In WT-4447 we introduced throttling if an io_capacity=(total=<size>) setting is used. The purpose is to throttle back IO internally, before systems like AWS otherwise delay IO when provisioned IOPS are exceeded. If AWS does it, it can often lead to long stalls in the application.

donald.anderson and I arrived at different capacity settings on the same machine, for two workloads. When he used total=30M he saw great results and when I used total=35M I saw great results. So our results were somewhat opposite.

Given that, it may be desirable for the throttling algorithm to detect when more/fewer stalls are happening and auto-adjust the capacity to see if we can find a sweet spot for the workload. The first and most important piece is figuring out what statistic or indicator can tell us we have stalls.