Our rolling upgrade tests were failing on many machines due to using the 5.x default of 4GB model_memory_limit, which then propagated forward to 6.1+ even though the default is now lower. Original commit: elastic/x-pack-elasticsearch@3b23d8fe9d