boti-data

boti-data is the data access and data transformation layer of the Boti ecosystem.

It builds on top of boti and gives teams a reusable interface for working with structured data across databases, parquet datasets, schema-controlled transformations, and distributed or partitioned loading workflows.

What boti-data is for

Many teams have the same recurring problem: business logic depends on data that lives in multiple places, arrives in slightly different shapes, and is loaded through a mix of notebooks, scripts, ad hoc SQL, and one-off helpers.

boti-data helps turn that into a more coherent data access layer.

It is designed for codebases that need to:

• connect to named data sources consistently
• reflect or model database tables without hand-writing everything up front
• load data through a gateway instead of bespoke query snippets everywhere
• normalise and validate schemas before downstream use
• combine parquet and database workflows in one library
• scale from simple local reads to partitioned or distributed loading

Problems boti-data solves

boti-data is useful when data code is suffering from issues like:

• repeated connection boilerplate across notebooks and services
• slow, fragile query code copied from place to place
• inconsistent schema assumptions between producers and consumers
• difficult transitions from exploratory analysis to reusable pipelines
• manual join and field-mapping logic repeated in many modules
• no common abstraction for loading data from SQL and parquet sources

By centralising those patterns, boti-data reduces duplicated plumbing and makes transformations easier to reason about.

Why boti-data can make a huge difference

The biggest benefit of boti-data is that it creates a shared data interface between infrastructure and business logic.

That means teams can spend less time rewriting access code and more time working on actual transformations, validation rules, and downstream decisions.

It can make a major difference when:

• analysts and engineers share the same source systems
• a notebook prototype needs to become production code
• multiple data products depend on the same tables or parquet layouts
• schema drift is a recurring source of errors
• large extracts need partitioning or distributed execution
• teams want a clean boundary between connection details and transformation logic

Domain areas where it is especially valuable

boti-data is intentionally general-purpose, but it is especially strong in domains where structured operational data must be transformed into reliable analytical or decision-ready datasets.

Examples include:

• analytics engineering: building reusable source loaders, schema maps, and standardised transformations
• business operations: consolidating data from transactional systems, planning tools, and operational databases
• finance and controlling: reconciling structured data with explicit schema expectations and repeatable joins
• risk, compliance, and audit: validating input shape, tracing transformations, and standardising access patterns
• customer and product analytics: joining behavioural and operational datasets with less custom plumbing
• supply chain and logistics: unifying inventory, movement, order, and status data from several systems
• data platform and internal tooling: giving teams a common gateway layer instead of ad hoc connectors
• ML feature preparation: building reliable dataset assembly steps from SQL and parquet sources

In those settings, the gains are not just convenience. They show up as better reuse, fewer integration bugs, and faster movement from exploration to production.

Core capabilities

• SQL database resources
• async and sync database access helpers
• SQLAlchemy model reflection and registries
• connection catalogues
• parquet resources and readers
• gateway-style loading APIs
• filter expressions
• schema normalisation and validation helpers
• field mapping and join helpers
• partitioned and distributed data workflows

Installation

Install directly:

pip install boti-data

Or install through the core package extra:

pip install "boti[data]"

boti-data doesn't bundle a database driver by default — SQLAlchemy resolves the driver package from your connection URL's scheme (mysql+pymysql://, postgresql+asyncpg://, ...), so install the one matching your database:

pip install "boti-data[mysql]"      # asyncmy + pymysql
pip install "boti-data[postgres]"   # asyncpg + psycopg[binary]

If a Postgres table has a pgvector vector column, install boti-data[pgvector] too — this registers pgvector's Vector type with SQLAlchemy's reflection so SqlAlchemyModelBuilder/SqlRepository/DataGateway map that column to a usable list[float] (with its dimension preserved) instead of an unreflectable NullType:

pip install "boti-data[pgvector]"

The base install is deliberately light: single-row transactional access (SqlRepository/AsyncSqlRepository, see below) works out of the box, with no dask/pandas/polars pulled in. The dataframe-shaped, bulk-oriented side of the package (DataGateway/DataHelper, ParquetPipeline, HybridDataset, the sink pipelines, ...) needs the dataframes extra:

pip install "boti-data[dataframes]"  # dask + pandas + polars + boti-dask

Every name importable from boti_data is loaded lazily on first access, so import boti_data alone never touches dask/pandas/polars — only accessing a dataframe-shaped name does.

