arrows

arrows is an universal data wrangling tool built around Apache Arrow, DuckDB, Polars, S3 for efficiently transferring and transforming data between different data sources.

Every integration is a component: an independently installable, independently loadable unit that owns its own dependencies and its own credentials. A job that only sends email never imports boto3, and never needs a Redshift password to exist.

Features

• 🧩 Pluggable components: load only what a job needs; add new ones without touching the core
• 🔄 Data format conversion: Apache Arrow, Pandas, Polars, DuckDB
• 📊 Google Sheets: read and write, with SQL query support
• 🗄️ Amazon Redshift: query, UNLOAD to S3, COPY from S3
• ☁️ AWS S3: Parquet datasets, queryable in place
• 📧 Gmail: send templated email
• 🔐 Layered secret resolution: env → .env → local files → OS keychain → AWS Secrets Manager
• 📝 SQL template rendering with Jinja2

Installation

pip install arrows        # every component
uv add arrows             # with uv

That is the whole installation. There are no extras to choose between: every component's libraries are ordinary dependencies.

Installing everything does not mean loading everything. import arrows pulls in none of it — boto3, psycopg2, the Google client and Polars are imported the first time a component that needs them is loaded, so start-up stays in the tens of milliseconds whatever a job actually uses.

Components

arrows components          # what exists, what it needs, what is loaded
arrows secrets redshift    # which secrets resolve, and from where (values redacted)
arrows doctor s3 redshift  # load them and run a real health check
arrows login redshift      # type the missing secrets, then verify

Loading

Nothing is loaded on import. Ask for what the job needs:

import arrows

arrows.load('s3', 'redshift')  # authenticates both, fails fast if a secret is missing
arrows.load('gmail')  # pulls in 'google' automatically

Loading is how a job states its requirements up front, so a missing credential surfaces at start-up instead of an hour into a run. In production, turn off implicit loading so that is enforced:

arrows.configure(autoload=False)  # or ARROWS_AUTOLOAD=0

With autoload left on (the default), using a data API loads its component on first call, which is what you want in a notebook.

Other entry points:

arrows.list_components()  # names, dependencies, load state
arrows.health()  # per-component health checks
arrows.unload('redshift')  # close connections, drop registered secrets
arrows.close()  # release everything

ARROWS_COMPONENTS=s3,redshift supplies the default set, so arrows.load() with no arguments loads it — useful for containers and scheduled jobs. Creating the session itself never touches a credential, so arrows.configure(secrets=...) can install a different secret source first.

Isolated sessions

A Session owns its components and its secret store, so one process can talk to two environments at once:

from arrows.core import Session, SecretStore
from arrows.core.secrets import AwsSecretsManagerProvider

prod = Session(secrets=SecretStore([AwsSecretsManagerProvider('prod/arrows')]))
with prod.activate():  # module-level APIs now use this session
    prod.load('redshift')
    arrow = arrows.redshift.fetch_arrow('select 1')

Adding a component

Copy src/arrows/components/sqlite.py — the reference implementation — and declare what you need:

from arrows.core.component import Component, HealthStatus


class ClickhouseComponent(Component):
    name = 'clickhouse'
    requires = ('CLICKHOUSE_URL', 'CLICKHOUSE_PASSWORD')  # resolved before setup runs
    optional = ('CLICKHOUSE_DATABASE',)
    depends_on = ()  # other components

    def setup(self, secrets):
        # A hint is only needed for a component distributed separately from
        # arrows; a built-in leaves it out, since its libraries always ship.
        driver = self.import_module('clickhouse_connect', hint='pip install arrows-clickhouse')
        self.client = driver.get_client(
            host=self.require_secret('CLICKHOUSE_URL'),
            password=self.require_secret('CLICKHOUSE_PASSWORD'),
        )

    def health_check(self):
        return HealthStatus(self.name, bool(self.client.ping()))

    def close(self):
        self.client.close()


COMPONENT = ClickhouseComponent

Register it in src/arrows/components/__init__.py for a built-in, or ship it from your own package with no change to arrows at all:

[project.entry-points."arrows.components"]
clickhouse = "arrows_clickhouse:SPEC"

SPEC = ComponentSpec(
    name='clickhouse',
    module='arrows_clickhouse.component',  # imported only when the component is loaded
    summary='ClickHouse over the native protocol',
    install_hint='pip install arrows-clickhouse',
)

The component's data API is an ordinary module; it reaches its component with arrows.core.session.get('clickhouse').

