Does anyone knows how accurate numbers on Stackdriver Monitoring GLB dashboard represents 95th percentile of Total Latency split by client country, and do they represent actual total latency that takes end user to receive response from request, especially located in different continents.
If not, what would be the best way to get real numbers and understand what region requires service deployment for improving latency
Since percentile of latency is based on statistics collected by Stackdriver at request of the user, it is at the user discretion how to influence the accuracy.
As with many statistic quantities, those depend on the volume of collected data. If production traffic is not high enough, you can't expect high accuracy. Under certain conditions, traffic to the application server could be generated to get enough data for statistically valid calculations.
Next, for accurate presentation of data user can tune bucket size.
There are some factors out of Stackdriver control that could limit data accuracy such as accuracy of event reporting provided by kernel.
Stackdriver just applies well established statistical methods to the collected data.
Please see:
Monitoring metrics for external HTTP(S) load balancer
