(In reply to Matej Cepl from comment #2) > Could you please *ATTACH* (not link) build log (preferably on x86_64) when > it fails, please? I am not sure what that is good for, but I attached it anyway. Even if it does not fail, it still consumes 8G of memory. So be it failed or successful build, the leak is still there... More interesting would be to run it locally and measure python memory footprint with 3.6/3.7 and 3.8. Maybe with some valgrind-for-python thing, if there is any...