Imports

boti-data uses the top-level Python package boti_data. Dask runtime/session/resilience helpers live in boti_dask and are not re-exported here.

from boti_data import (
    AsyncFrameEnricher,
    AttachmentSpec,
    CsvSink,
    CsvSinkConfig,
    ConnectionCatalog,
    DatacubeConfig,
    DatacubeContract,
    DataGateway,
    DataHelper,
    FieldMap,    
    HybridDataset,
    JsonlSink,
    JsonlSinkConfig,
    ParquetDataConfig,
    ParquetDataResource,
    ParquetPipeline,
    SinkRegistry,
    SinkPipeline,
    available_sinks,
    create_sink,
    register_sink,
    AsyncSqlRepository,
    SqlAlchemyModelBuilder,
    SqlDatabaseConfig,
    SqlDatabaseResource,
    SqlRepository,
)

For Dask runtime/session/resilience utilities, import from boti_dask directly:

from boti_dask import (
    DaskSession,
    UniqueValuesExtractor,
    apply_recommended_dask_config,
    async_safe_compute,
    async_safe_gather,
    async_safe_head,
    async_safe_persist,
    async_safe_wait,
    dask_is_empty,
    dask_is_probably_empty,
    dask_session,
    inspect_graph,
    safe_compute,
    safe_gather,
    safe_head,
    safe_persist,
    safe_wait,
)

Lower-level modules are also available:

from boti_data.db import SqlDatabaseConfig, SqlDatabaseResource
from boti_data.datacube import DatacubeConfig, DatacubeContract
from boti_data.gateway import DataGateway
from boti_data.parquet import ParquetDataConfig, ParquetDataResource
from boti_data.schema import validate_schema

Datacube backend

DataGateway also supports a callable-backed datacube backend for in-process cube loaders while keeping the same return-type API (pandas/dask/arrow/polars).

import pandas as pd

from boti_data import DatacubeConfig, DatacubeContract, DataGateway


def loader(request):
    frame = pd.DataFrame({"id": [1, 2], "status": ["active", "inactive"]})
    if request.filters.get("status__exact") == "active":
        frame = frame[frame["status"] == "active"]
    return frame


contract = DatacubeContract(
    request_transformer=lambda request: request.model_copy(update={"cube": request.cube or "orders_v2"}),
    frame_transformer=lambda frame, request: frame.assign(cube=request.cube),
)

with DataGateway(
    DatacubeConfig(loader=loader, default_cube="orders", contract=contract),
    table="orders",
    sticky_filters={"status__exact": "active"},
) as gateway:
    df = gateway.load(return_type="pandas")

You can also configure this path with DataGateway.from_config({"backend": "datacube", ...}) or DataHelper({"backend": "datacube", ...}).

See docs/DATACUBE_CONTRACT.md for hook ordering, loader precedence, and validator rejection guidance. See docs/SECURITY_HARDENING.md for SQL/raw-sql policy and distributed credential hardening defaults. For a runnable rejection flow, see examples/data_facade_datacube_contract_rejection.py.

DataHelper

DataHelper is the primary entry point for most use cases. It is a thin facade over DataGateway that provides a clean, consistent interface for loading data whether you are working locally, in a notebook, or inside a distributed Dask pipeline.

Creating a DataHelper

DataHelper accepts a DataGateway, a backend config object, or a plain dict:

from boti_data import DataHelper, SqlDatabaseConfig

# From a config object
config = SqlDatabaseConfig(
    connection_url="mysql+pymysql://user:pass@host/mydb",
    query_only=True,
)
helper = DataHelper(config, table="orders")

# From a dict (useful for config-driven setups)
helper = DataHelper({
    "backend": "sqlalchemy",
    "connection_url": "mysql+pymysql://user:pass@host/mydb",
    "table": "orders",
    "query_only": True,
})

# From keyword arguments
helper = DataHelper(
    backend="sqlalchemy",
    connection_url="mysql+pymysql://user:pass@host/mydb",
    table="orders",
)

Using DataHelper as a context manager ensures connections are properly closed:

with DataHelper(config, table="orders") as helper:
    df = helper.load(status="confirmed")

Async context managers are also supported:

async with DataHelper(config, table="orders") as helper:
    df = await helper.aload(status="confirmed")

For synchronous scripts that still need to call async load paths, use the explicit sync bridge:

with DataHelper(config, table="orders") as helper:
    df = helper.aload_sync(status="confirmed")

If an event loop is already running (for example notebooks or ASGI handlers), use await helper.aload(...) directly.

