Some applications have very long (several month) computations that generate large amounts of local state, and therefore are most efficiently done on a single host.
To accommodate such applications, BOINC allows workunits and results to be linked together into sequences. Each workunit and result may have a predecessor and a successor.
If a result A has a predecessor B, the scheduling server will dispatch A only to the host that has handled (or is currently handling) B. If result A times out, then the scheduling server will dispatch B to any eligible host.