Secrets

Components never read files or environment variables directly. They ask a SecretStore for a key; where that key comes from is a deployment decision. The default chain, first hit wins:

AwsSecretsManagerProvider is available for production, and is the recommended source once more than one machine is involved:

import arrows
from arrows.core import SecretStore
from arrows.core.secrets import AwsSecretsManagerProvider, EnvProvider

arrows.configure(secrets=SecretStore([EnvProvider(), AwsSecretsManagerProvider('prod/arrows')]))
arrows.load('redshift')

Guarantees worth knowing:

• Values are not copied into os.environ. The environment is inherited by every subprocess and appears in crash dumps. Use store.export([...]) only for a library that insists on reading the environment itself.
• Values cannot leak into a log by accident. Secret.__str__ and __repr__ render ***; the plaintext requires an explicit .reveal().
• Credential files are permission-checked. A file readable by group or others triggers a warning telling you to chmod 600.
• No secret is ever written to disk by arrows.

Notebooks: supplying secrets by hand

Two ways, mixable. In code, with named arguments per component:

import arrows

arrows.redshift.login(host='my-cluster...amazonaws.com', database='dev', user='analyst', password='...')

arrows.s3.login(profile='analytics', bucket='my-staging-bucket')
arrows.google_sheets.login(token_json=open('~/token.json').read())

By prompt, for anything not supplied — masked, and only for what is actually missing:

arrows.login('redshift')
# redshift: enter 2 secret(s), blank to skip
# REDSHIFT_HOST: ········
# REDSHIFT_PASSWORD: ········

arrows.redshift.login(user='analyst')  # asks only for host, database, password

The precedence rule differs between the two on purpose:

Calling login() again with different credentials rebuilds the affected components (and anything depending on them), so rotating a password mid-session takes effect immediately.

Keyword names come from each component's login_args; a raw UPPER_CASE secret key always works too, which is the escape hatch for keys a component has not named:

arrows.login('redshift', REDSHIFT_IAM_DURATION_SECONDS='900')

To stop retyping on every kernel restart, save to the OS keychain — needs arrows[keyring]:

arrows.redshift.login(password='...', save=True)  # found by KeyringProvider next time
arrows.login('redshift', save=True)  # same, for prompted answers

Same thing from a terminal: arrows login redshift --save.

Prompting is inert outside a terminal or an IPython kernel, and ARROWS_NON_INTERACTIVE=1 disables it everywhere — a scheduled job raises a missing-secret error naming the key, instead of hanging forever on a prompt nobody will answer. prompt=False forces that behaviour, prompt=True requires a human. So the same call works in both places:

arrows.redshift.login(host=..., user=..., password=...)  # script: no prompt, nothing missing
arrows.redshift.login()  # notebook: asks for all of it

A key ending in _JSON or _FILE accepts a path at the prompt, which is the sane way to supply a Google token:

arrows.login('google')
# GOOGLE_TOKEN_JSON: ~/Downloads/token.json     ← read from the file

Keys by component

Two credentials deserve a note:

• AWS: prefer no credentials in arrows at all. With nothing configured, boto3 resolves its own chain — AWS_PROFILE, SSO, EC2/ECS/EKS instance roles — all of which rotate automatically. Explicit keys are supported for laptops but are the weakest option, and they are the only case where key material has to be handed to DuckDB as a literal secret.
• Redshift: set REDSHIFT_CLUSTER_IDENTIFIER and omit REDSHIFT_PASSWORD. arrows then mints a short-lived password through redshift:GetClusterCredentials using the AWS identity already loaded, so no database password exists on disk.

