I went to a Map-Reduce tutorial at HLT-NAACL last year (one of the most respected natural language processing conferences). One of their main take home points?

I went to a Map-Reduce tutorial at HLT-NAACL last year (one of the most respected natural language processing conferences). One of their main take home points? For projects where results matter (e.g. science), don't use any homebrew Map-Reduce implementations. According to their evaluations of alternate implementations, it's just not worth the risk.