Hi Benoit,
yes, it does. Downscaling it at least as important as upscaling. When nodes scale down cleanly, they report the fact that they leave, and all nodes instantaneously take over the leaving node's portion of command handling.
When a node leaves without prior notice, there is an amount of time where the node is "suspect". That typically takes 5 seconds. During that time, some commands may fail to properly find their destination. Senders of these commands will get an exception and have the option to retry. Note that senders also get an exception if the node goes down after a command was sent, but before a result was reported. This exception states that the outcome of the command is unknown.
Cheers,
Allard