Service Mesh Platform · View 28 of 31 · 6 · Operations
Decisions
- Mirror before cohort, cohort before percentage. Each step exposes the new version to a larger and less forgiving audience, and each can be skipped by a declared exception, not by habit.
- The step that reads the applied share is part of the loop, not a debugging tool. It is what turns 'we set 1%' into 'the mesh sent 0.98%'.
Targets
- Weight granularity 1%. Revert is the fastest operation in the loop, and is measured on every abort.
Deliberately out
- Traffic mirroring of non-idempotent methods, and header-based cohorts built on user-supplied headers that the edge does not strip.