ParquetPipeline

ParquetPipeline is the first orchestration layer built on top of DataHelper and HybridDataset. It materializes a lazy source load into a parquet dataset directory and can optionally reload the written dataset through ParquetReader.

from boti_data import DataHelper, ParquetPipeline

helper = DataHelper(
    backend="sqlalchemy",
    connection_url=f"sqlite:///{db_path}",
    poolclass="sqlalchemy.pool.NullPool",
    query_only=False,
    table="source_events",
)

pipeline = ParquetPipeline(
    helper,
    {
        "backend": "parquet",
        "storage_path": "/tmp/source_events_dataset",
        "partition_on": ["partition_date"],
    },
    date_field="event_date",
)

result = pipeline.materialize(
    filters={"status__exact": "active"},
    reload=True,
    reload_options={"filters": {"partition_date__exact": "2026-04-17"}},
)

assert result.reloaded is True
frame = result.frame

Use materialize() / amaterialize() for the one-step workflow, or call to_parquet() and from_parquet() separately when you want explicit control over the write and reload phases. See examples/data_parquet_pipeline.py for a runnable end-to-end example.

SinkPipeline and Sinks

Phase 3 adds a minimal sink/plugin layer next to ParquetPipeline.

• SinkPipeline orchestrates DataHelper / HybridDataset loads into any sink implementing the pipeline sink contract
• CsvSink and JsonlSink are write-only dataset sinks (.csv / .jsonl shards)
• ParquetPipeline now uses the same sink abstraction internally via ParquetSink

Use the sink registry for named sink creation:

from boti_data import SinkPipeline, available_sinks

assert "jsonl" in available_sinks()

pipeline = SinkPipeline(
    helper,
    "jsonl",
    sink_config={
        "storage_path": "/tmp/source_events_jsonl",
        "partition_on": ["partition_date"],
    },
    date_field="event_date",
)
result = pipeline.write(filters={"status__exact": "active"})

from boti_data import CsvSink, DataHelper, SinkPipeline

helper = DataHelper(
    backend="sqlalchemy",
    connection_url=f"sqlite:///{db_path}",
    poolclass="sqlalchemy.pool.NullPool",
    query_only=False,
    table="source_events",
)

sink = CsvSink(
    {
        "storage_path": "/tmp/source_events_csv",
        "partition_on": ["partition_date"],
    }
)

pipeline = SinkPipeline(helper, sink, date_field="event_date")
result = pipeline.write(filters={"status__exact": "active"})

assert result.files

CsvSink/JsonlSink are intentionally write-only in this release. Use ParquetPipeline when you need write + reload orchestration. See examples/data_csv_sink_pipeline.py and examples/data_jsonl_sink_pipeline.py for runnable examples.

Enrichment v1

AsyncFrameEnricher adds declarative attachment-based enrichment before downstream writes. Select attachments by AttachmentSpec.key via cols=[...], and enforce bounded unique-value extraction with max_unique_values.

from boti_data import AsyncFrameEnricher, AttachmentSpec

enricher = AsyncFrameEnricher(
    [
        AttachmentSpec(
            key="customer_segment",
            required_cols={"customer_id"},
            attachment_fn=customer_segment_attachment,
            col_to_kwarg={"customer_id": "ids"},
            left_on=["customer_id"],
            right_on=["id"],
            drop_cols=["id"],
            max_unique_values=5000,
        )
    ]
)
enriched = await enricher.aenrich(base_frame, cols=["customer_segment"])

SinkPipeline accepts enricher=... and optional enrich_cols=[...] to apply enrichment right before sink writes.

Raw sql= safety policy (eager SQL only)

DataGateway supports convenience sql="..." reads for eager SQL paths, with explicit safety controls:

• raw SQL is validated as read-only single-statement SELECT/WITH
• mutating SQL (INSERT, UPDATE, DELETE, DROP, etc.) is rejected
• multi-statement SQL is rejected
• lazy SQL still requires statement + model (raw sql= is eager-only)

Use constructor policy raw_sql_policy to control availability:

• raw_sql_policy="readonly_opt_in" (default): call site must pass allow_raw_sql=True
• raw_sql_policy="disabled": raw sql= is blocked even if allow_raw_sql=True

from boti_data import DataGateway, SqlDatabaseConfig

config = SqlDatabaseConfig(connection_url="sqlite:///example.db", query_only=True)

with DataGateway(config, raw_sql_policy="readonly_opt_in") as gateway:
    df = gateway.load(
        sql="SELECT id, status FROM users WHERE status = :status",
        params={"status": "active"},
        as_pandas=True,
        allow_raw_sql=True,
    )

Prefer statement + model for production query paths whenever possible, especially when you need lazy/partitioned execution.

