Get results where you need them

Three ways to collect a result

Every finished job produces the same bundle. What differs is how it reaches you.

Download it directly when a person is collecting it. Pull it on your own schedule when your system decides when to read. Register a destination when results should be written into a store you already run.

Completion and delivery are different events

A job reaching SUCCEEDED means Compute finished the work and the bundle is terminal. It says nothing about whether bytes arrived anywhere else.

Delivery has its own states, its own failures and its own evidence. A destination outage is a delivery problem, and it never means the run needs doing again.

Pick your route

• Download and verify a result — collect the bundle yourself and check what you got.
• Read partial results with cursors — pull output on your own schedule while a job runs.
• Register a destination — plan, register, test and read back a store Compute writes to.

Write into a destination safely

• Write immutable commits — create-new or versioned writes, and how a conflicting replay is caught.
• Delivery states and acknowledgements — every state, and the next action for each failure.
• Shape your output partitions — file sizes, counts, compression and partition cardinality.

When delivery fails

• Retry delivery without rerunning the job — reauthorize the destination and resume the accepted bundle you already paid for.