Hamilton decorators, SDAX execution

sdax-hamilton brings Python SDAX's execution and resource lifecycle features to graphs written with Hamilton's real decorators. Keep your typed functions and familiar Hamilton imports. Add @execution for per-attempt timeouts and retries with exponential backoff and jitter, and @shutdown for cleanup sequenced in reverse dependency order.

For supported declarations, Hamilton still resolves configuration, bindings and decorator expansion. SDAX runs the resulting graph: consumers finish before their resources close, and dependent resources close before the resources they need. You describe the dependencies once; you do not write a second shutdown graph.

Get started Read the API guide

Two extra decorators

Add to your Hamilton graph	What it provides
@execution(timeout=5, retries=2)	Up to three attempts for retryable failures, with SDAX's backoff and jitter between attempts.
@shutdown(of=resource)	A cleanup function tied to an acquisition, ordered after its consumers and before its prerequisites are released.

Both come from sdax_hamilton. Continue importing graph decorators such as config, inject, parameterize and subdag from Hamilton itself. Cleanup can also have its own timeout and retry policy.

from hamilton.function_modifiers import inject, value
from sdax_hamilton import execution


@execution(timeout=5.0, retries=2, initial_delay=0.25, backoff_factor=2.0)
@inject(prefix=value("Hello"))
async def greeting(name: str, prefix: str) -> str:
    return f"{prefix}, {name}"

The quick start adds a resource and its shutdown function to a complete runnable example. Retries and shutdown explains jitter, cleanup ordering, cancellation and resource scopes.

Prepare once, run repeatedly

Driver(nodes).prepare(["greeting"]) compiles a reusable in-memory SDAX plan. Each invocation has separate values and acquisition records. Use execute() for ordinary results, or open() to keep owned results alive while the caller uses them. The frontend checks declared parameter/result bindings before execution and checks returned values by default.

Current compatibility

These pages describe the 0.2.0 alpha and follow subsequent development on main. The API is experimental and targets Python 3.11–3.13, Hamilton 1.90.0 and SDAX 0.7.2.

Version 0.2.0 covers the documented static decorator families and bounded optional profiles. It does not provide every Hamilton runtime feature: cache/Ray decorators retain metadata only, and dynamic graph execution is outside this scope. See Compatibility for exact limits. Timeouts are cooperative, and resource acquisitions with shutdown cannot use forward retries. These boundaries keep the execution and lifetime promises explicit.