Muscles Data SQL

Muscles SQL is a data layer package for SQL databases:

• model-to-table mapping
• engine and session management
• repository CRUD and advanced query API (filters/operators/joins/aggregates)
• transactions and Unit of Work (nested/savepoint/retry helpers)
• native SQL migrations through the shared muscles-data migration registry
• inspect/doctor support with machine-readable diagnostics

Related Repositories

• muscles - core schemas, actions, DI and canonical documentation.
• muscles-documents - document metadata/state can use SQL persistence in applications.
• muscles-ai - AI/RAG flows can use SQL-backed application state without making SQL generic storage.
• muscles-otel - observability hooks around SQL-backed flows.
• muscles-benchmarks - SQL transaction and mapping regression checks.

Quickstart

This low-level API is useful for standalone scripts and libraries that explicitly own their database lifetime. Application packages should use the shared app.resource integration below instead of constructing SqlDatabase objects.

The lightweight SQL API has no mandatory third-party dependencies. SQLite uses Python's standard library; PostgreSQL requires the postgres extra. The named-resource integration requires data; SQLAlchemy repositories and CLI require legacy. In this checkout install from the library directory using pip install -e '.[postgres]', .[data] or .[legacy] as needed. Publishing the renamed distribution is a separate step.

from muscles.data.sql import SqlDatabase

db = SqlDatabase.sqlite("app.db")
try:
    with db.transaction() as tx:
        tx.execute("CREATE TABLE IF NOT EXISTS users (id INTEGER, name TEXT)")
        tx.execute("INSERT INTO users VALUES (?, ?)", (1, "Denis"))
        users = tx.fetch_all("SELECT id, name FROM users WHERE id = ?", (1,))
finally:
    db.close()

SqlDatabase.postgres(dsn) uses psycopg and its native parameter syntax. A database owns one connection and is not a connection pool. Transactions from different threads are serialized on that connection; nested transactions in the same thread are rejected. Direct SQL does not enforce data policies; the caller must ensure access restrictions. Commit occurs on successful exit, and exceptions roll the transaction back.

muscles-data Resource Integration

The application composition root owns the catalog and DataRuntime. Register NativeSqlFactory there once, and load data.resources from the central application configuration:

from muscles.data import DataConfig, DataRuntime
from muscles.data.catalog import DataAdapterCatalog
from muscles.data.sql.data_adapter import NativeSqlFactory

class App:
    def __init__(self):
        # self.config is the application's shared Muscles configuration.
        catalog = DataAdapterCatalog.with_defaults()
        catalog.register(NativeSqlFactory())
        self.resource = DataRuntime(
            config=DataConfig.from_raw({
                "resources": self.config.get("data.resources").value() or {},
            }),
            catalog=catalog,
        )

The database resource itself is declared once in the application's central configuration, for example with type: native_sql, driver: postgres and url_env: DATABASE_URL. A package uses the existing resource and does not create another catalog, runtime or connection configuration:

from muscles.data.ports import SqlTransactionPort

sql = app.resource.connect("database", SqlTransactionPort)
with sql.transaction() as tx:
    tx.execute("CREATE TABLE IF NOT EXISTS audit (id INTEGER PRIMARY KEY)")

The adapter opens its native connection lazily on the first transaction and closes it with the owning DataRuntime. Direct SqlDatabase examples above are for standalone low-level use; application package code should resolve its port through app.resource.

python -m venv .venv
source .venv/bin/activate
pip install -e .

For local development with the test and legacy dependencies:

pip install -e ".[dev]"
pytest -q
muscles-data-sql doctor --url sqlite:///./app.db

Named SQL Connections

muscles-data-sql can manage multiple SQL connections without becoming a generic storage registry. The registry is SQL-only: it owns SQL connection configs, lazy SQLAlchemy EngineManager instances, sessions, inspect and doctor reports.

from muscles.data.sql import SqlConnectionConfig, SqlConnectionRegistry

registry = SqlConnectionRegistry(
    [
        SqlConnectionConfig(name="default", url="sqlite:///./app.db"),
        SqlConnectionConfig(name="analytics", url="sqlite:///./analytics.db", role="read"),
    ]
)

session = registry.session("analytics")
report = registry.inspect("analytics")

CLI diagnostics can read a JSON config:

{
  "connections": {
    "default": "sqlite:///./app.db",
    "analytics": {"url": "sqlite:///./analytics.db", "role": "read"}
  }
}

muscles-data-sql inspect --config sql-connections.json --connection analytics
muscles-data-sql doctor --config sql-connections.json --all

Diagnostic output uses safe URLs and does not print passwords from DSNs.

Advanced Query Example

from sqlalchemy import func
from muscles.data.sql import FilterClause, JoinClause, QuerySpec, SqlRepository

spec = QuerySpec(
    filters=[FilterClause("status", "eq", "active")],
    joins=[JoinClause(table=orders, on=users.c.id == orders.c.user_id)],
    select_columns=[users.c.id, func.count(orders.c.id).label("orders_total")],
    group_by=[users.c.id],
    order_by=[users.c.id.asc()],
)
rows = SqlRepository(session, users).aggregate(spec)

Data operation policies

SqlRepository accepts the policy pipeline and a request-scoped context from muscles-data. Policy predicates are combined with caller filters using AND; they apply to reads, counts, updates, deletes, creates, bulk creates and upserts. For update/delete, predicates are part of the SQL statement. Create and changed values are checked before execution. Under an active policy pipeline, upsert uses the constrained repository path instead of a native conflict statement.

from muscles.data import DataOperationContext
from muscles.data.sql import SqlRepository

repo = SqlRepository(
    session,
    documents,
    policies=data_runtime.policies,
    context=DataOperationContext(
        subject_type="account",
        subject_uid=account_uid,
        attributes={"workspace_uid": workspace_uid},
    ),
)
rows = repo.find({"owner_uid": account_uid}, limit=50)

The policy contract is backend-neutral; each adapter is responsible for translating predicates and enforcing them before pagination or mutation. Direct SQLAlchemy session use is outside this enforcement boundary.

Repository: https://github.com/butkoden/muscles-data-sql