Asynchronous schema enforcement, that's really interesting. How has it worked out?

Also I'm not sure I get the dropping part. Do you really throw out most of your data after a short time window? Is it not of trending/historical-analysis value?

Yes... pretty well - caught a few issues before they went out.

Dropping the data - this is just the raw unfiltered stream. We save the interesting process data elsewhere and then ditch the raw events.