Ask HN: What does your production machine learning pipeline look like?

Doing some design for an upcoming project and taking a survey. I'll go first. Model training happened nightly on a Spark cluster. This output a PMML-based SVM model. The model was instantiated on a cluster of compute servers running Openscoring. A thin Node web service wrapper used the Openscoring cluster to serve realtime client prediction requests. Dataset size in the hundred millions of examples with hundreds of features. Handled thousands of requests per second, no problem. Separating the training technology from the execution technology was nice but the PMML format is limiting in the kinds of models you can use that both you trainer and executor will support. What are people doing who use same tech for both? For something like Tensorflow, I assume you must have to save the model as binary from the train step and then send it off to the prediction cluster to be instantiated again for execution?