Nice. Thanks.
Do you think sampling the stack traces of millions of machines is a good dataset for improving code performance? Maybe sample android/jvm bytecode.
Maybe a sort of novelty sampling to avoid re-sampling hot-path?