I noticed that I was reproducibly getting incorrect results for some tests in test_mem_usage_scaling.py. E.g. q4.
I was able to bisect the problem to the commit 61fcb489745f3f0b3f1abbf9fbf666a29a6363de "
IMPALA-4300: Speed up BloomFilter::Or with SIMD"
I'm not sure why we're not seeing on the jenkins machines - maybe it's related to the CPU I'm running on.