Unfortunately we came across this same issue. We were trying to use MLFlow Serve to produce an API that could take text input and pass it through some NLP. In this instance we had installed a maven package on the cluster, so the experiment would run fine in a notebook, but MLFlow would fail as it couldn't install the maven package. As an alternative, it would help to be able to modify the job cluster that is provisioned to add additional libraries/packages that are required, that we can not specify in the conda definition.