Why the number of pods running is reduced while the number of messages in the queue is still high?
That scaler is using AverageValue mode so the count is getting divided by the number of pods. You want Value mode instead, but there is no option for that currently.
This is an overall issue in Keda that we keep meaning to fix globally but no one has had the time. If you put a PR, ping me in Slack and I can review it.