Secondary members can "self-heal" after being disconnected. Up to a configurable number of missed transactions, members download data until they catch up. This threshold defaults to 10,000 transactions, above which the member requires resyncing.
Unlike earlier versions of PAM, r3.3 and newer is using mySQL asynchronous group replication (with single primary) mechanism which guarantees delivery.
Once a member (re)joins the cluster it will receive all previous updates while it was unavailable.
Note, due to the new method in 3.3 it is highly recommended to have at least three nodes configured in the primary Cluster Site.
(else the possibility to run into quorum loss mode is given which then basically renders the complete PAM Cluster being unavailable)
For further details about mySQL asynchronous group replication please see its documentation