Python Client (`provisa-client`)¶

Python client for Provisa. Provides four interfaces:

Interface	Use case
`ProvisaClient`	GraphQL queries, Arrow Flight, DataFrame output
DB-API 2.0 (`connect`)	Standard Python database interface (PEP 249)
SQLAlchemy dialect	BI tools, ORM, Pandas `read_sql`
ADBC	Arrow-native columnar streaming via Flight

Install¶

pip install provisa-client                      # core (ProvisaClient + DB-API)
pip install "provisa-client[pandas]"            # adds pandas
pip install "provisa-client[sqlalchemy]"        # adds SQLAlchemy dialect
pip install "provisa-client[adbc]"              # adds ADBC over Arrow Flight

ProvisaClient¶

Quick Start¶

from provisa_client import ProvisaClient

client = ProvisaClient(
    "http://localhost:8001",
    token="provisa_pat_...",   # personal access token, or a provider bearer token
    role="analyst",
)

ProvisaClient takes a credential. It has no login step of its own, so it never asks for a username and password. When a script needs to run unattended, use a personal access token: you issue it from your own profile, it carries an expiry, and you can revoke it without touching the account. A provider bearer token works the same way. Either one goes in token, and the client presents it on both the HTTP and the Arrow Flight path.

To exchange a password for a token, POST to /auth/login and read access_token:

import httpx

body = httpx.post(
    "http://localhost:8001/auth/login",
    json={"username": "alice", "password": "secret"},
).json()
client = ProvisaClient("http://localhost:8001", token=body["access_token"])

The DB-API and ADBC entry points do this exchange for you, as described below.

GraphQL Queries¶

# Raw response dict
result = client.query("{ orders { id amount region } }")

# With variables
result = client.query(
    "query Q($region: String!) { orders(region: $region) { id amount } }",
    variables={"region": "west"},
)

# pandas DataFrame (first root field is flattened)
df = client.query_df("{ orders { id amount region } }")

Async¶

result = await client.aquery("{ orders { id amount } }")

Arrow Flight (high-throughput columnar)¶

Use Flight for large result sets. Data streams as Arrow record batches and is never materialized on the server.

import pyarrow as pa

table: pa.Table = client.flight("{ orders { id amount region } }")
df = client.flight_df("{ orders { id amount region } }")

Flight connects to port 8815 by default. Override with flight_port=:

client = ProvisaClient("http://prod.example.com", flight_port=8815)

Catalog Exploration¶

tables_df = client.list_tables()

Connection Reference¶

Parameter	Default	Description
`url`	`http://localhost:8001`	Provisa server base URL
`token`	`None`	Bearer credential — a provider token or a personal access token; omit for password auth
`role`	`None`	Role to act as, sent as `X-Provisa-Role` and honored only when assigned to the user; omit to run as the identity's own role
`flight_port`	`8815`	Arrow Flight gRPC port

Error Handling¶

query() raises httpx.HTTPStatusError on HTTP errors. query_df() raises RuntimeError if the response contains GraphQL errors.

Using Provisa in Jupyter¶

Install into the running kernel before importing. The pandas extra covers DataFrame output; add sqlalchemy if you also need pandas.read_sql through the dialect.

%pip install "provisa-client[pandas]"

Credentials¶

Notebooks persist cell output to the .ipynb file. Type a token into a cell and it will be saved to disk alongside your code. Don't do that.

ProvisaClient reads no environment: pass its credentials in explicitly, from environment variables. (provisa_client.query(), below, reads them for you.)

import os
from provisa_client import ProvisaClient

client = ProvisaClient(
    os.environ["PROVISA_URL"],           # choose your own variable name
    token=os.environ["PROVISA_TOKEN"],   # personal access token
    role=os.environ.get("PROVISA_ROLE"), # optional; omit to run as your own role
    org=os.environ.get("PROVISA_ORG"),   # required on a multi-tenant deployment; None otherwise
)

Set those variables in your shell before starting Jupyter, or load a .env file with python-dotenv. Either way, keep them out of the notebook.

One call: `query()`¶

provisa_client.query() takes a SQL, GraphQL or Cypher statement and returns a pandas DataFrame. The server works out which language it is. Nothing is constructed first: the connection comes from the environment or a profile file.

import provisa_client

df = provisa_client.query("SELECT region, amount FROM shop.orders")
admin_view = provisa_client.query("SELECT region, amount FROM shop.orders", role="org_admin")

role= acts as that role for one read, sent as X-Provisa-Role. The server honours it only for a role you hold. The read goes over Arrow Flight; set transport to http to send it to /data/query instead. await provisa_client.aquery(...) is the awaitable form.