---

Output engines: pandas, polars, and dask

DataHelper exposes three engine-bound views that pin the output type for a call chain:

helper = DataHelper(config, table="orders")

# Always returns pandas.DataFrame
df = helper.pandas.load(status="confirmed")

# Always returns polars.DataFrame
df = helper.polars.load(status="confirmed")

# Always returns dask.dataframe.DataFrame (lazy)
df = helper.dask.load(status="confirmed")

These are the cleanest way to use a single helper across different downstream contexts. You can also pass return_type explicitly to load or aload when you need more control:

# Explicit return_type on a single call
df = helper.load(status="confirmed", return_type="polars")
df = helper.load(status="confirmed", return_type="pandas")
df = helper.load(status="confirmed", return_type="dask")
df = helper.load(status="confirmed", return_type="arrow")  # pyarrow.Table

Choosing an output engine

Engine	Type returned	Best for
pandas	pandas.DataFrame	Small-to-medium data, notebooks, local analysis
polars	polars.DataFrame	CPU-intensive transforms, single-machine performance
arrow	pyarrow.Table	Zero-copy interchange, serialisation, ML pipelines
dask	dask.dataframe.DataFrame	Large datasets, distributed clusters, lazy evaluation
auto	decided at runtime	Unknown result size; boti-data uses backend-aware size heuristics

return_type="auto" uses pandas for small results and switches to Dask for larger scans, with backend-specific checks:

• SQL: probes up to 10,000 rows (and respects limit / explicit partitioned controls).
• Parquet: uses pandas when the selected scan is small enough to estimate eagerly (typically up to 4 files and <= 32 MB total), otherwise Dask.

Use auto when result size is uncertain and you want sensible defaults without hardcoding an engine.

---

Non-distributed usage

For local analysis, notebooks, or small-scale pipelines, use DataHelper without any Dask cluster. The default output is a Dask DataFrame, but you can force pandas or polars.

from boti_data import DataHelper, SqlDatabaseConfig

config = SqlDatabaseConfig(
    connection_url="sqlite:///local.db",
    query_only=True,
)

with DataHelper(config, table="orders") as helper:
    # Pandas — eager, in-memory
    df = helper.pandas.load(status="shipped")

    # Polars — eager, high-performance single-machine
    df = helper.polars.load(status="shipped")

    # Date range load with pandas output
    df = helper.pandas.load_period("created_at", "2024-01-01", "2024-03-31")

For async contexts (FastAPI, async services):

async def get_orders(status: str) -> pd.DataFrame:
    async with DataHelper(config, table="orders") as helper:
        return await helper.pandas.aload(status=status)

---

Dask resilience helpers

boti-data integrates with boti_dask for opt-in Dask runtime and resilience workflows.

Available helpers:

• inspect_graph(...)
• safe_compute(...)
• safe_head(...)
• safe_gather(...)
• safe_persist(...)
• safe_wait(...)
• async_safe_compute(...)
• async_safe_head(...)
• async_safe_gather(...)
• async_safe_persist(...)
• async_safe_wait(...)
• dask_is_probably_empty(...)
• dask_is_empty(...)
• UniqueValuesExtractor
• apply_recommended_dask_config(...)

Core rules

• Dask DataFrame-like graphs do not silently fall back to local threaded compute.
• Retry behavior is limited to recoverable communication failures.
• The helpers are opt-in and designed to complement explicit DaskSession ownership.

Minimal example

import asyncio
import dask
import dask.dataframe as dd
import pandas as pd

from boti_dask import (
    apply_recommended_dask_config,
    async_safe_compute,
    dask_session,
    inspect_graph,
    safe_compute,
    safe_persist,
    safe_wait,
)

frame = dd.from_pandas(pd.DataFrame({"id": [1, 2, 3]}), npartitions=2)

with apply_recommended_dask_config():
    with dask_session(cluster_kwargs={"n_workers": 1, "threads_per_worker": 1, "processes": False}) as client:
        metrics = inspect_graph(frame)
        persisted = safe_persist(frame, dask_client=client)
        safe_wait(persisted, dask_client=client)
        total = safe_compute(frame["id"].sum(), dask_client=client)
        preview = safe_head(frame, n=2, dask_client=client)

        async def run_async() -> int:
            delayed_value = dask.delayed(lambda: 6 * 7)()
            return await async_safe_compute(delayed_value, dask_client=client)

        async_total = asyncio.run(run_async())

