Lifecycle of a Batch Job's Output: All-or-Nothing Publication Diagram

All-or-Nothing Output

Job success means every task ran exactly once, even when some tasks failed and were retried
If the whole job fails, no output is produced at all
Partial shards stay hidden from queries while tasks are still writing

Read-Committed Reveal

Consumers keep the data they receive invisible to queries
A job-complete notification reveals the whole dataset in one step
The reveal mirrors committing a transaction at read-committed isolation

Recovery and Time Travel

A failed task's partial output is deleted and the task is retried from immutable input
Superseded outputs are kept for rollback and time travel
An aborted run's output is deleted or never announced downstream