However, reducing the number of replicas is unsafe. This only aggravates the inherent limitation of data replication: its failure resilience guarantees are fundamentally limited by the number of replicas. Having insufficient replicas leaves computations exposed to failures and this can severely impact performance. The reason is that without the use of data replication, failures can easily cause data loss which can trigger cascading job recomputations: several jobs need to be recomputed to regenerate the lost data. In the worst case, the recomputation may have to revert all the way to the beginningof the multi-jobcomputation.This suggests the need for devising efficient approaches to job recomputation.