Marko A. Rodriguez created TINKERPOP-1140:
---------------------------------------------
Summary: TraversalVertexProgramStep in support of OLAP/OLTP
conversions.
Key: TINKERPOP-1140
URL: https://issues.apache.org/jira/browse/TINKERPOP-1140
Project: TinkerPop
Issue Type: Improvement
Components: process
Affects Versions: 3.1.0-incubating, 3.1.1-incubating, 3.1.2-incubating
Reporter: Marko A. Rodriguez
Fix For: 3.2.0-incubating
This was moved from TINKERPOP-971. TINKERPOP-971 was about TraversalSource
fluency. This is a feature we can now do because of that work. Isolated this
feature in a new ticket.
--------------
Check this idea out:
{code}
g = graph.traversal().withComputer(graph ->
graph.compute(SparkGraphComputer.class).workers(10));
g.V().out().values('name')
{code}
This would compile to:
{code}
[TraversalVertexProgramStep([GraphStep,VerticesStep,PropertiesStep]),ComputerResultStep]
{code}
where,
{code}
TraversalVertexProgramStep<Graph,ComputerResult>
ComputerResultStep<ComputerResult,E>
{code}
Next, check this:
{code}
g = graph.traversal().withComputer(graph ->
graph.compute(SparkGraphComputer.class).workers(10));
g.V().hasLabel('person').pageRank(out('knows')).order().by('page.rank',decr)
{code}
This will compile to:
{code}
[TraversalVertexProgramStep([GraphStep,HasStep]),PageRankVertexProgramStep([VerticesStep]),TraversalVertexProgramStep([OrderStep]),ComputerResultStep]
{code}
How does this work?!
* TraversalVertexProgramStep will pass a {{ComputerResult}} to
PageRankVertexProgram.
** The ComputerResult.graph() will have HALTED_TRAVERSERS properties on all the
person vertices (as that is what was computed by the vertex program).
* PageRankVertexProgram will then pass a {{ComputerResult}} to the next
TraversalVertexProgram.
** That ComputerResult.graph() will have HALTED_TRAVERSER on all the person
vertices and PAGE_RANK on all the vertices.
* TraversalVertexProgram will then execute OrderStep which sorts the
person-vertices with HALTED_TRAVERSERS on them by their page.rank properties
computed previously.
* ComputerResultStep will then take get the ComputerResult.memory.reducing and
iterate it out.
{{ComputerResultStep}} (like now) is smart about either pulling from the graph
or from the sideEffect memory. What makes things different from now is that
{{TraversalVertexProgramStep}} will come into existence and pull out the
respective logic from {{ComputerResultStep}}. In this way, It will be possible
to chain {{{XXXVertexProgramStep}}-steps and thus, have a traversal that is
multiple OLAP jobs in sequence and thus, this ticket will fall naturally from
this https://issues.apache.org/jira/browse/TINKERPOP-570. The whole trick to
all of this is that (as we currently do) we save the state of the computation
in the graph (and memory) and thus, feeding program into the next just takes
over the computation. PageRankVertexProgram doesn't use HALTED_TRAVERSERS in
its computation, so it just ignores it. However, TraversalVertexProgram can
later access PAGE_RANK and thus, use the results of the previous OLAP
computation.
Finally, you want "lambda vertex programs?"
{code}
g = graph.traversal().withComputer(graph ->
graph.compute(SparkGraphComputer.class).workers(10));
myProgram = MyVertexProgram.build().create();
g.V().program(myProgram).values('myProgramCounters')
{code}
which compiles to:
{code}
[TraversalVertexProgramStep([GraphStep]),LambdaVertexProgramStep(MyVertexProgram),TraversalVertexProgramStep([PropertiesStep]),ComputerResultStep]
{code}
Thus, just like {{filter()}},{{flatMap()}}, etc....
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
