Why can't MapReduce put the output into in-memory?

i hava already knew that MapReduce output data to disk during the shuffle and after reducer, which is the reason why mapreduce is slower than Spark. I know it was designed this way in the early days because computers didn't have much memory and disks were cheap, But why not to change this design nowadays as the in-memory is bigger? Why not to change the source code to save output into in-memory like Spark?