12 May 2026
Lag windows are not latency
In clinic we keep meeting tiles titled “live” because a broker’s p99 is 12 milliseconds. That number is almost never the number a product lead is looking at.
Latency, in the vendor brochure, is how quickly a packet or a record can move through a hop you do not own. A lag window is a policy: from which event time is this tile allowed to be quoted as complete? Those are different sentences. Mixing them is how a morning GMV chart “wakes up” for an hour after the market already opened.
Write three clocks on the same scrap of paper. Event time: when the client thinks something happened. Ingest time: when your collector accepted it. Decision time: when a human is allowed to treat the tile as closed for a meeting. Real-time app KPI tracking starts to work when the third clock is explicit. Until then, “live” is a mood.
We ask seats to publish a histogram of ingest lag for one hero metric, then to pick a window they will actually defend. Forty-one milliseconds appeared in our teaching fixture because that is what the sandbox produced that week, not because it is a universal target. Your window might be eight minutes because mobile clients buffer. That can still be a live system. It is simply an honest one.
If you need a rule of thumb from the KPI Grid Clinic: never put the vendor’s latency in the tile subtitle. Put the watermark. Put the window. Put the name of the person who will change it.