Are you using TensorFlow for your production models as well?
Which target ? Pc or android device
Depends a lot on your infra ... if u r using gcp in ur org you can take advantage of their distributed gpu computation pipelines - beam and others . Else you have to manage that pipeline to distribute your tensorflow models in prod and also monitoring of the gpu cluster . Including tensorflow serving for prediction time
TF in production is easy to scale with containers. Spark in general is kinda buggy even for prototyping, let alone production..
TF is used for Deep Learning and Spark used for simpler models.
Also to add to this point spark is much more suited for data engineering pipelines using batch and streaming pipelines and any model support which can be parallelized and supported by spark ml lib , unless you want to find a approximate version of parallelizing the algorithm .
Numpy or go home Yes, only tf
Deep learning are not available really in spark. You can however use tensor(deep learning) and pipe the spark data to your trained tensorflow model at scale