Rolling back at the beginning of every request cause huge overhead

In reference of http://tinkerpop.apache.org/docs/current/reference/#transactions and https://github.com/JanusGraph/janusgraph/issues/500.

Adding a `rollback()` at the beginning of every request adds a large overhead of about 30s for retrieving only 4000~ vertices with indices enabled. Removing `rollback()` speeds up all requests to be in the range of milliseconds. However, removing `rollback()` causes inconsistencies with deleting edges.

We follow https://docs.janusgraph.org/latest/tx.html pretty closely for adding and updating vertices and edges and haven't encountered inconsistency issues without `rollback()`.

Are there no way to refresh the graph cache when deleting objects? Can there be more docs regarding this issue?

The issue names no source file or test. Start by reviewing the referenced TinkerPop transaction documentation and JanusGraph issue #500, then reproduce the rollback overhead and deletion inconsistency. Done means establishing supported transaction behavior and documenting or resolving how requests can avoid the costly rollback safely.

Written by the indexing model from the issue text.