[
https://issues.apache.org/jira/browse/NIFI-4475?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16290898#comment-16290898
]
ASF GitHub Bot commented on NIFI-4475:
--------------------------------------
Github user markap14 commented on a diff in the pull request:
https://github.com/apache/nifi/pull/2337#discussion_r156955954
--- Diff:
nifi-nar-bundles/nifi-framework-bundle/nifi-framework/nifi-framework-core/src/main/java/org/apache/nifi/controller/repository/StandardProcessSession.java
---
@@ -1533,13 +1533,11 @@ public FlowFile get() {
return Collections.emptyList();
}
- final Connection connection =
connections.get(context.getNextIncomingConnectionIndex() % connections.size());
-
- return get(connection, new ConnectionPoller() {
+ return get(new ConnectionPoller() {
@Override
public List<FlowFileRecord> poll(final Connection connection,
final Set<FlowFileRecord> expiredRecords) {
return connection.poll(new FlowFileFilter() {
- int polled = 0;
+ volatile int polled = 0;
--- End diff --
We don't need this to be volatile. The variable is not shared between
multiple threads and is not operated on atomically (i.e., we use ++polled,
which is not atomic for volatile variables).
> Processors that use session.get(batchsize) will yield if multiple inbound
> connections exist where at least one connection is empty.
> -----------------------------------------------------------------------------------------------------------------------------------
>
> Key: NIFI-4475
> URL: https://issues.apache.org/jira/browse/NIFI-4475
> Project: Apache NiFi
> Issue Type: Improvement
> Components: Core Framework
> Affects Versions: 1.3.0
> Reporter: Matthew Clarke
> Assignee: Joseph Percivall
> Labels: nifi
>
> There is a difference between how the NiFi framework handles batches of
> incoming data (session.get(batchsize)) versus 1 FlowFile (Session.get()) at
> a time.
> For example PutSyslog does batches and putUDP processes 1 FlowFile at a time.
> With the batch method, a thread is used to poll connection 1 and requests a
> batch of FlowFiles. If it gets at least 1 FlowFile, it sends that
> FlowFile(s) and ends that thread. On next thread it round-robins to the next
> connection (Looped failure relationship for example) and requests a batch
> again. If that connection is empty, the framework assumes there is no work
> to do and yields the processor for the configured "yield duration". So
> regardless of run schedule, this processor will not run again for the
> configured yield duration.
> With processors that only work on 1 FlowFile at a time. The thread will
> round-robin all the inbound connections until it finds a FlowFile. If it
> does not find a FlowFile in any connection the framework will yield the
> processor for the configured yield duration.
> The intent of yield duration is to keep processors with the default runs
> schedule of 0 sec from using excessive CPU doing nothing; however, in the
> case of batches it will yield even if FlowFiles exist on another connection.
> This can have a huge impact on throughput performance of processors that use
> session.get(batchsize)
> There are two possible work-arounds to this issue:
> 1. You should see improved performance when multiple inbound connections
> exist (where any connection may be normally empty) by reducing the configured
> yield duration. The result is better throughput but at the expense of more
> CPU usage when all connections are truly empty.
> 2. Only have one inbound connection to processor that work on batches. This
> can be accomplished by using a funnel.
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
