Sometimes a whole cluster stays on, while it obviously should be off.
Not sure what exactly causes this, but its probably some multi-threading issue. While I thought it is thread-safe.
When the cluster gets updated the next time, it fixes itself.
Difficult to reproduce.