I have this problem a lot with pandas Dataframes. I know I could engineer around it, but most of the time I just want to distribute some number crunching to a bunch of docker images quickly and a database is overkill for a one off analysis. Works fine until an image updates pandas. Not taking issue with pandas or pickle, but it's an issue of time trade-off. JSON is okay, but object conversion/Nan/None/inf can be a bear.