Make incremental scaling faster
I run a cloud service that handles resource-intensive processing of data. Data arrives from users. Processing each chunk of data takes about ten core-seconds. Sometimes as little as five hundred chunks per hour arrive. Sometimes several thousand chunks arrive is just several minutes. One thousand chunks takes 2,8 core-hours to process. This is why my service contains auto-scaling logic. If it sees that load has notably increased it issues a "Change Deployment Configuration" operation to scale the service out. It may so happen that users continue sending even more chunks of data and the service issues another "Change Deployment Configuration" operation before the first one completes. If this happens then "preemption" occurs - the earlier operation formally fails and the newer operation runs for very long time - sometimes up to 40 minutes. This causes new instances that should be started by the second operation to start unbelievably late - 30-50 minutes after the operation was issued.
That's too slow scaling out. It should be faster.