Migrating from load_credentials()

arrows.load_credentials() and the arrows.auth.load_* functions still work and now delegate to the component loader, with a DeprecationWarning:

The legacy ~/.credentials/ files are still read, so no file needs to move.

Usage

Amazon Redshift

Querying Data

import arrows
from arrows import redshift

arrows.load('redshift')

# Query using S3 Unload engine (default, suitable for large data)
arrow = redshift.fetch_arrow(sql='SELECT * FROM my_table', engine='s3')

# Query using ADBC engine (fast for smaller datasets)
arrow = redshift.fetch_arrow(sql='SELECT * FROM my_table WHERE date > %(date)s', engine='adbc', date='2024-01-01')

# Get Pandas DataFrame (defaults to ADBC engine)
df = redshift.fetch_dataframe(
    sql='SELECT * FROM my_table',
    engine='adbc',
    dtype_backend='numpy',  # or 'pyarrow'
)

Importing Data to Redshift

from arrows import redshift

# Import Arrow data to Redshift
redshift.arrow_to_redshift(
    arrow=arrow,
    table_name='schema.table_name',
    mode='append',  # or 'overwrite'
)

# Copy from S3 to Redshift
redshift.copy(table_name='schema.table_name', s3_path='s3://bucket/path/', mode='append')

Exporting Data to S3

from arrows import redshift

# Export Redshift query results to S3
dataset = redshift.unload(sql='SELECT * FROM my_table', s3_path='s3://bucket/path/')

Executing SQL

from arrows import redshift

# Execute SQL statement
redshift.execute_sql(sql='CREATE TABLE IF NOT EXISTS my_table (id INT, name VARCHAR)')

# Execute SQL file (supports Jinja2 templates)
redshift.execute_sql_file(sql_script_path='scripts/create_table.sql', table_name='my_table')

AWS S3

Storing and Reading Data

from arrows import s3

# Store Arrow data to S3
dataset = s3.arrow_to_s3(
    arrow=arrow,
    s3_path='s3://bucket/path/',
    bucket='my-bucket',  # Optional
    engine='duckdb',  # or 'pyarrow'
)

# Store Polars DataFrame to S3
dataset = s3.polars_to_s3(df=df, s3_path='s3://bucket/path/')

# Read data from S3
dataset = s3.get_dataset('s3://bucket/path/')
arrow = dataset.to_arrow()

# Convert to Pandas
df = dataset.to_duckdb().df()

# Convert to Polars
df = dataset.to_polars(lazy=False)

S3Dataset Operations

from arrows import s3


# Get dataset
dataset = s3.get_dataset('s3://bucket/path/')

dataset = s3.S3Dataset(s3_path='s3://bucket/path/')


# Create dataset
dataset = s3.create_dataset(s3_path='s3://bucket/path/')

# Write from Arrow
dataset.from_arrow(arrow)

# Write from Polars
dataset.from_polars(df)

# Import from Redshift
dataset.from_redshift(sql='SELECT * FROM my_table')

# Export to Redshift
dataset.to_redshift('schema.table_name', mode='append')

# Query S3 data with SQL
result = dataset.query('SELECT * FROM self WHERE id > 100')

# Delete dataset
dataset.delete()

# Clear dataset contents
dataset.clear_contents()

Google Sheets

Reading from Google Sheets

from arrows import google_sheets

# Read data from Google Sheet as Arrow format
arrow = google_sheets.fetch_arrow(
    spreadsheet_id='your_spreadsheet_id',
    sheet_name='Sheet1',
    sheet_range='A1:D100',  # Optional
    all_varchar=False,  # Optional: Treat all columns as VARCHAR
)

spreadsheet = google_sheets.get_spreadsheet(spreadsheet_id)
sheet = spreadsheet.get_sheet(sheet_name)
# or
sheet = google_sheets.get_sheet(spreadsheet_id, sheet_name)

arrow = sheet.to_arrow(self, sheet_range=None, all_varchar=False, sql=None)
# or
df = sheet.to_polars(self, sheet_range=None, all_varchar=False, sql=None)
# or
df = sheet.to_pandas(self, sheet_range=None, all_varchar=False, sql=None)
# or
duckdb_relation = sheet.to_duckdb(self, sheet_range=None, all_varchar=False, sql=None)


