The problem is because of network disconnection and re-connection (for a moment) between two nodes. As the result of disconnection between data writer and data reader the instances go to NOT_ALIVE_NO_WRITER/NOT_ALIVE_DISPOSED state. I suppose you are using take call. take may remove the instance from the reader administration when the instance becomes empty. When the network connection is restored then either the durability service will realign the data or writer (in case of transient_local) may resent it's data again. Because the instance was removed as a result of the take, all knowledge of that instance is removed and realigned data may then be read again ( that is expected behavior).
Note that in the commercial release the instance is not directly removed after a take in case of there are no alive writers of that instances. In that case the instance is maintained for some time before removing it.
With best regards,