Do not read an Application Map filter percentile as request latency

Source facts

Application Map’s average-duration connector filter measures the mean duration of calls across each connector. A P90 label on a suggested filter value describes the distribution of connectors, irrespective of how many calls each represents. Microsoft’s 200-millisecond example means that 90 percent of connectors have an average duration below that value, not that 90 percent of all requests completed within it. Microsoft Learn.

Map filters also apply in sequence from left to right. Once an earlier filter removes a node or connector, a later filter cannot bring it back. Microsoft Learn.

Applicability

Review Azure Monitor Application Map connector filters for distributed applications. Keep the map-selection statistic separate from a request-level latency objective or an application’s end-to-end response time.

DSE recommendation

DSE recommends labeling a saved investigation view with the connector statistic being selected. Compare low-volume and high-volume connectors before treating the percentile label as a workload-wide result. Use the filtered map to choose where to investigate, then examine the relevant request and dependency evidence. Preserve the filter order so another operator can reproduce why a connection was visible or absent.

Verification

Choose two connectors with different call volumes and compare their displayed averages with the selected filter boundary. Review the underlying calls for the actual latency question. Remove filters in a controlled comparison to identify excluded paths. Do not report a request percentile merely by copying the P90 label from a map filter.

Official references

Microsoft Learn: Application Map.