I see a slippery slope in ML engineering, where to make something “production ready” engineers layer a relatively simple workflow in about 8 different technologies. I sometimes wonder if it’s preferable to just deal with occasional issues as they occur than to anticipate every deployment bug under the Sun.