The connection is read from the first place that names each setting:

keywords to provisa_client.sign_in(...), which resolves the connection once for every later read;
environment variables: PROVISA_URL, PROVISA_TOKEN, PROVISA_USERNAME, PROVISA_PASSWORD, PROVISA_ROLE, PROVISA_ORG, PROVISA_FLIGHT_PORT, PROVISA_TRANSPORT;
a section of ~/.provisa/credentials: [default], or the one PROVISA_PROFILE (or sign_in(profile=...)) names.

[default]
url = https://provisa.example.com
token = provisa_pat_...
role = analyst

[staging]
url = https://staging.provisa.example.com
username = ana

A token is a personal access token or a session token. Give a username and password instead, and the client exchanges them for a session token at POST /auth/login. That works on a deployment whose sign-in provider takes a password (basic, simple, LDAP). Keep the password in PROVISA_PASSWORD, out of the file. A refused sign-in, a missing URL or a named profile that isn't in the file raises an error with the reason. Provisa has no sign-in flow a notebook can drive through a browser, so for single sign-on deployments use a personal access token.

Cell magics¶

%pip install "provisa-client[notebook]"
%load_ext provisa_client.magics

%%sql eu --role analyst_eu
SELECT region, customer_email FROM shop.orders

%%sql, %%graphql and %%cypher answer the cell with query(). The DataFrame is the cell's result. A name after the magic also binds it in the notebook (eu above); --role acts as that role for the cell.

JupySQL (%load_ext sql) defines a %%sql as well. Load both and %%sql belongs to whichever extension loaded last. Provisa's takes --role and reads over Arrow Flight. JupySQL works too, through the SQLAlchemy dialect: %sql provisa+http://user:password@host:port connects, %sql/%%sql run governed statements, and %sqlcmd tables -s <domain> and %sqlcmd columns -t <table> -s <domain> list the catalog.

In the browser (Pyodide)¶

Under Pyodide, query() reads from the server that served the page, as the page's signed-in user. The host page keeps that user on the kernel's global scope:

self.provisaAuth = { baseUrl: "", role: "analyst_eu", bearer: "<token>", org: null };

baseUrl is empty for the page's own origin. The client reads provisaAuth on every request, so a rotated token takes effect without a restart. Results arrive as Arrow when pyarrow is loaded; without it they come as JSON. A Cypher statement always comes back as JSON. query() blocks, which a Web Worker (where a notebook kernel runs) can do; on the page's main thread it raises and await provisa_client.aquery(...) is the call to use.

DataFrames from GraphQL, SQL, and Cypher¶

query_df() accepts all three query languages and returns a pandas DataFrame.

# GraphQL — first root field is extracted automatically
df = client.query_df("{ orders { id amount region } }")

# SQL
df = client.query_df("SELECT id, amount, region FROM orders WHERE region = 'west'")

# Cypher
df = client.query_df("MATCH (o:Order) RETURN o.id, o.amount, o.region")

GraphQL responses carry the data under a root field name; query_df() extracts it. SQL and Cypher responses carry explicit column names from the server and are mapped directly.

Arrow Flight for large pulls¶

flight_df() streams data as Arrow record batches instead of deserializing JSON. For wide tables or result sets over a few hundred thousand rows, it is faster and uses less memory.

df = client.flight_df("SELECT * FROM events")

Arrow preserves column types. Integers stay integers and timestamps stay timestamps, which JSON can't promise. When the schema matters, that alone is reason to prefer Flight.

Flight connects on port 8815 by default. Pass flight_port= to the ProvisaClient constructor if your server uses a different port.

`pandas.read_sql` through the SQLAlchemy dialect¶

Install the sqlalchemy extra into the kernel:

%pip install "provisa-client[sqlalchemy]"

Then build an engine with a provisa+http:// URL. Read the credentials from environment variables so they never sit in your notebook.

import os
import pandas as pd
from sqlalchemy import create_engine

host = os.environ["PROVISA_HOST"]       # e.g. "localhost:8001"
user = os.environ["PROVISA_USER"]
pwd  = os.environ["PROVISA_PASSWORD"]

engine = create_engine(f"provisa+http://{user}:{pwd}@{host}")

df = pd.read_sql("SELECT id, amount FROM orders", engine)

The dialect posts the username and password to /auth/login once and holds the resulting token for the life of the engine. The credentials aren't sent again with each query.

DB-API 2.0¶

Standard PEP 249 interface. Works with any tool that accepts a DB-API connection.

from provisa_client import connect

conn = connect(
    "http://localhost:8001",
    username="alice",
    password="secret",
    role="analyst",     # optional; omit to run as the role the login returns
)

connect posts the username and password to /auth/login and holds the access_token it gets back, so the connection carries a real credential. role requests a role, and the server honours it only if the identity is assigned that role. When you omit it, the connection runs as the role the login resolved.

