZ_DEV_PASSWORD and NZ_DEV_DATABASE (the legacy NZ_* names are accepted as fallbacks). It exits with the list of missing settings otherwise.

ODBC driver name can be set separately:

export NZ_ODBC_DRIVER=NetezzaSQL  # Linux/macOS — defaults to "NetezzaSQL"

2. Run the benchmark:

python tools/examples/performance_test.py

3. Adjust row count (default: 100 000):

set NZ_ROWS=100000     # Windows
export NZ_ROWS=100000  # Linux/macOS

What the script does

1. Connects using each driver: official_nzpy (always), pyodbc (via ANSI-ODBC ctypes fallback on Linux), nzpy_extended (async + sync, with and without native extension).
2. Runs six query categories: integer_types, numeric_types, string_types, datetime_types, boolean_types, and all_types (mixed).
3. Prints per-query timing, a compact DRIVER × TYPE comparison table (matching the README layout), and visual bar charts.

Saving results to a TXT file

Use the --output / -o flag or the NZ_OUTPUT environment variable:

python tools/examples/performance_test.py -o benchmark_results.txt

# or via env var:
set NZ_OUTPUT=benchmark_results.txt
python tools/examples/performance_test.py

Force pure-Python mode

set NZPY_EXTENDED_NO_NATIVE=1
python tools/examples/performance_test.py

Pytest benchmark (alternative)

A simpler pytest-based benchmark is also available (10 k rows, nzpy_extended only):

pytest tests/test_benchmark.py -v -m benchmark

License

Apache License 2.0 — see LICENSE.