Going back to the "federated" aspect of TinkerPop, I'm not sure that there's many on this list with much familiarity with Neo4j. I believe that Neo4j falls into a "Third Party" category and that that organization, or others more closely associated with it, provide the TinkerPop integration tools. There's a rather extensive list of these providers here:
http://tinkerpop.apache.org/providers.html. I think that they all fall into this category.
I have dabbled some with Docker and TinkerPop, in conjunction with building Gremlin training programs. Sometimes I have found that Docker causes as many problems as it solves, especially when it comes to managing the networking or dealing with high-IO images. But you might find this image helpful:
https://github.com/experoinc/gremlin-lang-intro. I believe that the README and Dockerfile are clear enough, though perhaps too wordy.
It sounds like you might be using Docker's volume capability to attempt to share state between the container and the OS. I have dabbled with this a little on a MacOS (a somewhat well-provisioned MacBook Pro). For my experience, the containers worked ok with ActiveMQ and ElasticSearch (because our use of those engines is very light), but when I tried to do the same with PostgreSQL and DSE Graph, attempting to map all data persistence directly to the underlying SSD, the containers could not handle the GBs of data I was trying to run through them. For both PostgreSQL and DSE Graph I had to switch to some sort of binary run directly by the OS. I did find that the generic Docker on Mac doesn't handle high-IO situations well on MacOS, though there's some command-line Docker approach which might work. Since the binaries work well enough, I've not had the need (or the leisure) to try anything else.
That might have nothing to do with the challenges you're facing in your setup. But as your approach seems unique, it is going to be difficult for any on this list to explain why it doesn't work without any error statement or more detailed description of the failure state.
For the training I'm planning in January, where I do plan to support the Python GLV, I'm planning to use a TinkerPop Gremlin Server binary and we will build a Python script which will interact with it through a network connection, or the "
Python Remote Connection" approach. It's my expectation that most folks will interact with their TinkerPop-compatible server in this type of client-server approach.
-Josh