I am struggling with this since months.
I have ca. 30 map services, serving data with identical schema, they just differ with the content. The data is being stored in cloud geodatabase. The services were published using ArcGIS Pro indentical maps. Brilliant example for comparing service performance. There are about 12 services with high traffic and they are configured with dedicated instances with minimum of 1. The rest of them are very low, infrequent traffic which I have configured with 8 shared instances.
All of them exprience "cold start" issue but the first 12, after a warm up, perform very well with response times below 1 sec. Others perform very bad with average response times reaching 20 - 30 seconds. I was hoping shared instances will help. But this is not the case.
IT guys say it is not a problem with the resources. That seems to be true: after warming up the services perform very well as long they are in use.
What the heck happens when inactive service is asked for the response from the client? Why do we exprience this issue? I am out of ideas. The geodatabase performs very well while accessing from ArcGIS Pro.