[jira] [Created] (NIFI-11677) DeleteHDFS is scheduled with empty incoming connection

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

Isha Lamboo created NIFI-11677:
----------------------------------

             Summary: DeleteHDFS is scheduled with empty incoming connection
                 Key: NIFI-11677
                 URL: https://issues.apache.org/jira/browse/NIFI-11677
             Project: Apache NiFi
          Issue Type: Bug
          Components: Extensions
    Affects Versions: 1.19.1, 1.18.0
         Environment: 64bit redhat linux, java 1.8.0_352, 3 node cluster

            Reporter: Isha Lamboo


The DeleteHDFS processor can operate either with or without an incoming connection. When operating *with* an incoming connection that has no flowfiles, it still registers as executing tasks and taking a small amount of time (in the range of 50-100 millis on our clusters). In our case this results in ~200 processors each running 800-1000 tasks per 5 minutes on a 3 node cluster with 16 CPUs per node.

This is in contrast to the PutHDFS with an incoming connection which shows as 0 tasks/0 millis.

Reproduce the situation by connecting a (stopped) ListHDFS processor to a DeleteHDFS processor with a variable for the path. Start the DeleteHDFS and refresh for statistics update.

Having compared the code for PutHDFS and DeleteHDFS I think I see the cause:

PutHDFS has this in its OnTrigger method:
{code:java}
        final FlowFile flowFile = session.get();
        if (flowFile == null) {
            return;
        } {code}
DeleteHDFS has this in the OnTrigger method:
{code:java}
        final FlowFile originalFlowFile = session.get();
        // If this processor has an incoming connection, then do not run unless a
        // FlowFile is actually sent through
        if (originalFlowFile == null && context.hasIncomingConnection()) {
            context.yield();
            return;
        } {code}
I'm guessing that the extra context.yield() causes the framework to execute some administration tasks and update counters, where it does not for PutHDFS.

I don't know how much overhead this causes, probably very little, but I'm trying to improve performance on a cluster with thread starvation issues and high CPU usage and these keep popping up in the summaries looking bad when I think they shouldn't.

The likely solution is to only perform the context.yield() when hasIncomingConnection() returns false.



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