Draft page

This page is an outline. It describes what will be covered and is not yet complete technical documentation.

Not all inference is interactive. Bulk classification, dataset annotation and document processing have no user waiting, which removes the latency constraint that shapes online serving and changes every configuration decision.

This page covers how to configure for pure throughput.

What you will learn

  • How offline and online serving objectives differ.
  • Which configuration choices change when latency does not matter.
  • How to make a long batch job restartable.
  • How to schedule batch work alongside interactive traffic.

This page is an outline. The subsections below are the planned structure; they are filled in as the handbook is written.

Offline versus online serving

Draft

Configuring for throughput

Draft

Input ordering and bucketing

Draft

Checkpointing and restartability

Draft

Sharing capacity with interactive traffic

Draft

Cost considerations

Draft