Dry-run and preview on gateway loads

Use dry_run=True to build and inspect a lazy Dask load graph without materializing it, and preview(...) / apreview(...) to safely sample rows inside the session:

with DataHelper.session(scheduler_address="tcp://scheduler:8786", verify_connectivity=True) as client:
    with DataHelper(config, table="transactions") as helper:
        ddf = helper.dask.load(year=2024, dry_run=True, diagnostics=True)
        preview = helper.dask.preview(year=2024, n=5)

Resilient joins

For large indexed distributed joins, opt into resilient persistence:

joined = helper.left_join(
    left,
    right,
    join_key="id",
    join_schema_map={"id": "Int64"},
    persist=True,
    resilient=True,
    diagnostics=True,
)

Parquet sources

from boti_data import DataHelper, ParquetDataConfig

config = ParquetDataConfig(
    parquet_storage_path="/data/orders/",
    parquet_start_date=date(2024, 1, 1),
    parquet_end_date=date(2024, 3, 31),
)

with DataHelper(config) as helper:
    df = helper.pandas.load()
    df = helper.polars.load()

---

Distributed usage with Dask

For large datasets or cluster workloads, DataHelper integrates natively with Dask. The DataHelper.session() factory creates a DaskSession that manages cluster and client lifecycle.

Local cluster (development)

from dask.distributed import LocalCluster
from boti_data import DataHelper, SqlDatabaseConfig

config = SqlDatabaseConfig(
    connection_url="mysql+pymysql://user:pass@host/mydb",
    query_only=True,
    worker_connection_env_var="DB_URL",  # see pickleable section below
)

with DataHelper.session(cluster_factory=LocalCluster) as client:
    with DataHelper(config, table="orders") as helper:
        # Returns dask.dataframe.DataFrame — lazy, partitioned
        ddf = helper.dask.load(status="confirmed")

        # Trigger computation
        df = ddf.compute()

Remote cluster

with DataHelper.session(scheduler_address="tcp://scheduler:8786") as client:
    with DataHelper(config, table="events") as helper:
        ddf = helper.dask.load(region="EU", return_type="dask")
        result = ddf.groupby("customer_id").agg({"amount": "sum"}).compute()

Persisting on the cluster

Use persist=True to push the loaded data into distributed memory before further computation. This avoids re-reading from the database on every downstream operation:

with DataHelper.session(scheduler_address="tcp://scheduler:8786") as client:
    with DataHelper(config, table="transactions") as helper:
        # Data is loaded and held in cluster memory
        ddf = helper.load(year=2024, persist=True)

        # Subsequent operations reuse the persisted graph
        monthly = ddf.groupby("month").agg({"amount": "sum"}).compute()
        by_region = ddf.groupby("region").size().compute()

        # Preview persisted data before the managed session closes.
        preview = helper.preview(year=2024, n=5, persist=True)

Shared sessions for repeated notebook cells

When repeated cells need to reuse the same scheduler connection, opt into shared-session reuse explicitly:

with DataHelper.session(
    scheduler_address="tcp://scheduler:8786",
    shared=True,
    shared_key="notebook-main-cluster",
) as client:
    with DataHelper(config, table="transactions") as helper:
        ddf = helper.dask.load(year=2024)

HybridDataset distributed pattern

Use this pattern when you split reads across historical and live sources but want one distributed execution context. HybridDataset composes two helpers, and DataHelper.session(...) keeps scheduler ownership explicit.

import asyncio

from dask.distributed import LocalCluster
from boti_data import DataHelper, HybridDataset

historical = DataHelper(backend="sqlalchemy", connection_url="sqlite:///events.db", table="events_hist")
live = DataHelper(backend="sqlalchemy", connection_url="sqlite:///events.db", table="events_live")
dataset = HybridDataset(historical, live, date_field="event_date", split_date="2026-04-18")

with LocalCluster(n_workers=1, threads_per_worker=1, processes=False, dashboard_address=":0") as cluster:
    with DataHelper.session(
        scheduler_address=cluster.scheduler_address,
        verify_connectivity=True,
        shared=True,
        shared_key="hybrid-dataset-distributed",
    ):
        ddf = dataset.dask.load(start="2026-04-14", end="2026-04-19", diagnostics=True)
        frame = ddf.compute()
        eager_frame = asyncio.run(
            dataset.aload(start="2026-04-16", end="2026-04-18", return_type="pandas")
        )

For a complete runnable script, see examples/data_hybrid_dataset_distributed.py.

