Failure handling guarantees

A dead letter queue is useless if the error handling itself can make the job fail. Asgarde guarantees the following behaviors, each covered by tests in both libraries.

Case	Behavior	Java	Python
Exception that can’t be serialized (holding a client, a lock…)	Replaced by a serializable copy keeping the type, the message and the stack trace	SerializableThrowable	SerializableException
Input element whose conversion to string fails (null, failing toString/__str__, dict with non JSON types)	Regular failure, never a crash	✅	✅
Memory errors	Re-raised to the runner, not sent to the failures: the worker is unstable	VirtualMachineError (OutOfMemoryError, StackOverflowError)	MemoryError
FlatMap whose iterable fails in the middle	Failure only: the outputs are materialized first, no partial outputs duplicated when the failure is replayed	✅	✅
Same DoFn instance applied in several steps	Each failure keeps the name of its own step	✅	✅
Transform names	Failures flattened once in a deterministic Get all failures of <last step name> transform	✅	✅

Dataflow streaming updates

Updating a running streaming job (--update) requires stable transform names. The failures transform has a deterministic name, and getting the failures several times doesn’t apply it twice.