[correction] The speed measurement of 20s is for a X86 Docker image with the platform field forced to amd64. For a native silicon image, the speed improvement is impressive, I am down to 4s.