Semi-join across distributed frames

import pandas as pd

active_customers = pd.Series([1001, 1002, 1003, 1099])

with DataHelper(config, table="orders") as helper:
    # Loads only rows where customer_id is in active_customers
    ddf = helper.semi_join(active_customers, on="customer_id")
    df = ddf.compute()

semi_join also accepts Dask Series, enabling fully lazy distributed joins:

with DataHelper(config, table="customers") as customer_helper:
    with DataHelper(config, table="orders") as order_helper:
        active_ids = customer_helper.dask.load(active=True)["customer_id"]

        # Lazy — no computation happens yet
        orders_ddf = order_helper.semi_join(active_ids, on="customer_id")

        # Single compute triggers both loads
        result = orders_ddf.compute()

---

The pickleable setting in distributed systems

When Dask distributes tasks across workers, it serialises (pickles) the task function and all its arguments to send them over the network. This creates a problem: database connection objects, engine pools, and credentials cannot be pickled.

boti-data addresses this through the worker_connection_env_var setting on SqlDatabaseConfig.

How it works

Instead of serialising the full SqlDatabaseConfig (which contains the connection URL and credentials), boti-data extracts a minimal WorkerSqlConfig for each worker task. If worker_connection_env_var is set, workers read the DSN from that environment variable instead of having it embedded in the task payload.

Scheduler                              Worker
─────────                              ──────
SqlDatabaseConfig (full config)        WorkerSqlConfig (minimal, safe to pickle)
  connection_url = "mysql://..."   →     connection_env_var = "DB_URL"
  pool_size = 10                         query_only = True
  ...                                    pool_recycle = 1800
                                         (reads DB_URL from os.environ on worker)

Setting it up

Step 1. Set the environment variable on all workers. For a local cluster:

export DB_URL="mysql+pymysql://user:pass@host/mydb"

For a Kubernetes-deployed cluster, inject it as a secret.

Step 2. Reference the variable in your config:

from boti_data import DataHelper, SqlDatabaseConfig

config = SqlDatabaseConfig(
    connection_url="mysql+pymysql://user:pass@host/mydb",
    query_only=True,
    worker_connection_env_var="DB_URL",  # workers use this instead of pickling credentials
)

Step 3. Use DataHelper normally. Credential serialisation is handled transparently:

with DataHelper.session(scheduler_address="tcp://scheduler:8786") as client:
    with DataHelper(config, table="orders") as helper:
        ddf = helper.dask.load(status="confirmed")
        result = ddf.compute()

Why this matters

Without worker_connection_env_var, using a real database DSN with distributed Dask will either:

• fail with a pickle error (connection pool objects are not serialisable)
• embed plaintext credentials in task payloads that flow through scheduler memory and worker logs

Setting worker_connection_env_var prevents both problems and is the recommended approach for any distributed SQL workflow.

Parquet in distributed settings

Parquet resources use fsspec for filesystem access. The filesystem object is not pickled directly; instead, ParquetDataResource uses a fs_factory callable or a filesystem_profile name that workers can use to reconstruct the filesystem independently. When filesystem_profile is used, filesystem creation is routed through ConnectionCatalog adapters so retry/cache behavior stays consistent.

from boti_data import DataHelper, ParquetDataConfig, ConnectionCatalog

catalog = ConnectionCatalog()
catalog.load_filesystem("s3_prod", prefix="S3_")  # reads S3_ENDPOINT, S3_KEY, etc.

config = ParquetDataConfig(
    filesystem_profile="s3_prod",  # workers resolve filesystem from catalog
    parquet_storage_path="s3://my-bucket/orders/",
)

with DataHelper.session(cluster_factory=LocalCluster) as client:
    with DataHelper(config) as helper:
        ddf = helper.dask.load()
        result = ddf.compute()

---

Choosing between distributed and non-distributed

Use the following as a guide:

Scenario	Recommended approach
Exploratory analysis in a notebook	helper.pandas.load() — simple, no overhead
Single-machine pipeline, large-ish data	helper.polars.load() — fast, low memory
Result size unknown at design time	helper.load(return_type="auto") — adapts
Data does not fit in one machine's RAM	helper.dask.load() + local or remote cluster
Heavy transforms over millions of rows	helper.dask.load() + Dask cluster
Async service (FastAPI, ASGI)	await helper.pandas.aload() or await helper.dask.aload()
Sync script that needs async load path	helper.aload_sync(...)
Joining two large tables on a cluster	helper.semi_join(series, on="key")
Scheduled overnight batch job	Dask cluster + persist=True for multi-pass jobs

Rule of thumb: start with pandas, switch to polars when single-machine performance matters, and move to dask when data size exceeds available RAM or when the task benefits from parallelism across workers.

