[ 
https://issues.apache.org/jira/browse/TINKERPOP-2944?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17721818#comment-17721818
 ] 

ASF GitHub Bot commented on TINKERPOP-2944:
-------------------------------------------

FlorianHockmann commented on PR #2058:
URL: https://github.com/apache/tinkerpop/pull/2058#issuecomment-1544154709

   I made the change you suggested, @vkagamlyk. But I also had to make another 
change as my fix wasn't complete. It did solve the problem that the 
`Connection` class had an ever increasing list of cancellation registrations, 
but those registrations weren't disposed and were therefore still stored in the 
underlying `CancellationTokenSource`.
   We've found that by trying out the fix in the service where we initially 
found the memory leak. The earlier fix here already improved the situation a 
lot but a memory dump revealed that the `CancellationTokenSource` had a linked 
list of all cancellation registrations.
   The version that is now in this PR doesn't show any increasing memory usage 
any more (at least right now after executing the application for ~30min which 
was already enough earlier to see the problem, but I will update this comment 
tomorrow to confirm this).




> Memory leak in Gremlin.Net driver if CancellationToken is used
> --------------------------------------------------------------
>
>                 Key: TINKERPOP-2944
>                 URL: https://issues.apache.org/jira/browse/TINKERPOP-2944
>             Project: TinkerPop
>          Issue Type: Bug
>          Components: dotnet
>    Affects Versions: 3.6.2, 3.5.5, 3.6.3, 3.5.6
>            Reporter: Florian Hockmann
>            Assignee: Florian Hockmann
>            Priority: Major
>
> We have noticed in my team at G DATA that a .NET service has been using an 
> increasing amount of memory since we began to propagate the 
> {{CancellationToken}} into the Gremlin.Net traversal executions which was 
> introduced in TINKERPOP-2794.
> We could track this down with memory profiling to [this place in the 
> driver|https://github.com/apache/tinkerpop/blob/5c8af70c7fac3ee98df36d250b05320e67ff1b96/gremlin-dotnet/src/Gremlin.Net/Driver/Connection.cs#L94]
>  where we register a callback to execute when cancellation is requested. 
> These registered callbacks are only cleaned up when the whole {{Connection}} 
> object gets disposed. Since the callback contains a reference to the full 
> {{RequestMessage}} of the traversal, it can contain a lot of data. This is 
> something that I've completely overlooked when I added support for 
> cancellation.
> The driver should clean up those callbacks when they are not needed any more 
> because the request has been processed completely (either successfully or 
> with an error).



--
This message was sent by Atlassian Jira
(v8.20.10#820010)

Reply via email to