The current "tranaction manager" in Gremlin Server is just like Rexster's and it's not very smart. It makes no distinction about what graphs were actually affected when it issues its auto-commits/rollbacks at the end of a sessionless request. For those with a number of different Graph instances configured in Gremlin Server, that's a lot of extra empty commits if the intent is to just mutate a single graph in the set. I'm not sure what that time amounts to, but it seems sensible that if we could only commit when needed then it would be better than lots of extra commits for nothing.
As it so happens we have the rebindings feature (wonder why didn't call that "alias") in Gremlin Server: http://tinkerpop.incubator.apache.org/docs/3.0.1-incubating/#_rebinding So we could use that to tell the transaction manager which graph the user is working with and thus allow the correct commit on the right graph. If no rebindings are supplied, I guess we could stick with the current model. It is a potentially breaking change to people already using rebindings only in the sense that if they have two graphs in Gremlin Server, then chose to use a rebind for one, but then issued scripts that directly referenced the other, obviously, under this new model, the transaction would be uncommitted on the direct reference - they would have to alter the code to rebind both. Generally speaking, I think we want to encourage rebinding as a pattern as it makes code more flexible as your scripts don't get bound to the name of the graph in Gremlin Server - they can be bound to a more general alias. So this change would perhaps lead folks in this direction. Anyone against using rebindings in that fashion?
