An ATM cell may encounter congestion and suffer variable delay due to bufferring within the ATM switches, and may even be dropped either due to congestion control or due to header checksum error. However an ATM connection always obeys causality, the cells in a connection (i.e. cells with the same VCI label) arrive inorder at the destination. This is so because there is no store and forwarding in the network, cells travel over a single virtual circuit path, the ATM switches do not switch the cells in the same VCI out of order, and no retransmissions is done at any point in the ATM network.
Connectionless services are also supported on ATM networks, but these are implemented as a higher layer service layered over the ATM datalink layer. Thus cells in a connectionless service may arrive out-of-order because there might be multiple VCIs over multiple paths setup to deliver the connectionless datagrams and cells may arrive over different paths in different order. Thus the fragmentation reassembly engine which implements the connectionless datagrams, and which is layered on top of the basic connection oriented service of the ATM layer, must carry sequence numbers in the adaptation layer in each cell and correct any reordering of the cells at reassembly time. This is what the IEEE 802.6 protocol for MAN does to support its connectionless service class.
There is no end-to-end reliable delivery service at the ATM layer. The ATM layer does not do any retransmissions and there are no end-to-end acknowledgements for what has been received. Reliable delivery service can be implemented as a layer on top of the basic connection oriented ATM layer, where acknowledgement of received data and retransmission of missing data can be done for connections requiring reliable delivery. Thus a TCP type transport layer protocol (layer 4 in the OSI model) layered on top of the ATM layer is required for guaranteed delivery.