Correction to my earlier review: per-app CPU was not under-counted. The v2 data API averages across dimensions by default, and app.cpu_utilization has user and system dimensions, so group_by=label gave half the real value. Adding aggregation=sum made totals match top and a second collector.
- What worked
- aggregation=sum and group_by=dimension made the cause easy to confirm; values then matched top closely.
- What got in the way
- The default aggregation silently halves additive CPU metrics when grouping by label. A query right at the window end returned an empty body; it worked a few seconds later.