[
https://issues.apache.org/jira/browse/FLINK-9506?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16499947#comment-16499947
]
Stefan Richter commented on FLINK-9506:
---------------------------------------
>From what I can see, the problem is purely related to garbage collection. When
>you use reducing state, your json objects are accumulated on the heap for
>longer time and are no longer short lived. The observed performance variance
>is caused by the GC activity, and the job is indeed producing and releasing
>big amounts of nested structures of small objects. You can think about tuning
>GC and memory/allocation related settings, or think about a different
>in-memory representation (object layout) of your state data. With RocksDB, the
>GC problem would surely go away, but passing your large objects through ser/de
>on every access/update will not be a piece of cake as well for performance.
> Flink ReducingState.add causing more than 100% performance drop
> ---------------------------------------------------------------
>
> Key: FLINK-9506
> URL: https://issues.apache.org/jira/browse/FLINK-9506
> Project: Flink
> Issue Type: Improvement
> Affects Versions: 1.4.2
> Reporter: swy
> Priority: Major
> Attachments: KeyNoHash_VS_KeyHash.png, flink.png
>
>
> Hi, we found out application performance drop more than 100% when
> ReducingState.add is used in the source code. In the test checkpoint is
> disable. And filesystem(hdfs) as statebackend.
> It could be easyly reproduce with a simple app, without checkpoint, just
> simply keep storing record, also with simple reduction function(in fact with
> empty function would see the same result). Any idea would be appreciated.
> What an unbelievable obvious issue.
> Basically the app just keep storing record into the state, and we measure how
> many record per second in "JsonTranslator", which is shown in the graph. The
> difference between is just 1 line, comment/un-comment "recStore.add(r)".
> {code}
> DataStream<String> stream = env.addSource(new GeneratorSource(loop);
> DataStream<JSONObject> convert = stream.map(new JsonTranslator())
> .keyBy()
> .process(new ProcessAggregation())
> .map(new PassthruFunction());
> public class ProcessAggregation extends ProcessFunction {
> private ReducingState<Record> recStore;
> public void processElement(Recordr, Context ctx, Collector<Record> out) {
> recStore.add(r); //this line make the difference
> }
> {code}
> Record is POJO class contain 50 String private member.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
