What I often do when I analyze logs is removing timestamps and changing unique identifiers to something more predictable and diff them to see when things started diverting from the norm, because it is often earlier than the usual error/crash. Is there anything that does something like it?