[
https://issues.apache.org/jira/browse/TINKERPOP-2393?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17178364#comment-17178364
]
redtree edited comment on TINKERPOP-2393 at 8/15/20, 6:43 PM:
--------------------------------------------------------------
[~spmallette] Thank you for the reply.
Yeah in another thread I noticed that having more barrier steps will break
Aggregate SideEffect's locality more, so it's not so simple to have the batch
call functionality in more steps natively...
I think for now we should aim for the 1st approach. But now I noticed that
currently BranchStep determines if it has a barrier during construction
[https://github.com/apache/tinkerpop/blob/534746b6019d0775e628375a3dfeba8a18df90fe/gremlin-core/src/main/java/org/apache/tinkerpop/gremlin/process/traversal/step/branch/BranchStep.java#L92]
This doesn't work for our case as we swap PropertiesStep with our own
BulkPropertiesStep which would implement Barrier when TraversalStrategies are
applied, so it is too late.
I think BranchStep should defer this check during query run-time, luckily it
already has the flag "first" , so only for the first time we can call this
{{if (!this.hasBarrier &&
!TraversalHelper.getStepsOfAssignableClassRecursively(Barrier.class,
traversalOption).isEmpty())}}{\{ }}
{{ this.hasBarrier = true;}}
in StandardAlgorithm. By this change BranchStep also detects any
NoOpBarrierStep added for optimization but it should be ok.
Does it make sense to you ? I first want to make this change then.
was (Author: redtree1112):
[~spmallette] Thank you for the reply.
Yeah in another thread I noticed that having more barrier steps will break
Aggregate SideEffect's locality more, so it's not so simple to have the batch
call functionality in more steps...
I think for now we should aim for the 1st approach. But now I noticed that
currently BranchStep determines if it has a barrier during construction
[https://github.com/apache/tinkerpop/blob/534746b6019d0775e628375a3dfeba8a18df90fe/gremlin-core/src/main/java/org/apache/tinkerpop/gremlin/process/traversal/step/branch/BranchStep.java#L92]
This doesn't work for our case as we swap PropertiesStep with our own
BulkPropertiesStep which would implement Barrier when TraversalStrategies are
applied, so it is too late.
I think BranchStep should defer this check during query run-time, luckily it
already has the flag "first" , so only for the first time we can call this
{{if (!this.hasBarrier &&
!TraversalHelper.getStepsOfAssignableClassRecursively(Barrier.class,
traversalOption).isEmpty())}}{{ }}
{{ this.hasBarrier = true;}}
in StandardAlgorithm. By this change BranchStep also detects any
NoOpBarrierStep added for optimization but it should be ok.
Does it make sense to you ? I first want to make this change then.
> BranchStep should respect the barrier size of BarrierStep when calling
> applyCurrentTraverser
> --------------------------------------------------------------------------------------------
>
> Key: TINKERPOP-2393
> URL: https://issues.apache.org/jira/browse/TINKERPOP-2393
> Project: TinkerPop
> Issue Type: Improvement
> Components: process
> Affects Versions: 3.4.7
> Reporter: redtree
> Priority: Minor
>
> Right now in BranchStep, if children contain a BarrierStep it keeps calling
> applyCurrentTraverser until this.starts is exhausted.
> [https://github.com/apache/tinkerpop/blob/534746b6019d0775e628375a3dfeba8a18df90fe/gremlin-core/src/main/java/org/apache/tinkerpop/gremlin/process/traversal/step/branch/BranchStep.java#L139-L140]
> This means the child step may hold all upstream solutions in memory at once,
> and potentially memory usage is bloated.
> I think here we can change so that BarrierStep can have a method like
> getMaxBarrierSize, and BranchStep examines the size of barrier that child
> steps can handle at once then break the while loop either when
> `this.starts.hasNext` becomes false or we exceed the barrier size to save the
> potential memory consumption.
>
> When there are multiple barrier steps as children, we may pick either max or
> min among them.
> I am happy to implement this, but please let me know if you have any comments
> or concerns on this change.
> Thanks
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
