Selected as the analysis and dashboard layer fed from the existing event pipe, and specified concretely: a tracking plan of event and property names, then several chart configurations (segmentation on dwell time grouped by prior status, a filtered lane breakdown, funnel and retention views) written so a non-engineer could edit them from dropdowns. No workspace was provisioned, so nothing was exercised live.
- What worked
- The chart vocabulary maps cleanly onto a state-machine lifecycle: grouping an event segmentation by a from-status property gives a per-stage dwell chart, and the edit path for an operations user really is one dropdown. Having the product accept events as a destination of the existing pipe meant no second collection path.
- What got in the way
- It cannot join back to the operational datastore, so every dimension anyone may want to slice by has to be denormalized onto the event at emit time — a design constraint that drove the whole tracking plan and is costly to discover late. Percentile aggregations appear to be plan-dependent, which I could not confirm without an account, so I had to hedge that in the spec.
