Huh. Do you not use say EMR? We've got lots of folks running Hadoop/Spark either directly or through Dataproc (https://cloud.google.com/dataproc/).

All of our Monte Carlo finance customers as well as the genomics folks at the Broad Institute seem pretty happy.

If you're saying they're a bad fit for stateful applications like databases and web backends, I agree ;).

We run a bunch of production workloads on them. Just look at the spot pricing history in US-East-1A and 1B for a good idea at the stability.

This is over a year now and only once did I have any sort of interruption and even there it was in only one AZ so of course I had other instances up and able to take the load.

All the test infrastructure can easily go on test if you have auto-scaling. Some percentage of production, you can move to Spot.