---

Examples

Repository (lightest-weight, no dataframe dependencies)

For single-row transactional access — get/insert/update/delete by primary key — without pulling in dask/pandas/polars at all. Rows come back as plain dict[str, Any], never a DataFrame.

query_only defaults to True on SqlRepository/AsyncSqlRepository themselves and always wins over whatever the passed-in config.query_only says — insert/update/delete raise until you opt in explicitly with query_only=False. This is deliberate: a config object might be shared with or reused from a context where writes are already enabled, and a repository shouldn't silently inherit write access nobody asked it for.

from boti_data import SqlDatabaseConfig, SqlRepository

config = SqlDatabaseConfig(connection_url="sqlite:///example.db")

with SqlRepository(config, "users", query_only=False) as users:
    row = users.insert({"id": 1, "name": "Ada"})
    row = users.get(1)              # -> dict | None
    row = users.update(1, {"name": "Grace"})
    deleted = users.delete(1)       # -> bool

Omit query_only=False (or read-only access is all you need) and writes raise instead of silently succeeding:

with SqlRepository(config, "users") as users:
    users.get(1)                    # fine
    users.insert({"id": 2, "name": "Bob"})  # raises SQLAlchemyError

An async counterpart, AsyncSqlRepository, mirrors the same four methods as coroutines and is used as an async context manager:

from boti_data import AsyncSqlRepository

async with AsyncSqlRepository(config, "users", query_only=False) as users:
    row = await users.get(1)

Multi-table atomic transactions

By default each SqlRepository owns its own connection and commits after every write — fine for independent single-row operations, but not when several writes (potentially across different tables) need to succeed or fail together. Pass an already-open session instead of config to bind a repository to a transaction you control; writes flush() instead of committing, so nothing is durable until you commit the session yourself:

with SqlDatabaseResource(config) as db, db.session() as session:
    with SqlRepository(session=session, table_name="audit_log") as audit_log:
        with SqlRepository(session=session, table_name="embeddings") as embeddings:
            audit_log.insert({"event": "enroll"})
            embeddings.insert({"subject_id": 1, "vector": [...]})
    session.commit()  # both rows commit together, or neither does on an exception

SqlUnitOfWork wraps this pattern for the common case — one config, several tables, one commit/rollback:

from boti_data import SqlDatabaseConfig, SqlUnitOfWork

config = SqlDatabaseConfig(connection_url="postgresql+psycopg://user:pass@host/mydb")

with SqlUnitOfWork(config, query_only=False) as uow:
    uow.repository("audit_log").insert({"event": "enroll"})
    uow.repository("embeddings").insert({"subject_id": 1, "vector": [...]})
# commits both on clean exit; rolls back both on any exception

query_only on a session-bound SqlRepository is inert — the session's own class (e.g. a ReadOnlySession) already determines writability, since this path never builds a new guarded session itself. AsyncSqlUnitOfWork mirrors SqlUnitOfWork, with an async repository() method (returns an already-set-up AsyncSqlRepository, so callers don't need to async with each one individually):

from boti_data import AsyncSqlUnitOfWork

async with AsyncSqlUnitOfWork(config, query_only=False) as uow:
    audit_log = await uow.repository("audit_log")
    embeddings = await uow.repository("embeddings")
    await audit_log.insert({"event": "enroll"})
    await embeddings.insert({"subject_id": 1, "vector": [...]})

Nearest-neighbour search (pgvector)

boti_data.db.nearest_neighbors/vector_distance build an ORDER BY <distance> LIMIT k statement against a pgvector column, with the query vector bound as a driver parameter — never string-interpolated into the SQL, regardless of how large the vector is. Requires boti-data[pgvector] (see Installation) and a table reflected via SqlAlchemyModelBuilder/SqlRepository/DataGateway.

from boti_data import DataGateway, SqlDatabaseConfig
from boti_data.db import SqlAlchemyModelBuilder, nearest_neighbors

config = SqlDatabaseConfig(connection_url="postgresql+psycopg://user:pass@host/mydb", query_only=True)
gateway = DataGateway(config)

model = SqlAlchemyModelBuilder(gateway.resource.engine, "embeddings").build_model()
query_vector = [...]  # e.g. a 512-float face embedding

stmt = nearest_neighbors(model, model.vector, query_vector, k=10, metric="cosine")
matches = gateway.load(statement=stmt, model=model, as_pandas=True)

Supported metric values map directly to pgvector's operators: "cosine" (<=>), "l2" (<->), "inner_product" (<#>), "l1" (<+>). Add ordinary filtering with .where(...) on the statement before passing it to gateway.load(...).

