Roadmap

The Java and Python libraries evolve together: each feature is released in both, with the same concepts. Feedback and contributions are welcome: open an issue on Asgarde Java or Asgarde Python.

1.4.0 (released)

Representation of the input element, without breaking change: the string for reading and searching, the bytes for replaying exactly.

Feature	Java / Kotlin	Python
Input element format: a function converting the input element to a string (JSON, masking sensitive data, identifier)	✅ 1.4.0 (Failure)	✅ 1.4.0
Encoded elements (opt-in): the input and origin elements encoded with the coder of their step, to replay them exactly	✅ 1.4.0 (Encoded elements)	✅ 1.4.0

1.3.0 (released)

Enrichment of the Failure model and integration with the Beam native error handling, without breaking change.

Feature	Java / Kotlin	Python
Richer Failure: exception type, message and stack trace computed from the exception, and a timestamp	✅ 1.3.0 (Failure)	✅ 1.3.0
Failures to BigQuery: conversion to Beam Rows (Java) or dicts (Python) with a documented schema, and a BigQuery recipe	✅ 1.3.0 (Write the failures)	✅ 1.3.0
Beam native error handling: bridge to the Beam ErrorHandler (BadRecord), and any Beam transform in the Python composer (with_exception_handling)	✅ 1.3.0 (Beam native error handling)	✅ 1.3.0

Why no breaking change on Failure

Every enrichment is additive: computed getters, nullable fields and a fixed serialVersionUID keep the failures of a previous version readable, including in flight during a Dataflow streaming update. The schema for BigQuery is an explicit conversion, not a default schema on Failure: a default schema would change its coder and lose the exception.

2.0.0 (later)

A major version for the platform, not for the Failure model:

• Java 11 bytecode, when the minimum supported Beam version is raised to 2.74.0 or later (Beam requires Java 11 since this version).
• Kotlin 2 for the extensions, whose metadata can’t be read by Kotlin 1.x projects.

Ideas, on demand

Not planned: they’ll be considered if users need them.

• Custom Failure type (Java): inject a Failure factory once for all the steps. It needs a generic composer, a breaking and complex change.
• Ready-to-use failure sinks module (BigQuery, GCS, Pub/Sub): a sink module would bring the GCP dependencies; the 1.3.0 conversion plus a one line recipe with the Beam IOs covers most needs.

Delivered

Version	Highlights
1.4.0	Input element format and encoded elements
1.3.0	Richer Failure, write the failures to BigQuery, Beam native error handling
1.2.0	Origin element in the failures (lazy evaluation)
1.1.0 (Java)	Failure handling guarantees and failure metrics
1.0.0	Not tied to a Beam version anymore: Beam is provided (Java) or a minimum version (Python), tested every week against the latest Beam release. The Python 1.0.0 also brings the guarantees and the metrics