Tempest test on prometheus fails on fake metrics because there are ceilometer_cpu metrics both faked and from real ceilometer backends, and the datasource driver is taking data just from the real one instead of aggregatin both. That leads to getting metrics near to 0:

Jun 10 09:56:54.117054 np0041073325 watcher-decision-engine[100130]: DEBUG watcher.decision_engine.strategy.strategies.workload_balance [None req-984e8be0-e22a-43f9-893b-524133ea4525 None None] Host usage for (c74a714c-5d1d-4402-ae19-5b32ad636414): host_cpu_usage_percent is 0.684483, lower than threshold 10.000000 (pid=100130) group_hosts_by_cpu_or_ram_util /opt/stack/watcher/watcher/decision_engine/strategy/strategies/workload_balance.py:286

Jun 10 08:20:10.622963 np0041072552 watcher-decision-engine[100094]: DEBUG watcher.decision_engine.strategy.strategies.workload_balance [None req-8739e54d-61ca-402e-acbf-2bc13745bc26 None None] Host usage for (7604851e-eca6-4406-a5fe-f5dd8b114e83): host_cpu_usage_percent is 0.782753, lower than threshold 10.000000 (pid=100094) group_hosts_by_cpu_or_ram_util /opt/stack/watcher/watcher/decision_engine/strategy/strategies/workload_balance.py:286

This is because the function:

https://github.com/openstack/watcher/blob/15981117ee28627f235264e505e1e0d5956cf4e4/watcher/decision_engine/datasources/prometheus.py#L345

is aggregating by instance instead of resource.