SQL resource (low-level)

from boti_data import SqlDatabaseConfig, SqlDatabaseResource

config = SqlDatabaseConfig(connection_url="sqlite:///example.db", query_only=True)

with SqlDatabaseResource(config) as db:
    with db.session() as session:
        rows = session.execute(...)

Gateway (mid-level)

from boti_data import DataGateway, SqlDatabaseConfig

gateway = DataGateway(
    backend="sqlalchemy",
    config=SqlDatabaseConfig(connection_url="sqlite:///example.db", query_only=True),
)

DataHelper — local pandas

from boti_data import DataHelper, SqlDatabaseConfig

config = SqlDatabaseConfig(
    connection_url="postgresql+asyncpg://user:pass@host/mydb",
    query_only=True,
)

with DataHelper(config, table="sales") as helper:
    df = helper.pandas.load(year=2024, region="EMEA")
    print(df.head())

DataHelper — local polars

with DataHelper(config, table="sales") as helper:
    df = helper.polars.load(year=2024)
    summary = df.group_by("region").agg(pl.col("amount").sum())

DataHelper — lazy Dask, no cluster

with DataHelper(config, table="sales") as helper:
    ddf = helper.dask.load(year=2024)
    # Graph is not executed yet; chain transforms lazily
    result = ddf.groupby("region")["amount"].sum().compute()

DataHelper — distributed Dask cluster

from dask.distributed import LocalCluster
from boti_data import DataHelper, SqlDatabaseConfig

config = SqlDatabaseConfig(
    connection_url="mysql+pymysql://user:pass@host/mydb",
    query_only=True,
    worker_connection_env_var="DB_URL",
)

with DataHelper.session(cluster_factory=LocalCluster, n_workers=4) as client:
    with DataHelper(config, table="events") as helper:
        ddf = helper.dask.load(event_type="purchase", persist=True)
        result = ddf.groupby("user_id").size().compute()

DataHelper — async service

from boti_data import DataHelper, SqlDatabaseConfig

config = SqlDatabaseConfig(
    connection_url="mysql+asyncmy://user:pass@host/mydb",
    query_only=True,
)

async def load_orders(status: str) -> pd.DataFrame:
    async with DataHelper(config, table="orders") as helper:
        return await helper.pandas.aload(status=status)

DataHelper — date-range load

with DataHelper(config, table="transactions") as helper:
    # Inclusive date range; dt_field is the semantic field name
    df = helper.pandas.load_period("created_at", "2024-01-01", "2024-06-30")

HybridDataset — historical + live composition

historical = DataHelper(backend="parquet", storage_path="/data/historical", parquet_filename="orders")
live = DataHelper(backend="sqlalchemy", connection_url="sqlite:///live.db", table="orders_live")

dataset = HybridDataset(
    historical,
    live,
    date_field="order_date",
    split_date="2026-04-18",
)

# mixed window (auto -> dask for hybrid composition)
ddf = dataset.load(start="2026-04-15", end="2026-04-20", return_type="auto")

# explicit source override through engine-bound view
pdf = dataset.pandas.load(start="2026-04-18", end="2026-04-20", source="live")

DataHelper — parquet source

from boti_data import DataHelper, ParquetDataConfig
from datetime import date

config = ParquetDataConfig(
    parquet_storage_path="/data/warehouse/orders/",
    parquet_start_date=date(2024, 1, 1),
    parquet_end_date=date(2024, 6, 30),
)

with DataHelper(config) as helper:
    df = helper.pandas.load()
    ddf = helper.dask.load()  # lazy, partitioned read

Connection catalog

from boti_data import ConnectionCatalog, DataHelper

catalog = ConnectionCatalog()
catalog.load_sql("prod", prefix="PROD_DB_")  # reads PROD_DB_URL, PROD_DB_POOL_SIZE, etc.
catalog.load_sql("reporting", prefix="REPORT_DB_")

prod_config = catalog.sql_config("prod")
report_config = catalog.sql_config("reporting")

with DataHelper(prod_config, table="orders") as helper:
    df = helper.pandas.load(status="confirmed")

---

Relationship to boti

boti-data depends on boti, and reuses:

• logging
• resource lifecycle
• secure I/O helpers
• project/environment utilities

If you only need the runtime primitives, install boti. If you need a stronger data access and transformation layer, install boti-data or boti[data].

Development & Deployment

See docs/DEPLOYMENT.md for publishing instructions.