A single global latency chart hides regional problems in the average. Build one dashboard that makes region-by-region health obvious at a glance.

A service running in four regions with one region silently degraded looks completely healthy on a dashboard that only shows the global average, the three healthy regions pull the number back down into a range nobody would flag. Splitting every panel by region fixes that, and it costs one extra field in the group-by clause.

1. Add region to every group-by, not just one panel

The habit that actually matters here is discipline, not cleverness: every panel on the dashboard, not just the headline one, needs the same regional breakdown, or the exception becomes the blind spot during the next incident.

from traces
| where service.name == "checkout-api"
| summarize p95(duration) by bin(timestamp, 1m), region

2. Add a region-comparison panel, not just per-region lines

A chart with four overlapping lines is readable when they track together and unreadable the moment one diverges. A dedicated panel showing the spread between the best and worst region makes the divergence itself the metric, easier to alert on than eyeballing four lines.

from traces
| summarize p95(duration) by region, bin(timestamp, 1m)
| summarize spread = max(p95_duration) - min(p95_duration) by bin(timestamp, 1m)

3. Account for the follow-the-sun traffic shape itself

Regional traffic volume rises and falls on its own daily cycle as usage follows daylight hours. A raw request-count panel per region will look like a false alarm every single day at the same local time unless the on-call rotation for that region already expects the shape and reads past it.

from traces
| summarize count() by region, bin(timestamp, 1h)
| sort by timestamp asc

4. Hand the dashboard off with the rotation

For a follow-the-sun on-call model, the dashboard itself should be part of the handoff, not just a link. The outgoing region calls out anything currently sitting at the edge of normal before the next region picks up, using the exact panels above as the shared vocabulary for that conversation.

  • Break every panel down by region, not only the headline latency chart.
  • Add a spread panel that turns divergence between regions into a single alertable number.
  • Expect and read past the natural daily traffic cycle per region before calling it an anomaly.
  • Use the dashboard itself as the shared language during a follow-the-sun handoff.