First-class DAG workflow authoring across Java, TypeScript, and Python

[zhongkechen]

Request

Support authoring workflows with DAG (Directed Acyclic Graph) semantics as a first-class construct in the Durable Functions SDKs for Java, TypeScript, and Python.

Context

The Durable Functions SDK currently supports parallel execution through explicit orchestration code, such as stepAsync and parallel branches. However, it has no native declarative DAG abstraction where developers define nodes and their dependencies and the runtime automatically schedules eligible nodes for concurrent execution.

There is currently no standard AWS abstraction available for this pattern.

Why this is necessary

DAGs are a core workflow pattern

Many real-world workflows have complex dependency graphs: tasks that can run in parallel once their upstream dependencies complete. Declarative DAGs are a natural and expressive way to model these workflows.

Imperative code is error-prone for complex dependency graphs

Without a native DAG construct, developers must manually write orchestration logic to determine when tasks can run in parallel and when they must wait for dependencies. As the number of nodes and dependencies grows, this logic becomes increasingly complex, harder to maintain, and prone to scheduling bugs.

The runtime cannot optimize scheduling automatically

With a declarative DAG, the runtime has full visibility into the dependency graph and can schedule all eligible nodes as soon as their dependencies complete. In the imperative model, the developer bears this burden, which can lead to suboptimal execution ordering and underutilized parallelism.

Declarative DAGs improve operational visibility

A DAG construct makes workflow structure explicit and inspectable, enabling dependency-graph visualization, per-node progress tracking, and easier debugging. Imperative parallel logic obscures this structure.

Desired capabilities

• A native declarative DAG construct, for example ctx.dag(...), where nodes and their dependencies are declared and the runtime automatically schedules nodes whose dependencies are satisfied
• Deterministic and replayable dependency resolution, with the checkpoint model efficiently capturing DAG execution state
• Fan-out support for parallel branches with no dependencies
• Fan-in support for barrier synchronization across multiple upstream dependencies
• Idiomatic APIs for Java, TypeScript, and Python