We run a number of mission-critical MapReduce jobs daily in our production cluster, mostly on top of HBase. In the past, we've hit a number of Hadoop bugs, and found it difficult to maintain a solid SLA.
We are now moving to CDH5 and evaluating if we should move to YARN or keep running Hadoop 1. YARN is very compelling, but it's also relatively young. I know that Cloudera recommends YARN over Hadoop 1 in CDH5, but I could use a second opinion :)
Can someone running YARN in a mission-critical Production environment share their experience, specifically as it relates to stability? I realize that this is a question that lends itself to a somewhat subjective answer.