Java SDK: Serialize native Dags to DagSerialization v3

[jason810496]

• depends on: Java SDK: Group a native Dag's tasks with task groups #74230
• Diff for early review: jason810496/airflow@feature/java-sdk-task-group...feature/java-sdk-dag-serialization

Merge order:

1. Java SDK: Group a native Dag's tasks with task groups #74230 — Group a native Dag's tasks with task groups
2. Java SDK: Serialize native Dags to DagSerialization v3 #71190 — Serialize native Dags to DagSerialization v3 (current one)
3. Java SDK: Pack each native Dag's source file into the bundle JAR #74096 — Pack each native Dag's source file into the bundle JAR

Why

Make the runtime can answer the coordinator's DagFileParseRequest itself.

How

• parseDags serializes every Dag registered on the Bundle to DagSerialization v3.
• Server.dispatchTask gains a DagFileParseRequest branch alongside StartupDetails, so a bundle process serves either a task run or a parse request.
• A field the Dag leaves unset stays out, including the ones Python reads from [core] config, such as max_active_tasks and catchup. Airflow fills those in from its own config when it receives the Dag.
• Task groups serialize into the nested task_group object Python writes, with each group carrying its own edges.
• Every task carries is_stub, and one the Dag called with arguments also carries the _arg_bindings spec ADR-0007 defines, so the serialized Dag records what each task is called with.
• A Java task still reads its arguments from the Dag in its own bundle rather than from that spec: the bundle that runs the task also holds the call that wired it, so nothing has to travel at run time.
• A cron preset expands before it is written, and the timezone comes from the Dag's start date, both as Python does. A schedule that is neither a preset nor a cron expression is rejected where the Dag is written.
• A Dag that cannot be serialized becomes an import error keyed by the bundle-relative path, so the other Dags in the same JAR still parse.

Example

load(transform(extracted, lit(1.5))) in a wiring class serializes as:

"_arg_bindings": [
  {"name": "extracted", "kind": "xcom", "task_id": "extract"},
  {"name": "factor", "kind": "literal", "value": 1.5}
]

The spec travels as JSON, so lit(...) takes a string, number, boolean, list or map; anything else is rejected when the Dag is parsed.

Cross-language validation

check-java-sdk-serialization-conformance builds the shared test Dags of scripts/ci/lang_sdk_serialization/test_dags.yaml with both this SDK and Airflow's own serializer, loads the Java output through DagSerialization.validate_schema and from_dict, and compares the two field by field.

Known limitation

Cron schedules map to CronTriggerTimetable only, as the TypeScript SDK does, until the supervisor forwards the [scheduler] timetable flags over the coordinator protocol. There is a TODO at the timetable serializer.

---

Was generative AI tooling used to co-author this PR?

• Yes, with help of Claude Code Opus 5 following the guidelines

[henry3260]

Just want to help unblock the technical side here, not a formal review.

[jason810496]

Thanks for the review.