[jira] [Commented] (TINKERPOP-2396) TraverserSet should be extendable for GraphDB provider

["redtree (Jira)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/TINKERPOP-2396?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17176083#comment-17176083 ] 

redtree commented on TINKERPOP-2396:
------------------------------------

Ok I think the idea that a provider should have some strategy to swap the TraverserSet if necessary does make sense to me.
To do so, I think we still need my initial implementation + AbstractStep and ExpandableStepIterator should have some method to replace the logic to spawn a new TraverserSet.

So in my latest pull request, I added *AbstractStep#setTraverserSetSupplier* and *ExpandableStepIterator#setTraverserSet*.

The tests failed because I initially added a pure Supplier which is not serializable. I fixed the issue and now tests passed.
Not sure if TraversalVertexProgram can have the same benefit so I didn't add the same logic there - I am not familiar with TraversalVertexProgram but it is for BigData analysis like Spark right ?
The reason for an internal Id is something OLTP specific (you know, there is a performance benefit by having a fixed length identifier in some Storage engine for instance).

Please let me know if there are any test needed are missing or have any comments on the pull request.

> Can you say which graph database you intend this change for?
You can see my organization in Github...

> TraverserSet should be extendable for GraphDB provider
> ------------------------------------------------------
>
>                 Key: TINKERPOP-2396
>                 URL: https://issues.apache.org/jira/browse/TINKERPOP-2396
>             Project: TinkerPop
>          Issue Type: Improvement
>          Components: process
>            Reporter: redtree
>            Priority: Minor
>
> Currently in many TinkerPop steps, TraverserSet is internally created and used. 
>  However GraphDB provider may want to use their own TraverserSet to change the logic  how it populates an inner Map.
> Just let me give one example
> Say internally we have a different representation of id (which won't be exposed to client when the result is returned) for each element and want to use it for hash key.
> Because we can't overwrite ElementHelper's behavior so first we tried overriding Element#hashCode but after that we saw that this test effectively fails.
>  [https://github.com/apache/tinkerpop/blob/cff4c161615f2b50bda27b6ba523c7f52b833532/gremlin-test/src/main/java/org/apache/tinkerpop/gremlin/structure/util/detached/DetachedVertexTest.java#L75]
>  Here v is our own custom vertex instance which uses internal id for hashCode but DetachedVertex just uses a user facing id copied from v, so they are not considered as equal when used as hash key.
> We may skip this test but it's better to keep all Tinkerpop tests passed as GraphDB provider I think.
> So instead, I will propose to change so that we can add a capability to swap TraverserSet as needed.
>  The change should be simple, 
>  1. ExpandableIterator should have constructor like ExpandableIterator(final Step<S, ?> hostStep, final TraverserSet<S> traverserSet) and use the traverserSet as its own traverserSet value.
>  2. AbstractStep should have construct like AbstractStep(final Traversal.Admin traversal, final Supplier<TraverserSet<S>> traverserSetSupplier) and use traverserSetSuppler.get() for all places that currently we instantiate new TraverserSet.
>  Also this supplier.get() may be passed when AbstractStep instantiates ExpandableIterator.
> So still GraphDB provider needs to extends all Steps that deals with TraverserSet and overwrite their constructor to have their own traverserSetSupplier, but after that they can freely use their own TraverserSet, so the logic how the key of TraverserSet is determined is up to GraphDB provider.
>  Or if we add some helper method like getTraverserSetSupplier under, for instance, Graph class, then instead of extending all steps we can just override that method under our own Graph class. But I am not sure which is better in this case, but either way the goal is the same.
> Any thoughts or objections on this idea ?



--
This message was sent by Atlassian Jira
(v8.3.4#803005)