Thermos Observer Takes up to 150% CPU on default settings

The Thermos Observer when default configured with 5 second polling interval takes up to 150% cpu on an aws c3.4xlarge. This is with

~10k executors
~1 run per executor

Haven't done any profiling, but the overall design seems to be inefficient given it has to list read the entire directory tree every time and do a fair amount of regex parsing.