Allow defining buckets where each metric & span can be marked

Use cases, pain points, and background

Training frameworks both (pre & post) are complex and evolving, the number of unique metrics and spans for a given training framework can only grow. Calculating training efficiency currently entails knowing whether each span indicates productive or unproductive work.

Description

In order to minimize the complexity in calculating training efficiency it would be very helpful if we allow defining buckets where each of the span falls under, as an abstraction. This allows developers to tag their respective instrumented code to be categorized consistently.

Design

I propose 4 buckets: productive , overhead , idle , wasted to be defined by default and allow users to extend or rename it. Each span has only one of the buckets as an attribute, as mandatory if buckets are defined

Out of scope

Acceptance Criteria

• Centrally define efficiency buckets in a file or config, allow defining a default bucket
• Each function that generates span should allow attributing it with a bucket name
• Optionally allow mapping each span group to a bucket, centrally, so that users dont have to define it for each corresponding span