[
https://issues.apache.org/jira/browse/FLINK-6301?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16097510#comment-16097510
]
Rahul Yadav commented on FLINK-6301:
------------------------------------
[~stevenz3wu]: This issue is only in gzip as gzip uses off heap memory and
since close() method (which releases this off heap memory) of GZipInputStream
doesn't get called in older kafka consumers, we get this leak. By this logic,
if we have another compression scheme which also use off-heap memory, then we
will get the same behavior there as well. One example of such compression codec
in "Zip".
I went through the code of
https://github.com/xerial/snappy-java/blob/master/src/main/java/org/xerial/snappy/SnappyInputStream.java
and it uses Java byte array for byte buffers, hence we won't get the leak
there.
> Flink KafkaConnector09 leaks memory on reading compressed messages due to a
> Kafka consumer bug
> ----------------------------------------------------------------------------------------------
>
> Key: FLINK-6301
> URL: https://issues.apache.org/jira/browse/FLINK-6301
> Project: Flink
> Issue Type: Bug
> Components: Kafka Connector
> Affects Versions: 1.2.0, 1.1.3, 1.1.4
> Reporter: Rahul Yadav
> Assignee: Rahul Yadav
> Fix For: 1.2.2, 1.4.0
>
> Attachments: jeprof.24611.1228.i1228.heap.svg,
> jeprof.24611.1695.i1695.heap.svg, jeprof.24611.265.i265.heap.svg,
> jeprof.24611.3138.i3138.heap.svg, jeprof.24611.595.i595.heap.svg,
> jeprof.24611.705.i705.heap.svg, jeprof.24611.81.i81.heap.svg,
> POSTFIX.jeprof.14880.1944.i1944.heap.svg,
> POSTFIX.jeprof.14880.4129.i4129.heap.svg,
> POSTFIX.jeprof.14880.961.i961.heap.svg, POSTFIX.jeprof.14880.99.i99.heap.svg,
> POSTFIX.jeprof.14880.9.i9.heap.svg
>
>
> Hi
> We are running Flink on a standalone cluster with 8 TaskManagers having 8
> vCPUs and 8 slots each. Each host has 16 GB of RAM.
> In our jobs,
> # We are consuming gzip compressed messages from Kafka using
> *KafkaConnector09* and use *rocksDB* backend for checkpoint storage.
> # To debug the leak, we used *jemalloc and jprof* to profile the sources of
> malloc calls from the java process and attached are the profiles generated at
> various stages of the job. As we can see, apart from the os.malloc and
> rocksDB.allocateNewBlock, there are additional malloc calls coming from
> inflate() method of java.util.zip.inflater. These calls are innocuous as long
> as the inflater.end() method is called after it's use.
> # To look for sources of inflate() method, we used Btrace on the running
> process to dump caller stack on the method call. Following is the stackTrace
> we got:
> {code}
> java.util.zip.Inflater.inflate(Inflater.java)
> java.util.zip.InflaterInputStream.read(InflaterInputStream.java:152)
> java.util.zip.GZIPInputStream.read(GZIPInputStream.java:117)
> java.io.DataInputStream.readFully(DataInputStream.java:195)
> org.apache.kafka.common.record.MemoryRecords$RecordsIterator.makeNext(MemoryRecords.java:253)
> org.apache.kafka.common.record.MemoryRecords$RecordsIterator.makeNext(MemoryRecords.java:210)
> org.apache.kafka.common.utils.AbstractIterator.maybeComputeNext(AbstractIterator.java:79)
> org.apache.kafka.common.utils.AbstractIterator.hasNext(AbstractIterator.java:45)
> org.apache.kafka.common.record.MemoryRecords$RecordsIterator.innerDone(MemoryRecords.java:282)
> org.apache.kafka.common.record.MemoryRecords$RecordsIterator.makeNext(MemoryRecords.java:233)
> org.apache.kafka.common.record.MemoryRecords$RecordsIterator.makeNext(MemoryRecords.java:210)
> org.apache.kafka.common.utils.AbstractIterator.maybeComputeNext(AbstractIterator.java:79)
> org.apache.kafka.common.utils.AbstractIterator.hasNext(AbstractIterator.java:45)
> org.apache.kafka.clients.consumer.internals.Fetcher.handleFetchResponse(Fetcher.java:563)
> org.apache.kafka.clients.consumer.internals.Fetcher.access$000(Fetcher.java:69)
> org.apache.kafka.clients.consumer.internals.Fetcher$1.onSuccess(Fetcher.java:139)
> org.apache.kafka.clients.consumer.internals.Fetcher$1.onSuccess(Fetcher.java:136)
> org.apache.kafka.clients.consumer.internals.RequestFuture.fireSuccess(RequestFuture.java:133)
> org.apache.kafka.clients.consumer.internals.RequestFuture.complete(RequestFuture.java:107)
> org.apache.kafka.clients.consumer.internals.ConsumerNetworkClient$RequestFutureCompletionHandler.onComplete(ConsumerNetworkClient.java:380)
> org.apache.kafka.clients.NetworkClient.poll(NetworkClient.java:274)
> org.apache.kafka.clients.consumer.internals.ConsumerNetworkClient.clientPoll(ConsumerNetworkClient.java:320)
> org.apache.kafka.clients.consumer.internals.ConsumerNetworkClient.poll(ConsumerNetworkClient.java:213)
> org.apache.kafka.clients.consumer.internals.ConsumerNetworkClient.poll(ConsumerNetworkClient.java:193)
> org.apache.kafka.clients.consumer.KafkaConsumer.pollOnce(KafkaConsumer.java:908)
> org.apache.kafka.clients.consumer.KafkaConsumer.poll(KafkaConsumer.java:853)
> org.apache.flink.streaming.connectors.kafka.internal.KafkaConsumerThread.run(KafkaConsumerThread.java:227)
> {code}
> The end() method on Inflater is called inside the close() method of
> *InflaterInputSteam* (extended by *GZIPInputStream*) but looking through the
> Kafka consumer code, we found that RecordsIterator is not closing the
> compressor stream after use and hence, causing the memory leak:
> https://github.com/apache/kafka/blob/23c69d62a0cabf06c4db8b338f0ca824dc6d81a7/clients/src/main/java/org/apache/kafka/common/record/MemoryRecords.java#L210
> https://issues.apache.org/jira/browse/KAFKA-3937 was filed for this and the
> issue was fixed in 0.10.1.0 but not back-ported to previous versions.
> So, I would assume that we have to two paths from here:
> 1. Wait for the changes to be back-ported to 0.9.x Kafka consumer and then,
> update the Kafka-clients dependency:
> https://github.com/apache/flink/blob/release-1.2/flink-connectors/flink-connector-kafka-0.9/pom.xml#L40
> 2. Update the kafka-connector10 to use 0.10.1.0 clients library instead of
> 0.10.0.1.
> https://github.com/apache/flink/blob/release-1.2/flink-connectors/flink-connector-kafka-0.10/pom.xml#L40
> Apart from the master, also back-port the changes to 1.2.x for Kafka
> connector 10 and all the 1.x dependencies for Kafka connector 09.
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