Executing queries¶

The cursor accepts GraphQL or SQL and detects which one it got.

cur = conn.cursor()

# GraphQL
cur.execute("{ orders { id amount region } }")
rows = cur.fetchall()           # list of tuples
one  = cur.fetchone()           # single tuple or None
many = cur.fetchmany(size=50)   # up to N tuples

# SQL (routed through Stage 2 governance)
cur.execute("SELECT id, amount FROM orders WHERE region = 'west'")
rows = cur.fetchall()

Column metadata¶

cur.execute("{ orders { id amount } }")
print(cur.description)
# [('id', None, ...), ('amount', None, ...)]
print(cur.rowcount)

Named parameters¶

cur.execute(
    "SELECT * FROM orders WHERE region = :region",
    {"region": "west"},
)

Context managers¶

with connect("http://localhost:8001", username="alice", password="secret") as conn:
    with conn.cursor() as cur:
        cur.execute("{ orders { id amount } }")
        print(cur.fetchall())

SQLAlchemy Dialect¶

pip install "provisa-client[sqlalchemy]"

URL scheme: provisa+http:// or provisa+https://

from sqlalchemy import create_engine, text

engine = create_engine("provisa+http://alice:secret@localhost:8001")

with engine.connect() as conn:
    result = conn.execute(text("{ orders { id amount region } }"))
    for row in result:
        print(row)

With pandas¶

import pandas as pd

df = pd.read_sql("{ orders { id amount } }", engine)

URL parameters¶

Parameter	Description	Default
`role`	Role to request; validated server-side	the role the login resolves

engine = create_engine(
    "provisa+http://alice:secret@localhost:8001?role=analyst"
)

Schema introspection¶

The dialect reflects the catalog your role is served, read from GET /data/catalog as the connection's user, role and org. Each domain is a schema, and every table lives in one: get_schema_names() lists the domains, get_table_names(schema="shop") lists that domain's tables by their SQL names, and get_columns("orders", schema="shop") gives each column's type, nullability and description. With no schema named, get_table_names() lists none. An unknown table raises NoSuchTableError; a catalog the server refuses raises.

ADBC¶

Arrow Database Connectivity, backed by Arrow Flight. It returns a pyarrow.Table directly, with no JSON deserialization.

pip install "provisa-client[adbc]"

from provisa_client.adbc import adbc_connect

conn = adbc_connect(
    "http://localhost:8001",
    user="alice",
    password="secret",
    role="analyst",   # optional; server validates the requested role
    port=8815,        # Arrow Flight port
)

adbc_connect logs in over HTTP first and puts the resulting token in every Flight ticket, so the Flight server authenticates the connection the same way the REST surface does. The role argument is a request that the server validates against the identity's assignments. It never becomes the identity.

Fetch as Arrow Table¶

with conn.cursor() as cur:
    cur.execute("{ orders { id amount region } }")
    table = cur.fetch_arrow_table()   # pyarrow.Table
    df = table.to_pandas()

Fetch as tuples¶

with conn.cursor() as cur:
    cur.execute("{ orders { id amount } }")
    rows = cur.fetchall()    # list of tuples
    one  = cur.fetchone()    # single tuple or None

Column metadata¶

cur.execute("{ orders { id amount } }")
print(cur.description)
# [('id', None, ...), ('amount', None, ...)]

Context manager¶

with adbc_connect("http://localhost:8001", user="alice", password="secret") as conn:
    with conn.cursor() as cur:
        cur.execute("{ orders { id amount } }")
        table = cur.fetch_arrow_table()

ADBC connects to the Flight server on port 8815 by default. Pass port= to reach a Flight server bound to a non-default port.

Python Client (provisa-client)¶

Install¶

ProvisaClient¶

Quick Start¶

GraphQL Queries¶

Async¶

Arrow Flight (high-throughput columnar)¶

Catalog Exploration¶

Connection Reference¶

Error Handling¶

Using Provisa in Jupyter¶

Credentials¶

One call: query()¶

Cell magics¶

In the browser (Pyodide)¶

DataFrames from GraphQL, SQL, and Cypher¶

Arrow Flight for large pulls¶

pandas.read_sql through the SQLAlchemy dialect¶

DB-API 2.0¶

Executing queries¶

Column metadata¶

Named parameters¶

Context managers¶

SQLAlchemy Dialect¶

With pandas¶

URL parameters¶

Schema introspection¶

ADBC¶

Fetch as Arrow Table¶

Fetch as tuples¶

Column metadata¶

Context manager¶

Python Client (`provisa-client`)¶

One call: `query()`¶

`pandas.read_sql` through the SQLAlchemy dialect¶