The handling of output from MapReduce jobs follows the same philosophy. By treating inputs as immutable and avoiding side effects (such as writing to external databases), batch jobs not only achieve good performance but also become much easier to maintain:
