[
https://issues.apache.org/jira/browse/TINKERPOP-1254?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15308248#comment-15308248
]
Ted Wilmes commented on TINKERPOP-1254:
---------------------------------------
Not sure if this makes sense, but I'm in the thick of this right now and taking
an approach where I build a dependency tree for the match traversals.
Contrived example here to illustrate:
{code}
match(
__.as('a').out().as('b'), (1)
__.as('b').out().as('c'), (2)
__.as('c').out().as('d'), (3)
__.as('b').out().as('e') (4)
).select('d', 'e').dropLabels('d', 'e')
{code}
Traversal (2) and (4) depend on (1), Traversal (3) depends on (2). The
results of (3) and (4) are referenced in the {{select}}. This info is used to
insert the following dropLabel steps:
{code}
match(
__.as('a').out().as('b'),
__.as('b').dropLabels('b').out().as('c'),
__.as('c').dropLabels('c').out().as('d')
__.as('b').dropLabels('b').out().as('e')
).select('d', 'e').dropLabels('d', 'e')
{code}
My thinking is even though the order execution is determined at runtime, we can
figure out beforehand which match traversals depend on each other and insert
the drop steps accordingly. Does this make sense or have I made some incorrect
assumptions on dependencies between match traversals?
> Support dropping traverser path information when it is no longer needed.
> ------------------------------------------------------------------------
>
> Key: TINKERPOP-1254
> URL: https://issues.apache.org/jira/browse/TINKERPOP-1254
> Project: TinkerPop
> Issue Type: Improvement
> Components: process
> Affects Versions: 3.1.1-incubating
> Reporter: Marko A. Rodriguez
> Assignee: Ted Wilmes
>
> The most expensive traversals (especially in OLAP) are those that can not be
> "bulked." There are various reasons why two traversers at the same object can
> not be bulked, but the primary reason is {{PATH}} or {{LABELED_PATH}}. That
> is, when the history of the traverser is required, the probability of two
> traversers having the same history is low.
> A key to making traversals more efficient is to do as a much as possible to
> remove historic information from a traverser so it can get bulked. How does
> one do this?
> {code}
> g.V.as('a').out().as('b').out().where(neq('a').and().neq('b')).both().name
> {code}
> The {{LABELED_PATH}} of "a" and "b" are required up to the {{where()}} and at
> which point, at {{both()}}, they are no longer required. It would be smart to
> support:
> {code}
> traverser.dropLabels(Set<String>)
> traverser.dropPath()
> {code}
> We would then, via a {{TraversalOptimizationStrategy}} insert a step between
> {{where()}} and {{both()}} called {{PathPruneStep}} which would be a
> {{SideEffectStep}}. The strategy would know which labels were no longer
> needed (via forward lookahead) and then do:
> {code}
> public class PathPruneStep {
> final Set<String> dropLabels = ...
> final boolean dropPath = ...
> public void sideEffect(final Traverser<S> traverser) {
> final Traverser<S> start = this.starts.next();
> if(this.dropPath) start.dropPath();
> else start.dropLabels(labels);
> }
> }
> {code}
> Again, the more we can prune historic path data no longer needed, the higher
> the probability of bulking. Think about this in terms of {{match()}}.
> {code}
> g.V().match(
> a.out.b,
> b.out.c,
> c.neq.a,
> c.out.b,
> ).select("a")
> {code}
> All we need is "a" at the end. Thus, once a pattern has been passed and no
> future patterns require that label, drop it!
> This idea is related to TINKERPOP-331, but I don't think we should deal with
> manipulating the species. Thus, I think 331 is too "low level."
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