# Use SQL query
arrow = google_sheets.fetch_arrow(
    spreadsheet_id='your_spreadsheet_id',
    sheet_name='Sheet1',
    sql="""
        SELECT
            *
        FROM 
            self
        WHERE column1 > 100
        """,
)

Writing to Google Sheets

from arrows import google_sheets

# Write Arrow data to Google Sheet
sheet = google_sheets.arrow_to_googlesheet(arrow=arrow, spreadsheet_id='your_spreadsheet_id', sheet_name='Sheet1')

Managing Spreadsheets and Sheets

from arrows import google_sheets

# Create a new Spreadsheet
spreadsheet = google_sheets.create_spreadsheet(
    spreadsheet_name='My Spreadsheet',
    parent_id='parent_id',  # Optional
)

# Get a Spreadsheet
spreadsheet = google_sheets.get_spreadsheet('spreadsheet_id')

# Create a new Sheet
sheet = spreadsheet.create_sheet('New Sheet')

# Get a Sheet
sheet = spreadsheet.get_sheet('Sheet1')

# Share Spreadsheet
spreadsheet.share(email='user@example.com', role='writer')

# Delete a Sheet
spreadsheet.delete_sheet('Sheet1')

Gmail

Sending Emails

from arrows import gmail

# Simple email
gmail.send_email(to=['user@example.com'], subject='Report', content='<h1>Hello</h1>', cc=['manager@example.com'])

# Advanced usage with Email class
email = gmail.Email(subject='Monthly Report', to=['user@example.com'], sender='Data Team')

# Set content from template
email.from_template('path/to/template.html', variable='value')

# Send email
email.send()

SQL Template Rendering

from arrows.template_renderer import render_template

# Render SQL template
sql = render_template('path/to/template.sql', table_name='my_table', date='2024-01-01')

Core API

Google Sheets

• fetch_arrow() - Read data from Google Sheet as Arrow format
• arrow_to_googlesheet() - Write Arrow data to Google Sheet
• get_sheet() - Get Sheet object
• get_spreadsheet() - Get Spreadsheet object
• create_spreadsheet() - Create new Spreadsheet

Redshift

• fetch_arrow() - Query data from Redshift as Arrow format
• fetch_dataframe() - Query data from Redshift as DataFrame
• arrow_to_redshift() - Import Arrow data to Redshift
• unload() - Export Redshift query results to S3
• copy() - Copy data from S3 to Redshift
• execute_sql() - Execute SQL on Redshift
• execute_sql_file() - Execute SQL file

S3

• arrow_to_s3() - Store Arrow data to S3
• polars_to_s3() - Store Polars DataFrame to S3
• get_dataset() - Get S3 dataset
• create_dataset() - Create new S3 dataset

Gmail

• send_email() - Send an email
• Email - Class for constructing and sending emails

Development

uv sync                      # install with every component
uv run pytest                # unit tests, no credentials needed
uv run ruff check .          # lint
uv run ruff format .         # format; CI checks this
uv run arrows doctor         # check real connectivity for what is configured

Publishing a new version is docs/RELEASING.md.

Notes

1. Load what you need: arrows.load(...) fails fast on a missing credential; arrows doctor checks a real round-trip
2. Google Sheets permissions: the OAuth token needs the spreadsheets and drive scopes; Gmail needs gmail.send
3. S3 permissions: Redshift UNLOAD/COPY needs the cluster's IAM role to reach the same bucket your session does
4. Data formats: everything is Apache Arrow in the middle; check type compatibility at the edges
5. Breaking change in 0.2: redshift.fetch_arrow(..., engine='s3') now renders its SQL with Jinja2 ({{ name }}) like every other entry point, instead of str.format ({name})