When a timer process puts a host into maintenance, it updates both configuration cache and database. However, when it takes a host out of maintenance, it only updates the database, so the host remains in no data maintenance from the perspective of configuration cache. So, technically, data collection stops on that host. What saves us, is that host's maintenance status is set back to HOST_MAINTENANCE_STATUS_OFF when configuration cache is synced again and configuration cache update frequency is probably set to a relatively low value in our users' configuration files, so that is probably the reason issue has gone unnoticed until now.