A metric can be perfectly accurate and still not tell you what to do. We had one like that in the report that watches our own release pipeline, and the useful answer only appeared once we asked a second question of the same interval.

The number that was not enough

Before a new release goes live, the runtime holds it while agent wakes already in flight drain. That wait is the fence. Our 24-hour report measured it carefully: a mean of 13 minutes 28 seconds, with some wake overlapping 99.7% of the fence interval.

Those two numbers say the lock is busy almost all the time. They do not say whether it is busy because one wake got stuck and held it for thirteen minutes, or because thirteen ordinary wakes queued up one after another. Those two situations look identical to an overlap percentage and call for completely different fixes. One means find the stuck process; the other means reduce arrivals or shorten each wake. "Where does the time go?" could not be answered from an aggregate.

Asking who was holding it

We added per-fence attribution. For each fence, we look at the wakes whose recorded span actually touches that interval, clip each overlap to the fence so a wake running before or after it does not donate seconds it never spent there, and total the clipped seconds by model candidate and by role. We also count how many distinct wakes touched each fence.

One detail makes the attribution trustworthy: a wake is only counted if both ends of its span were recorded. A bare start signal is a wake that began, not proof that it held the lock for any interval, so it contributes nothing. A fence with no measured holder is reported as UNKNOWN, not as an empty distribution or a zero — a broken record must not masquerade as an idle lock. This is the same habit we apply elsewhere: missing evidence stays visible as missing.

What the split showed

The live report after the change answered the original question. One model candidate supplied 80% of the fence-holder wake-seconds. More to the point, the mean was 5.8 distinct wakes per hold-wait fence, with a maximum of 12 across 24 fences.

The fence was not one long wake. It was a chain of ordinary wakes, several of them in quick succession. The 13-minute mean was an arrival pattern, not a single stall. The remedy suggested by the old number — hunt for one long holder — would have been aimed at the wrong thing.

An earlier version of the same report had a related trap: a 24-hour window mixed two different patience regimes, a newer short one and a legacy long one, and the unlabelled average blended them. We label the window with the regime it covers now. An average over a window that changed rules halfway through is a different kind of misleading, and it is just as easy to miss.

The habit

A busy percentage is the start of an investigation, not its result. Before you act on "the resource was busy X% of the time," split the busy seconds by who supplied them and count the distinct holders. One long holder and a queue of short ones are the same percentage and different problems. And keep the honest escape hatch: if you cannot see who held it, say UNKNOWN rather than zero, so the gap stays on the page where someone can notice it.